Kernel methods and their derivatives: Concept and perspectives for the earth system sciences.

Kernel methods are powerful machine learning techniques which use generic non-linear functions to solve complex tasks. They have a solid mathematical foundation and exhibit excellent performance in practice. However, kernel machines are still considered black-box models as the kernel feature mapping cannot be accessed directly thus making the kernels difficult to interpret. The aim of this work is to show that it is indeed possible to interpret the functions learned by various kernel methods as they can be intuitive despite their complexity. Specifically, we show that derivatives of these functions have a simple mathematical formulation, are easy to compute, and can be applied to various problems. The model function derivatives in kernel machines is proportional to the kernel function derivative and we provide the explicit analytic form of the first and second derivatives of the most common kernel functions with regard to the inputs as well as generic formulas to compute higher order derivatives. We use them to analyze the most used supervised and unsupervised kernel learning methods: Gaussian Processes for regression, Support Vector Machines for classification, Kernel Entropy Component Analysis for density estimation, and the Hilbert-Schmidt Independence Criterion for estimating the dependency between random variables. For all cases we expressed the derivative of the learned function as a linear combination of the kernel function derivative. Moreover we provide intuitive explanations through illustrative toy examples and show how these same kernel methods can be applied to applications in the context of spatio-temporal Earth system data cubes. This work reflects on the observation that function derivatives may play a crucial role in kernel methods analysis and understanding.

1 Introduction

Kernel methods (KMs) constitute a standard set of tools in machine learning and pattern analysis [1, 2]. They are based on a mathematical framework to cope with nonlinear problems while still relying on well-established concepts of linear algebra. KMs are one of the preferred tools in applied sciences, from signal and image processing [3], to computer vision [4] and geosciences [5]. Since its introduction in the 1990s through the popular support vector machines (SVMs), kernel methods have evolved into a large family of techniques that cope with many problems in addition to classification. Kernel machines have also excelled in regression, interpolation and function approximation problems [3], where Gaussian Processes (GPs) [6] and support vector regression [7] have provided good results in many applications. Furthermore, many kernel methods have been engineered to deal with other relevant learning problems; for example, density estimation via kernel decompositions using entropy components [8]. For dimensionality reduction and feature extraction, there are a wide family of multivariate data analysis kernel methods such as kernel principal component analysis [9], kernel canonical analysis [10] or kernel partial least squares [11]. Kernels have also been exploited to estimate dependence (nonlinear associations) between random variables such as kernel mutual information [12], or the Hilbert-Schmidt Independence Criterion [13]. Finally in the literature, we find kernel machines for data sorting [14], manifold learning and alignment [15], system identification [16], signal deconvolution and blind source separation [3].

However, understanding a model is more difficult than just applying a model, and kernel methods are still considered black-box models. Little can be said about the characteristics of the feature mapping which is only implicit in the formulation. Several approaches have been presented in the literature to explore the kernel feature mapping and to understand what the kernel machine is actually learning. One way to analyze kernel machines is by visualizing the empirical feature maps but this is very challenging and only feasible in low-dimensional problems [1, 17]. Another approach is to study the relative relevance of the input features (covariates) on the output. This is commonly referred to as feature ranking and it typically reduces to evaluating how the function varies when an input is removed or perturbed. Automatic relevance determination (ARD) kernels [6] or multiple kernel learning [18] allow one to study the relevance of the feature components indirectly. While this approach has been extensively used to improve the accuracy and understanding of supervised kernel classifiers and regression methods, they only provide feature ranking and nothing is said about the geometrical properties of the feature map. In order to resolve this, two main approaches are available in the kernel methods literature. For some particular kernels one can derive the metric induced by the kernel to give insight into the surfaces and structures [19]. Alternatively, one can study the feature map (in physically meaningful units) by learning the inverse feature mapping; a group of techniques known as kernel pre-imaging [20, 21]. However, the current methods are computationally expensive, involve critical parameters, and very often provide unstable results.

Function derivatives is a classical way to describe and visualize some characteristics of models. Derivatives of kernel functions have been introduced before, yet mostly used in supervised learning as a form of regularization that controls fast variations of the decision function [22]. However, derivatives of the model’s function with regards to the input features for feature understanding and visualization has received less attention. A recent strategy is to derive sensitivity maps from a kernel feature map [23]. The sensitivity map is related to the squared derivative of the function with respect to the input features. The idea was originally derived for SVMs in neuroimaging applications [24], and later extended to GPs in geoscience problems [25-28]. In both cases, the goal was to retrieve a feature ranking from a learned supervised model.

In this paper, we analyze the kernel function derivatives for supervised and unsupervised kernel methods with several kernel functions in different machine learning paradigms. We show the usefulness of the derivatives to study and visualize kernel models in regression, classification, density estimation, and dependence estimation with kernels. Since differentiation is a linear operator, most kernel methods have a derivative that is proportional to the derivative of the kernel function. We provide the analytic form of the first and second derivatives of the most common kernel functions with regards to the inputs, along with iterative formulas to compute the m-th order derivative of differentiable kernels, and for the radial basis function kernel in particular; where m is the number of successive derivatives. In classification problems, the derivatives can be related to the margin, and allow us to gain some insight on sampling [29]. In regression problems, a models’ function derivatives may give insight about the signal and noise characteristics that allow one to design regularization functionals. In density estimation, the second derivative (the Hessian) allows us to follow the density ridge for manifold learning [30], whereas in dependence estimation squared derivatives (the sensitivity maps) allows one to study the most relevant points and features governing the association measure [31]. All in all, kernel derivatives allow us to identify both examples and features that affect the predictive function the most, and allow us to interpret the kernel model behavior in different learning applications. We show that the solutions can be expressed in closed-form for the most common kernel functions and kernel methods, they are easy to compute, and we give examples of how they can be used in practice.

The remainder of the paper is organized as follows. Section 2 briefly reviews the fundamentals of kernel functions and feature maps, and concentrates on the kernel derivatives for feature map analysis where we provide the first and second order derivatives for most of the common kernel functions. We also review the main ideas to summarize the information contained in the derivatives. Section 3 and Section 4 study popular discriminative kernel methods, such as Gaussian Processes for regression and support vector machines for classification. Section 5 analyzes the interesting case of density estimation with kernels, in particular through the use of kernel entropy component analysis for density estimation. Section 6 pays attention to the case of dependence estimation between random variables using the Hilbert-Schmidt independence criterion in cases of dependence visualization maps and data unfolding. Section 7 illustrates the applicability of kernel derivatives in the previous kernel methods on spatio-temporal Earth system science data. We conclude in section 8 with some final remarks.

2 Kernel functions and the derivatives

2.1 Kernel functions and feature maps

In this section, we briefly highlight the most important properties of kernel methods, needed to understand their role of the kernel methods mentioned in the subsequent sections. Recall that kernel methods rely on the notion of similarity between points in a higher (possibly infinite) dimensional Hilbert space. Let us consider a set of empirical data , whose elements are defined in a d-dimensional input space, , 1 ≤ i ≤ n. In supervised settings, each input feature vector is associated with a target value, which can be either discrete in the classification case, or real in the regression case, , i = 1, …, n. Kernel methods assume the existence of a Hilbert space with an inner product where samples in are mapped into with a feature map , 1 ≤ i ≤ n. The mapping function can be defined explicitly (if some prior knowledge about the problem is available) or implicitly, which is often the case in kernel methods. The similarity between the elements in can be estimated using its associated dot product via reproducing kernels in Hilbert spaces (RKHS), , such that pairs of points (, ′) ↦ k(, ′). So we can estimate similarities in without the explicit definition of the feature map φ, and hence without having access to the points in . This kernel function k is required to satisfy Mercer’s Theorem [32].

Definition 1 Reproducing kernel Hilbert spaces (RKHS) [33]. A Hilbert space is said to be a RKHS if: (1) The elements of are complex or real valued functions f(⋅) defined on any set of elements ; And (2) for every element , f(⋅) is bounded.

The name of these spaces comes from the so-called reproducing property. In a RKHS , there exists a function k(⋅, ⋅) such that by virtue of the Riesz Representation Theorem [34]. In particular, for any

A large class of algorithms have originated from regularization schemes in RKHS. The representer theorem gives us the general form of the solution to the common loss function formed by the loss term and a regularization term.

Theorem 1 (Representer Theorem) [34, 35] Let be a strictly monotonic increasing function; let be an arbitrary loss function; and let be a RKHS with reproducing kernel k. Then: admits a space of functions f defined as which is expressed as a linear combination of kernel functions. Also note that the previous theorem states that solutions imply having access to an empirical risk term V and a regularizer Ω. In the case of not having labels y, alternative representer theorems can be equally defined. A generalized representer theorem was introduced in [36], which generalizes Wahba’s theorem to a larger class of regularizers and empirical losses. Also, in [37], a representer theorem for kernel principal components analysis (KPCA) was used: the theorem gives the solution as a linear combination of kernel functions centered at the input data points, and is called the representer theorem of learning theory [38], whereby the coefficients are determined by the eigendecomposition of the kernel matrix [9, 36]. Should the reader want more literature related to kernel methods, we highly recommend this paper [39] for a more theoretical introduction to Hilbert-Spaces in the context of kernel methods and [3] for a more applied and practical approaches.

2.2 Derivatives of linear expansions of kernel functions

Computing the derivatives of function f can give important insights about the learned model. Interestingly, in the majority of kernel methods, the function f is linear in the parameters , cf. Eq (4) derived from the representer theorem [35] [Th. 1]. For the sake of simplicity, we will denote the partial derivative of f w.r.t. the feature x as , where j denotes the dimension. This allows us to write the partial derivative of f as: where and . It is possible to take the second order derivative with respect to feature x twice which remains linear as well with : where . Inductively, the m-th partial derivative w.r.t the j-th feature is also linear with and it follows the following equation:

The gradient of f gives information about the slope (increase rate) of the function and reduces to where ∇ denotes the vector differential operator, and . The Laplacian accounts for the curvature, roughness, or concavity of the function itself, and can be easily computed as the sum of all the unmixed second partial derivatives, which for kernels reduces to where and is a column vector of ones of size d. Another useful descriptor is the Hessian matrix of f, which characterizes its local curvature. The Hessian is a d × d matrix of second-order partial derivatives with respect to the features x, x:

The equations listed above have shown that the derivative of a kernel function is linear with . Once the is computed, the problem reduces to (1) computing the derivatives for a particular kernel function, and (2) to summarize the information contained within the derivatives.

Derivatives of common kernel functions. Kernel methods typically use a set of positive definite kernel functions, such as the linear, polynomial (Poly), hyperbolic tangent (Tanh), Gaussian (RBF) kernel, and the automatic relevance determination (ARD) kernel. We give the partial derivative for all of these kernels in Table 1, and the (mixed) second derivatives in Table 2. For the most widely used kernels (RBF and ARD), one can recognize a linear relation between the kernel derivative and the kernel function itself. It can be shown that the m-th derivative of some kernel functions can be computed recursively using Faà di Bruno’s identity [40].

Table 1

Partial derivatives for some common kernel functions: Linear, Polynomial (Poly), Radial Basis Functions (RBF), Hyperbolic tangent (Tanh), and Automatic Relevance Determination (ARD).

Kernel	Kernel function, k(x, y)	Partial derivative, ∂k(x,y)∂x^j
Linear	x⊤ y	yj
Poly	(γx⊤ y + c0)p	γpyj (γx⊤ y + c0)p−1
RBF	exp(−γ∥x − y∥2)	-2γ(xj − yj)k(x,y)
Tanh	tanh(γ x⊤ y + c0)	γyj sech2 (γ x⊤ y + c0)
ARD	ν^2exp(-12∑d=1D(x^d-y^dλ_d)^2)	(x^j-y^jλj2)k(x,y)

Table 2

Second derivatives for some common kernel functions.

Kernel	2nd partial derivative, ∂^2k(x,y)∂x^j^2	Mixed partial derivative, ∂^2k(x,y)∂x^j∂x^k
Linear	0	0
Poly	(p − 1)p (γyj)2 (γ x⊤ y + c0)p−2	(p − 1)pγ2 yj yk (γ x⊤ y + c0)p−2
RBF	2γ [2γ (xj − yj)2 − 1] k(x, y)	4γ2 (xk − yk) (xj − yj) k(x, y)
Tanh	−2(γyj)2 sech2(γ x⊤ y + c0)k(x, y)	−2γ2 yj yk sech2(γ x⊤ y + c0)k(x, y)
ARD	(1λj2+(x^j-y^jλj2)^2)k(x,y)	(x^j-y^jλj2)(x^k-y^kλk2)k(x,y)

Summarizing function derivatives. Summarizing the information contained in the derivatives is not an easy task, especially in high dimensional problems. The most obvious strategy is to use the norm of the partial derivative, that is ‖∂ f‖, which summarizes the relevance of variable x. A small norm implies a small change in the discriminative function f with respect to the j-th dimension, indicating the low importance of that feature. This approach was introduced as sensitivity maps (SMs) in [24] for the visualization of SVM maps in neuroimaging and later exploited in GPs for ranking spectral channels in geosciences applications [26]. The SM for the j-th feature, is the expected value of the squared derivative of the function with respect the input argument x: where p(x) is the probability density function (pdf) over dimension j of the input space . In order to avoid the possibility of cancellation of the terms due to its signs, the derivatives are squared. Other transformations like the absolute value could be equally applied. The empirical sensitivity map approximation to Eq (11) is obtained by replacing the expected value with a summation over the available n samples which can be grouped together to define the sensitivity vector as s = [s1, …, s]⊤.

This can be thought of as studying the relevance of the sample points. Similarly,one can average over the features to obtain a point sensitivity: which can be grouped to define the point sensitivity vector as q = [q1, …, q]⊤. The information contained in q is related to the robustness to changes of the decision in each point of the space.

Now we are equipped to use the derivatives and the corresponding sensitivity maps in arbitrary kernel machines that use standard kernel functions. In the following sections, we study its use in kernel methods for both supervised (regression and classification) and unsupervised (density estimation and dependence estimation) learning.

3 Kernel regression

3.1 Gaussian Process Regression

Multiple proposals to use kernel methods in a regression framework have been done during the last few decades. Gaussian Processes (GPs) is perhaps the most successful kernel method for discriminative learning in general and regression in particular [6]. Standard GP regression approximates observations as the sum of some unknown latent function f() of the inputs plus some additive Gaussian noise, y = f() + ε, where . A zero mean GP prior is placed on the latent function f() and a Gaussian prior is used for each latent noise term ε, in other words , where m() = 0, and K is a covariance function, [K] = k(, ), parameterized by a set of hyperparameters (e.g. = [λ1, …, λ] for the ARD kernel function).

If we consider a test location * with the corresponding output y*, a prior induces a prior distribution between the observations and y*. Collecting all available data in , it is possible to analytically compute the posterior distribution over the unknown output y* given the test input * and the available training set , , which is a Gaussian with the following mean and variance: where contains the kernel similarities of the test point * to all training points in , K is a n × n kernel (covariance) matrix whose entries contain the similarities between all training points, , k** = k(*, *) is a scalar with the self-similarity of *, and is the identity matrix. The solution of the predictive mean for the GP model in (14) is expressed in the same way as equation (4), where . This expression is exactly the same as in other kernel regression methods like the Kernel Ridge Regression (KRR) [2] or the Relevance Vector Machine (RVM) [2]. The derivative of the mean function can be computed through Eq (5) and the derivatives in Table 1.

3.2 Derivatives and sensitivity maps

Let us start by visualizing derivatives in simple 1D examples. We used GP modeling with a standard RBF kernel function to fit five regression data sets. We show in Fig 1 the first and second derivatives of the fitted GP model, as well as the point-wise sensitivities. In all cases, first derivatives are related to positive or negative slopes, while the second derivatives are related to the curvature of the function. Since the derivative is a linear operator, a composition of functions is also the composition of derivatives as can be seen in the last two functions. This could be useful for analyzing more complex composite kernels. See Table 3 for a comparison with other kernel methods derivatives.

Fig 1

Different examples of functions, derivatives and sensitivity maps.

Original data (red), the GP predictive function (black), high derivative values are in yellow, close-to-zero derivative values are in gray and negative derivative values are in blue.

Table 3

Summary of the formulation for each of the main kernel methods GPR (Gaussian Process Regression, section 3), SVM (Support Vector Machines, section 4), KDE (Kernel Density Estimation, section 5), HSIC (Hilbert-Schmidt Independence Criterion, section 6).

The derivative formulation as well as some related analysis procedures in the literature as well as demonstrated in this paper.

Method	Function	Derivative	Analysis
GPR	k*⊤α	∂k*⊤α	Sensitivity, Ranking, Regularization
SVM	g(yα k* + b)	(1-g^2(x_*))∂k*⊤yα	Sensitivity, Feature Ranking, Margin
KDE	n^-1k_*1_n	∇p^(x_*)^⊤E_r(x_*)	Principal Curves
HSIC	n−2Tr(KHLH)	2n−2 Ai ∂q k(xi)	Leverage, Feature/Point Relevance

3.3 Derivatives and regularization

We show an example of applying the derivative of the kernel function as a regularization parameter for the noise. We modeled the function f(x) = sin(3πx) with an additive white Gaussian noise (AWGN) using a Kernel Ridge Regression (KRR) model with RBF kernel. Different amounts of noise power was used resulting in different values of the signal to noise ratio (SNR), SNR = , SNR∈[0, 50] dB. Two different settings were explored to analyze the impact of the standard regularizer, , and the derivatives in KRR modeling: (1) either using the optimal amount of regularization in (14), , or (2) assuming no regularization was needed, .

Four scenarios were explored in this experiment: , , , and , where K is a matrix with entries [K] = k(, ) (for definitions of gradients see Eqs (8) and (9)). The resulting SNR curves were then normalized in such a way that they are comparable. We explore two scenarios; the regularized and unregularized. Since the maximum SNR was subtracted from all norm values, in Fig 2a any norm greater than zero signifies the need to regularize more and in Fig 2b any norm less than zero signifies the need to regularize less.

Fig 2

Signal-to-Noise Ratio (SNR) versus the expected normalized value of different norms () to act as regularizers.

A unregularized (left) and an regularized (right) Kernel Ridge Regression (KRR) model was fitted. The top row shows a few examples of these fitted KRR models with a different quantity of noise added. The red data points are the data with different noise levels, the true function is black and the fitted KRR model is in blue. The second row shows the norm for the different regularizers. All lines were normalized in such a way that they are comparable. The norm of the true signal (SNR = 50 dB) is subtracted from all points so any curve with values below zero require less regularization and any points above zero require more regularization.

Fig 2 shows the effect of the noise on the norm for different regularization terms. All four regularization functions give the user information about how noisy the signal is for the unregularized case and the regularized case. The graph for the regularized case has the norms of the functions below zero, except for the , when the SNR is extremely low. Since the norm of the functions are increasing as one increases the SNR, this says that there needs to be less regularization. The has a straight line because the ‘optimal’ parameter for using the norm of the weights for regularization has already been chosen. However, the norm of the first and second derivative still give us information that the problem needs to be regularized less. So both cases showcase the functionality of the first and second derivative as viable regularizers.

4 Kernel classification

4.1 Support vector machine classification

The first effective and influential kernel method introduced was the Support Vector Machine (SVM) [1, 41–43] classifier. Researchers and practitioners have used it to solve problems in speech recognition [44], computer vision and image processing [45-47], or channel equalization [48]. The binary SVM classification algorithm minimizes a weighted sum of a loss and a regularizer where the cost function is called the ‘hinge loss’ and is defined as , ∈ {−1, + 1}, and is the RKHS of functions generated by the kernel k, and λ is a parameter that trades off accuracy for smoothness. The norm is generally interpreted as a roughness penalty, and can be expressed as a function of kernels, . The decision function for any test point * is given by where α are Lagrange multipliers obtained from solving a quadratic programming (QP) problem, being the support vectors (SVs) of those training samples with non-zero Lagrange multipliers α ≠ 0 [1]. See [49] for more details on the formulation and more practical examples.

4.2 Function derivatives and margin

The SVM decision function in (16) uses a mask function g(x) = sgn(⋅) to decide between the two classes, which is inherited from the hinge loss used. Since the sgn(⋅) function is not differentiable at 0 and for the sake of analytic tractability we replaced it with the hyperbolic tangent, g(⋅) = tanh(⋅). Now one can simply compute the derivative of the model by applying the chain rule: where the leftmost term in the product can be seen as a mask function on top of the derivative of the regression function and allows us to study the model in terms of decision and estimation separately. See Table 3 for a comparison to other kernel methods derivatives.

Three datasets were used to illustrate the effect of the derivative in the SVM classifier. We used a SVM with RBF kernel in all cases, and hyperparameters were tuned by 3-fold cross-validation and the results are displayed in Fig 3. The mask function only focuses on regions along the decision boundary. However the derivative of the kernel function displays a few regions along the decision boundary along with other regions outside of the decision boundary. The composite of the derivative of the masking function and kernel function showcases a combination of the two components: the high derivative regions along the decision boundary. The two half moons and two circles examples have a clear decision boundary and the derivative of the composite function is able to capture this. However, the two ellipsoid example is less clear as the decision boundary passes through two overlapping classes. This is related to the density within the margin as the regions with less samples have a smaller slope and the regions with more samples have a higher slope, which results in wider and thinner margin, respectively. This fact could be used to define more efficient sampling procedures.

Fig 3

Visualizing three examples of sensitivity maps in SVM classification.

The top row shows a figure has red and green points to showcase the classes, black points showing the support vectors chosen by the SVM classifier, and a contour map showcasing the same color scheme for the decision function. In the subsequent plots, we plot the sensitivity measures where the high derivative values are in yellow and negative derivative values are in gray. The leftmost column showcases which derivative is plotted.

5 Kernel density estimation

The problem of density estimation is difficult in machine learning and statistics and it has been widely studied via kernels [50-52]. Kernel density estimation (KDE) is a classical non-parametric method for estimating a probability density function (pdf) [53]. In KDE, the choice of the kernel function is key to properly approximating the underlying pdf from a finite number of samples. The KDE kernel must be a non-negative function that integrates to one (i.e. a proper pdf), yet does not need to be positive semi-definite (PSD). KDE is versatile in that sense. However, if the kernel is PSD, there are close relations between density estimation and RKHS learning via the kernel eigendecomposition. Many KDE kernels are PSD, and some well-known examples include the Gaussian kernel, the Student kernel and the Laplacian kernel [54] functions.

5.1 Density estimation with kernels

For the Parzen window expression, KDE defines the pdf as a sum of kernel functions defined on the training samples, where * is the vector of kernel evaluations between the point of interest *, and all training samples (see section 3.1). KDE kernel functions have to be non-negative and integrate to one to ensure that is a valid pdf. When a point-dependent weighting, β, is employed, then the above expression can be modified as , where the β have to be positive and sum to one, i.e. β ≥ 0 and . In [51] a solution to find a suitable vector based on kernel principal components analysis was proposed. If the decomposition of the un-centered kernel matrix follows the form K = EDE⊤, where E is orthonormal and D is a diagonal matrix, then the kernel-based density estimation can be expressed as where E is the reduced version of E by keeping r < n top eigenvectors. If we keep all the dimensions, i.e. r = n the solution reduces to (18). By reducing the number of components we restrict the capacity of the density estimator and hence obtain a smoother approximation of the pdf as r reduces.

The retained kernel components should be selected by keeping the dimensions that maximize a sensible pdf characteristic, e.g. the variance. However, other criteria can be used to select the retained components. For instance, the kernel entropy component analysis (KECA) method uses the information potential as criterion to select the components from the eigenvector decomposition [8]. In this case, the decomposition method is already optimized to maximize the variance, therefore the solution will be sub-optimal. A more accurate way of finding a decomposition was presented in [55] where the features are directly optimized to maximize the amount of retained information. This method was named optimized KECA (OKECA), and showed excellent performance using very few extracted components.

The relevant aspect for this paper is that, by doing , Eqs (18) and (19) can be cast in the general framework of kernel methods we proposed in Eq (4). Through this equality the derivatives and the second derivatives (and therefore the Hessian) can be obtained in a straightforward manner using Eqs (5) and (6). This information can be used for different problems, such as computing the Fisher’s information matrix, optimizing vector quantization systems, or the example in the following section where we use them to find the points that belong to the principal curve of the distribution.

5.2 Derivatives and principal curves

This example illustrates the use of kernel derivatives in the KDE framework. In particular, we use the gradient and the Hessian of the pdf, to find points that belong to the principal curve along the data manifold [56]. A principal curve is defined as the curve that passes through the middle of the data. How to find this curve in practice is an important problem since multiple data description methods are based on drawing principal curves [30, 57–60]. In [30], they characterize the principal curve as the set of points that belong to the ridge of the density function. These points can be determined by using the gradient and the Hessian of the pdf: a point * is an element of the d-dimensional principal curve iff the inner product of the gradient, , and at least r eigenvectors of the Hessian, H(*), is zero: where E(*) are the top r eigenvectors of the matrix H(*). Note that applying this definition using our framework is straightforward as we can use the KDE to describe the probability density function, and Eqs (5) and (6), as well as formulas in Table 1, to find the gradient and the Hessian of the defined pdf with respect to the points. See Table 3 for a comparison to other kernel methods derivatives.

In Fig 4, we show an illustrative example of this application in three different toy datasets. The pdf can be obtained from the data points by using the OKECA method and the derivative lines describe the direction to which the density changes the most. The last row shows the points of the dataset with smaller dot products between the gradient and the last eigenvector of the Hessian, see Eq (20). Note that these points belong to the ridge of the distribution, and thus to the principal curve.

Fig 4

First row: Original data points. Second, third and fourth row: probability density in gray scale (brighter means denser). Second row: derivative direction of the pdf for some data points is represented using red lines. Third row: Hessian eigenvectors for some points represented with blue lines (first eigenvector) and green lines (second eigenvector). Fourth row: points on the ridge computed using the formula proposed in [30], different brightness of green has been computed using the Dijkstra distance over the curve dots (see text for details).

6 Kernel dependence estimation

6.1 Dependence estimation with kernel methods

Measuring dependencies and nonlinear associations between random variables is an active field of research. The kernel-based dependence estimation defines a covariance and cross-covariance operators in RKHS, and the subsequent statistics from these operators allows one to measure dependence between functions therein.

Let us consider two spacess and , which we jointly sample observation pairs (, ) from distribution . The covariance matrix is , where is the expectation with respect to , and . A statistic that summarizes the content of the covariance matrix is its Hilbert-Schmidt norm. This quantity is zero if and only if there exists no second order dependence between and .

The nonlinear extension of the notion of covariance was proposed in [13] to account for higher order statistics. Essentially, let us define a (possibly non-linear) mapping such that the inner product between features is given by a PSD kernel function k(, ′). The feature space has the structure of a RKHS. Similarly, we define with associated kernel function l(, ′). Then, it is possible to define a cross-covariance operator between these feature maps, and to compute the squared norm of the cross-covariance operator, , which is called the Hilbert-Schmidt Independence Criterion (HSIC) and can be expressed in terms of kernels [61, 62]. Given a sample dataset of size n drawn from , an empirical estimator of HSIC is [13]: where Tr(⋅) is the trace operation, K, L are the kernel matrices for the input random variables and y (i.e. [K] = k(, )), respectively, and centers the data in the feature spaces and , respectively. HSIC has demonstrated its capability to detect dependence between random variables but, as for any kernel method, the learned relations are hidden behind the kernel feature mapping. To address this issue, we consider the derivatives of HSIC.

6.2 Derivatives of HSIC

HSIC empirical estimate is parameterized as a function of two random variables, so the function derivatives given in section 2 are not directly applicable. Since HSIC is a symmetric measure, the solution for the derivative of HSIC wrt will have the same form as the derivative wrt . For convenience, we can group all terms that do not explicitly depend of X as A = HLH, which allows us expressing (21) simply as:

Note that the core of the solution is the same as in the previous sections; a weighted combination of kernel similarities. However, now we need to derive both arguments of the kernel function k with respect to entry that appears twice. By taking derivatives with regards to a particular dimension q of sample , i.e. , and noting that the derivative of a kernel function is a symmetric operation, i.e. , one obtains where A is the i-th row of the matrix A. For the the RBF kernel we obtain [31]: where entries of matrix are (1 ≤ j ≤ n), and zeros otherwise, and where the symbol ∘ is the Hadamard product between matrices.

Recently [63] extended the notion of leverage scores for the ridge regression problem. Leverage is a measure of how points with low density neighbours are enforcing the model for passing through them. By definition, the leverage (of a regressor) is the sensitivity of the predictive function w.r.t. the outputs. There is no definition of leverage in the case of HSIC as it is not a regression model but a dependence measure. However, HSIC could be interpreted in a similar way by fixing one of the variables and taking the derivative w.r.t. the other. By this interpretation, one can think of the HSIC sensitivity as a measure of how individual points are affecting the dependence measurement, i.e. how sensitive HSIC is to the perturbations for each particular point. This interpretation allows us to link the concepts of leverage and sensitivity in kernel dependence measures.

In this case, the derivatives of HSIC report information about the directions that impact the dependence estimate the most. This allows one to evaluate the measure as a vector field representation of two components. As in the previous kernel methods analyzed, the derivatives here are also analytic, just involving simple matrix multiplications and a trace operation. See Table 3 for a comparison to other kernel methods derivatives.

6.3 Visualizing kernel dependence measures

HSIC derivatives give information about the contribution of each point and feature to the dependence estimate. Fig 5 shows the directional derivative maps for three different bi-dimensional problems of variable association. We show the different components of the (sign-valued) vector field as well as its magnitude. In all problems, arrows indicate the strength of distortion to be applied to points (either in directions x, y, or jointly) such that the dependence is maximized. For the first example (top row), the map pushes the points into the 1-1 line and tries to collapse data into 2 different clusters along this line. In the second example (middle row), the distribution is a noisy ring: here the sensitivity map tries to collapse the data into clusters in order to maximize the dependence between the variables. In the last third experiment (bottom row), both variables are almost independent and the sensitivity map points towards some regions in the space where the dependence is maximized. In all cases, the S and S are orthogonal in direction and form a vector field whose intensity can be summarized in its norm |S| (columns in the figure).

Fig 5

Visualizing the derivatives and the modulus of the directional derivative for HSIC in three toy examples.

6.4 Unfolding and independization

We have seen that the derivatives of the HSIC function can be useful to learn about the data distribution and the variable associations. The derivatives of HSIC give information about the directions most affecting the dependence or independence measure.

Fig 6 shows an example of how the derivatives of the HSIC can be used to modify the data and achieve either maximum dependence or maximum independence. We embedded the derivatives in a simple gradient descent scheme, in which we move samples iteratively to maximize or minimize data dependence. Departing from a sinusoid, one can attain dependent or independent domains.

Fig 6

Modification of the input samples to maximize of minimize HSIC dependence between their dimensions (see text for details).

Note that HSIC can be understood as a maximum mean discrepancy (MMD) [64] between the joint probability measure of the involved variables and the product of their marginals, and MMD derivatives are very similar to those of HSIC provided here. The explicit use of the kernel derivatives would allow us to use gradient-descent approaches in methods that take advantage of HSIC or MMD, such as in algorithms for domain adaptation and generative modeling.

7 Analysis of spatio-temporal earth data

Kernel methods are widely applied in the Earth system sciences [5], where they have proven to be effective when dealing with low numbers of (potentially high dimensional) training samples. Data of this kind are characteristic for hyperspectral data, multidimensional sensor information, and different noise sources in the data. The most common applications in Earth system sciences are anomaly and target detection [65], the estimation of biogeochemical or biophysical parameters [66-68], dimensionality reduction [15, 69, 70], and the estimation of data interdependence [31]. However, so far multivariate spatio-temporal data problems have received comparable little attention [71, 72], and in particular regarding the use of the derivatives of kernel methods [25, 26]. This is surprising, given the high-dimensional nature of most spatio-temporal dynamics in most sub-domains of the Earth system, e.g. land-surface dynamics, land-atmosphere interactions, ocean dynamics, etc. [73]. Hence, this section explores the added value of kernel derivatives for analyzing multivariate spatio-temporal Earth system data. We showcase applications considering the four studied problems of classification, regression, density estimation and dependence estimation. Please see github.com/IPL-UV/sakame for a working implementation of the algorithms as well as the subsequent ESDC experiments.

7.1 Spatio-temporal earth data

Today, data-driven research into Earth system dynamics has gained momentum and complements global modelling efforts. Much of Earth data is generated by a wide range of satellite sensors, upscaled products from in-situ observations, and model simulations with constantly improving spatial and temporal resolutions. The question is whether using kernel derivatives may help in (1) choosing the appropriate space and time scales to analyze phenomena, (2) visualize the most informative areas of interest, and (3) detect anomalies in spatio-temporal Earth data. We will work with products contained in the Earth System Data Lab (ESDL) [73]. The analysis-ready data-cube contains and harmonizes more than 40 variables relevant to monitor key processes of the terrestrial land-surface and atmosphere. The data streams contained in the ESDL are grouped in three data streams: land surface, atmospheric forcings and socio-economic data. Here we focus on three land-surface variables which exhibit nonlinear relations in space and time. The following three variables; the gross primary productivity (GPP), root-zone soil moisture (SM), and land surface temperature (LST); are outlined below:

GPP is the rate of fixation of carbon dioxide through the photosynthesis and one of the largest single flux in the global carbon cycle. However, the process is sensitive to climate variability. For instance, it has been shown that regional extreme events like droughts, heatwaves, and other types of disturbances may even influence the inter-annual variability of the globally integrated GPP [74]. Hence, it is key to understand the spatial and temporal dynamics of GPP at regional and global scales. Here, we consider the GPP FLUXCOM (http://www.fluxcom.org/) product, computed as described in [75, 76].

SM plays a fundamental role for the environment and climate system, as it influences hydrological and agricultural processes, runoff generation and drought development processes, and land-atmospheric feedbacks [77] There are two products of soil moisture in our experiments. Standard SM products carry information limited to a few centimeters below the surface (±5 cm), and do not allow access to the whole zone from where water can be absorbed by roots. This is why we used root-zone soil moisture (RSM) [78-80] in the dependence estimation problem instead, a product from GLEAM that is a more sensitive variable to monitor water stress and droughts in vegetation.

LST is an essential variable within the Earth climate system as it influences processes such as the exchange of energy and water between the land surface and atmosphere, and influences the rate and timing of plant growth. The LST product contained in the ESDL is the result of an ESA project called GlobTemperature, that developed a merged LST data set from thermal infrared (geostationary and polar orbiters) and passive microwave satellite data to provide best possible coverage.

The data is organized in 4-dimensional data cube x(u, v, t, k) involving (latitude, longitude) spatial coordinates (u, v), time sampling t, and the variable k. The data in ESDL contains a spatial resolution (high 0.083° resolution and coarser grid aggregation at 0.25°) and a temporal resolution of 8 days spanning the years 2001-2011. In our experiments, we focus on the lower resolution products, during 2008-2010, and over Europe only. In the year 2010, a severe combination of spring and summer drought combined with a summer heat stress event affected large parts of Russia which can be observed in the three variables under study here [81], and we expect that also their interrelations must be affected. We use this well known event to provide a proof of concept for our suggestion approaches to interpret regressions, principal curves, and dependence estimation.

7.2 Sensitivity analysis in GP modelling

Studying time-varying processes with GPs is customary. Designing a GP becomes more complicated when dealing with spatio-temporal datasets. This can be cumbersome when the final goal is to understand and visualize spatial dependencies as well as to study the relevance of the features and the samples. Sensitivity analysis can be useful for either scenario. In this experiment, we study the impact of features in the GP modeling of the GPP and LST variables during 2010. To do so, we developed GP regression models trained to predict a pixel from their neighbourhood pixels. This is similar to geographically weighted regression [82] which can be used to model the local spatial relationships between these features and the outputs. From this framework, we can get sensitivity values for each of the contributing dimensions. We further split the data into subsets of spatial ‘minicubes’ which ranged in size from 2 × 2 until size 7 × 7. We use a GP model on a training subset of minicubes whereby the neighbours were used as input dimensions to predict the center pixel for both GPP and LST. For metrics, we used the R2-value to measure the goodness of fit between our model and the real data.

Fig 7 show how the sensitivity changes according to the mean prediction of the GPP and LST for two neighbourhood spatial window sizes (3 × 3 and 5 × 5). It also shows the spatial sensitivity maps for both settings and the R2-value and the average sensitivity for each GP model. What’s reassuring is that we see consistently low sensitivity values in areas (e.g. near the Black and Caspian Sea) for the GPP and LST regardless of the spatial window size as these are typically areas of low GPP and SM. For GPP, we see that sensitivities tend to become smoother as the neighbourhood size increases. These particular maps for GPP reach an R2 value of 0.93 and 0.95 for each respective window size. Unlike the small differences in goodness of fit (+2% in R2), the sensitivity curves show a wider variation and suggest that bigger windows are more appropriate to capture smoother areas; this is expected. Although we get a better model with a higher spatial window size, the sensitivity of neighbouring points become more dispersed over larger areas over Europe instead of just staying within small clusters. A similar pattern of dispersion of the sensitive points is observed for the LST maps w.r.t. the spatial window size. For LST, we notice that there is not a large difference in the R2 as we increase the spatial window size. The most sensitive regions mostly stay the same but there is a small shift from the northern regions of Europe from more sensitive to less sensitive. So it’s clear that the number of spatial-pixels used as input features would be different depending upon the input variable, e.g. one can use a higher neighbourhood size for LST because we get the same R2 and similar sensitivity maps whereas the GPP could have a lower window size to ensure that we capture the local variability.

Fig 7

Visualizing the spatial maps for the senstivity of the Gaussian process (GP) regression model under different spatial sampling sizes for the Gross Primary Productivity (GPP) [top] and land surface temperature (LST) [bottom] for the summer of 2010 (Jun-Jul-Aug).

The rightmost column shows the summary R2 and Sensitivity for each spatial window size for the GP model.

7.3 Classification of drought regions

Support vector machines (SVMs) is a very common classification method widely used in numerous applications in the field of machine learning in general and remote sensing in particular [5]. The derivatives of the SVM function, however, have not been used before to understand the model, nor linked to the concept of margin. The derivatives of SVMs can be broken into two components via the product rule (section 4): 1) the derivative of the mask function and 2) the derivative of the kernel function. Typically one would use a sgn function as the mask but we used the tanh function to allow us to observe how the boundary or margin behaves w.r.t. the inputs.

In this experiment, we chose to study the relationship between gross primary productivity and root soil moisture during the year 2010 affected by a severe heatwave. There was a severe drought that occurred over this region for the entire summer of 2010 (June, July and August) [81]. Drought classification is an unsupervised problem and so there is a lot of debate about how to detect droughts within different scientific communities. We use a pre-defined drought mask of the countries affected by the 2010 heatwave found from the EM-DAT database [83] which reports all drought events which follow at least one of the criteria: 10 or more people dead, 100 or more people affected, declared state of emergency, or a call for international assistance. The region where the droughts are reported is just over the region of Eastern Europe, as shown in the binary classification maps in Fig 8. We chose this pre-defined drought area to simplify the problem which would allow us to see if we can indeed classify a drought region spatially and then look at the derivatives. We did a simple binary classification problem over the spatial coordinates using the two input variables (GPP and RM). We sampled only from the month of July at different time intervals within the month to make the samples more varied as the GPP and RM can still fluctuate within a monthly span. This is an unbalanced dataset as there are more non-drought regions than drought regions in the spatial subsample. While there are numerous advanced methods to deal with imbalanced datasets, we only used the standard SVM as that complexity is out of the scope for this experiment. The ESDC is very dense so we used 500 randomly selected points for the drought region and 1,000 randomly selected points for the non-drought regions. The remaining points (∼3800) were considered for calculating test statistics, while the visualizations include all of the points for the dataset. We applied a standard cross-validated SVM classification algorithm with an RBF kernel function. For metrics, we used the standard precision, recall, F1-score and Support for the predictions of drought over non-drought. Table 4 shows the classification results compared to the labels of the trained SVM algorithm and Fig 8 shows the classification maps.

Fig 8

Visualizing the (a) labels, (b) predictions (b), and (c) the 2D representation space for the predictions.

This is the classification problem of drought (red) versus no-drought (green) with the support vectors (black) for the SVM formulation (section 4).

Table 4

This table summarizes classification results for the drought and non-drought regions over Eastern Europe using the SVM (Support Vector Machines, section 4) formulation.

Class Label	Precision	Recall	F1-Score	Support
Non-Drought Regions	0.90	0.91	0.91	28590
Drought Regions	0.69	0.67	0.68	8268
Accuracy			0.86	36858

Fig 9 shows the sensitivity spatial maps as well as the 2D latent space for the outputs of the SVM classification model. We show the full derivative and the mask and kernel product components. The mask derivative has high sensitivity values for almost all regions where the decision function is unsure about the classification region. We see that the highest yellow regions are near the Caspian Sea which is also the area where there is a lot of overlap between the classes. Recall that the mask used is based on the reported regions and not the actual GPP or SM values. So naturally the SVM algorithm is probably picking up the inconsistencies with the data given. Nevertheless, the kernel derivative indicates where there are regions of little data and in regions where there is significant overlap. Ultimately, the combination of the products represents a good balance between lack of data and the width of the margin between the two classes.

Fig 9

Visualizing the scatter plot of the drought (red) versus no-drought (green) and the support vectors (black) using the SVM section 4 classification algorithm.

We also display the sensitivity of the full derivative and its components: the mask function (tanh) and the kernel function (∂k* αy) based on the predictive mean of the SVM classification results.

7.4 Principal curves of the ESDC

In this experiment we analyze GPP spatial-temporal patterns for different seasons for the year 2010 using the principal curve (PC) framework in Section 4. Each sample consists of a vector with the variable value for a particular location and all the time dimensions in the season: (05-Jan to 05-May), (21-May to 08-Aug), and (17-Aug to 31-Dec) For each season we have around 28, 000 samples of size 1 × T. Fig 10 shows the results. For each data set we plot the mean GPP value of the season in each point. The location of the points that belong to the PC are plotted in green using the Dijkstra distance inside the curve (as in the toy examples in Fig 4). The points belonging to the PC can be interpreted as the landmarks of the whole dataset, similar to a centroid of a cluster. But in this example, they refer to the points on the probability ridge of the data manifold (i.e. similar to the points closer to the first eigenvector in PCA). These points could be used for multiple purposes, e.g. as a summary to analyze the behaviour of the whole manifold or used for a temporal analysis of their evolution. One one hand, the location of the points is quite independent of the mean values, so they give different, alternative information. On the other hand, the location depends on the time of the year represented.

Fig 10

Principal curves on the ESDC.

Each figure represents the results for GPP at different time periods during the 2010. In each image the mean value of the variable for each location is shown in colormap (minimum blue, maximum red), and the points that belong to the principal curves are represented in green. Different brightness of green has been computed using the Dijkstra distance over the curve dots.

Most of the GPP ‘representative’ points are scattered around the manifold which depends on the season. For instance during the colder season (Jan-May), the dots are concentrated in the middle and low latitudes. During this period, the dots in northern Germany have a similar temperature and GPP than in the North-West part of Europe. Therefore, there is no need to add extra landmarks in these regions. Points in Morocco represent the warmer part of the manifold and Balcans area and Turkey represent the central part of the manifold. During the warmest period (May-Aug) the distribution of the dots follow an opposite direction, Southern regions are weighted less while Northern regions have more representation. In the case of mild temperatures (Aug-Dec), more landmarks in different regions are needed.

7.5 Sensitivity analysis of kernel dependence measures

HSIC is a dependence measure which can show differences in the higher-order relations between random variables. The derivatives of HSIC w.r.t. the input features are related to the change of the dependence measure which summarizes the relevance of the input features in the dependence. Therefore, these derivative maps can be related to the sensitivity of the inputs.

In this experiment, we chose to study the relation between GPP and RM for Europe and Russia during the years 2008, 2009 and 2010. We apply the HSIC with a linear kernel and compute the sensitivity maps, which is an estimation of how much the dependency criterion changes. We take spatial segments of GPP and RM at each time stamp, T and compute the HSIC value for each T independently for Russia and Europe. We also computed the derivative of HSIC for the same T time stamps independently for Russian and Europe. We computed the modulus to summarize the impact of each dimension to act as a proxy for the total average sensitivity. The final step involved computing the expectation between the modulus of the derivative of HSIC between Russian and Europe. Europe acts as a proxy stable environment and Russia is the one we would like to compare to. We estimated the expected value for three time periods (before: 05-Jan, 20-May; during: 28-May, 01-Sep; after: 09-Sep, 30-Dec) for each year individually. Then we compared each of the values to see how the expectation changes between Europe and Russia for each period across the years. The expected value of the HSIC derivatives summarize the change of association between variables differently than the HSIC measure itself.

The experiment focuses on studying the coupling/association between RM and GPP during the Russian drought in 2010. The HSIC algorithm captures an increased difference in dependencies of GPP and RM for Russia relative to Europe in 2010 if we compare this relationship to the years 2008 and 2009, see Fig 11a. However, HSIC only captures instantaneous instances of dependencies and not how fast these changes occur. The derivatives of HSIC (Fig 11b) allow us to quantify and capture when these changes actually occur. The gradients of HSIC do not show obvious differences in magnitude or shape across years between Russia and Europe. By taking the expected value of specific time periods of interest (before-during-after drought), we can highlight the contrast in the dependency trends between different periods with respect to their previous years, both in terms of HSIC and HSIC derivatives. We observe in Fig 11c, a change the mean value of the difference in the derivative of HSIC in Fig 11d which reveals a noticeable change in the trend for the springtime and summertime of 2010 compared to 2008 and 2009.

Fig 11

Each figure represents different summaries of how HSIC can be used to capture the differences in dependencies between Europe and Russia for GPP and RSM.

(a) shows the HSIC value for Europe and Russia at each time stamp, (b) shows the derivative of HSIC for Europe and Russia at each time stamp, and the mean value for the difference in the (c) HSIC between Europe and Russia for different periods (Jan-May, Jun-Aug, and Sept-Dec), and (d) in the derivative of HSIC for the same periods.

8 Conclusions

The use of Kernel methods is very popular in pattern analysis and machine learning and have been widely adopted because of their performance in many applications. However, they are still considered black-box models as the feature map is not directly accessible and predictions are difficult to interpret. In this note, we took a modest step back to understand different kernel methods by exploiting partial derivatives of the learned function with respect to the inputs.

To recap, we have provided intuitive explanations for derivatives of each kernel method through illustrative toy examples, and also highlighted the links between each of the formulations with concise expressions to showcase the similarities. We show that 1) the derivatives of kernel regression models (such as GPs) allows one to do sensitivity analysis to find relevant input features, 2) the derivatives of kernel classification models (such as SVMs) also allows one to do sensitivity analysis and visualize the margin, 3) the derivatives of kernel density estimators (KDE) allows one to describe the ridge of the estimated multivariate densities, 4) the derivatives of kernel dependence measures (such as HSIC) allows one to visualize the magnitude and change of direction in the dependencies between two multivariate variables. We have also given proof-of-concept examples of how they can be used in challenging applications with spatial-temporal Earth datasets. In particular, 1) we show that we can express the spatial-temporal relationships as inputs to regression algorithms and evaluate their relevance for prediction of essential climate variables, 2) we show that we can assess the margin for classification models in drought detection, as a way to identify the most sensitive points/regions for detection, 3) we show that the ridges can be used as indicators of potential regions of interest due to their location in the PDF, which could be related to anomalies, and 4) we show that we can detect changes in dependence between two events during an extreme heatwave event.

A Higher order derivatives of kernel functions

It can be shown that the m-th derivative of some kernel functions can be computed recursively using Faà di Bruno’s identity [40] for the multivariate case: where the sum is over all m-tuples and . It is also useful the expression for mixed derivatives: where Π is the ensemble all the partitions sets in 1…m, π is a particular partition set, B ∈ π runs over the blocks of the partition set π, and |π| is the cardinality of π.

For the RBF kernel we can identify f = exp(⋅) and g = −γ∥ − ∥2. The derivatives for the f(g()) are always the same ∂ f/∂g() = f(g()) = exp (g()), and the derivatives for the g() are: ∂g/∂x = −2γ(x − y), ∂2 g/∂x2 = −2γ, ∂ g/∂x = 0, for m ≥ 3, and .

Applying the previous formula for m = 1 the first derivative is:

The second derivative is:

The mixed derivative is:

B Custom regression function

In this example we show the behaviour of the first and second derivatives for a multivariate input. A GP model is fitted over the dataset using the RBF kernel function. The experiment uses a custom linear multivariate function with two inputs, x1 and x2, as inputs: where the coefficients a and b have varying values. Both x1,2 were generated along the same range uniform distribution but there was a linear transformation a = 5, b = 1 from ([0, 20]) and constant everywhere else, i.e. a = b = 1 from ([−20, 0]).

The GP model smooths the piece-wise continuous function which results in some additional slopes than the original formulation. This is visible (see Fig 12) from the derivatives of the kernel model as the first derivative for the x1 and x2 components have positive values for the sensitivities of the slopes in the regions where a and b are equal to some constant, respectively. The second derivative for both x1 and x2 show the same effect except for curvature. This experiment successfully highlights the derivatives of the individual components as well as their combined sensitivity.

Fig 12

First row: The original toy data is displayed as well as the predicted GP model which presents a smoother curve. Second row: the first derivative in the x1,x2 direction and combined direction (the sensitivity) respectively. Third row: the second derivative in the x1, x2 direction and combined direction (the sensitivity) respectively. The yellow colored points represent the regions with positive values, the blue colored points represent the regions with negative values and the gray colored points represent the regions where the values are zero.

10 Jul 2019

Submitted filename: sakame_rebuttal_letter.pdf

Click here for additional data file.

7 Feb 2020

PONE-D-19-17584

Kernel Methods and their Derivatives: Concept and Perspectives for the Earth System Sciences

PLOS ONE

Dear Mr Johnson,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

==============================

ACADEMIC EDITOR:

Please revise the manuscript follow the reviewers' comments

Please revise the manuscript in terms of languages and correct the typos

==============================

We would appreciate receiving your revised manuscript by Mar 15 2020 11:59PM. When you are ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter.

To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'.

Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.

We look forward to receiving your revised manuscript.

Kind regards,

Qichun 'Kit' Zhang, PhD

University of Bradford, UK

Academic Editor, PLOS ONE

Journal Requirements:

1. When submitting your revision, we need you to address these additional requirements.

Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

http://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Please explain the rationale for the development of your approach in light of recent research in this area, clearly indicating which problem with existing methods you are addressing.

3. Please clearly report at the beginning of your methods or results section which the key performance measures were to establish validity and utility of your method. Please also report clearly which statistical analysis was used to establish robustness of performance measures.

4. Please note that PLOS ONE requires that experiments, statistics, and other analyses must be performed to a high technical standard and described in sufficient detail to allow for reproducibility of the study (http://journals.plos.org/plosone/s/criteria-for-publication#loc-3). To demonstrate the performance of the method, we would expect comparisons to be drawn between existing state-of-the-art methods.

Additional Editor Comments:

This paper has been well-written with interesting topic. The reviewers return some critical comments, especially the novelty of the contribution as the kernel methods have been widely used. Moreover, some typos have to be corrected in the revised manuscript. Thus I believe that a major revision is necessary to improve the quality of the manuscript.

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: I have read the paper, “Kernel Methods and their derivatives: Concept and

perspectives for the Earth system sciences”. The paper is well organized, well written and of high academic quality. I am happy to see that the authors have made a really good effort to apply the kernel methods and their derivatives for the Earth system sciences.

The majority of the work is very well, but there are some minor issues that I would recommend to address in the final version of the paper. These issues are provided as follows;

(1) The citation of Figure is not consistent. Cite the figure as (Fig. xx) or (Figure xx).

(2) At line # 216 citation is missing

(3) At line # 61 what mean of m-th order? As it is not defined before.

(4) At line # 83 change the text, Section 7.4 with Section 6.

(5) The detail is missing in conclusion section, What was done in this paper related to Earth system sciences?

Reviewer #2: Manuscript Draft: PONE-D-19-17584 - Kernel Methods and their Derivatives: Concept and Perspectives for the Earth System Sciences

I. Summary

This article surveys the application of kernel methods to several supervised and unsupervised machine learning tasks, including Gaussian processes for regression, SVMs for classification, density estimation, and dependence estimation, with specific focus on showing the potentials of using the derivatives of kernel functions to gain insight of the function learned and to design kernel machines. This study claims the following contributions:

• To show a way to interpret the functions learned by various kernel methods by using their derivatives;

• To provide the explicit analytic form of the 1st and 2nd derivatives of some kernel functions and generic formulas to compute arbitrary order derivatives;

• To provide illustrative toy examples as well as a real-world problem with earth system data to demonstrate the usefulness of the derivatives.

While the topic of the article is interesting and the paper is well organized, I don’t think the current version of the article is ready for publication because the paper provides neither a synergistic framework for better understanding the kernel methods nor sufficient evidences to demonstrate their effectiveness in earth system science, failing to well justify the claimed contribution. I provide more specific comments in the following section.

II. Major issues

1. As the derivations of kernel functions’ derivatives are not new, one key theoretical contribution of this paper is to give more insights into the supervised and unsupervised kernel learning methods by relating the derivatives to the associated characteristic information, such as margin, sensitivity, and leverage. However, the new development of the present article is not clear to me compared to the previous study, i.e. [28,29,51,30], instead of summarizing the findings in these papers and running new toy examples. To make the article clearer, can the authors provide more details and discussion about the differences compared to the above mentioned study?

2. The definition 1 and Theorem 1 in Page 5 seem not relevant to the main material. Are they serve any purpose?

3. Following last comment, the kernel classification in Section 4 is not used in the earth system problem.

4. Section 7 looks interesting. But it is too short without sufficient evidences to demonstrate the effectiveness of kernel methods. Besides, the readers may want to know the background of spatio-temporal earth system data and the state-of-art of kernel methods in this domain, but this information is missing in the article.

5. What does the color bar in Figure 7 mean? Also, the font size of the numeric values in the plots is too small.

6. What are the values of a and b to generate the data in Figure 10? There are typos in the caption.

III. Minor issues

1. In Abstract, the phrase “…various kernel methods in a much more intuitively that commonly assumed…” is strange and should be reworded.

2. In the last paragraph of Introduction, the sentence “Section 7.4 pays attention to …” seems to have a wrong reference to the section number. Please double check.

3. Please double check the symbol in Eq. (9)

4. The reference is missing in Section 3.1.

5. In the caption of Figure 2, it seems the regularized and unregularized cases are wrongly referred.

6. A symbol error in Line 260, Page 11

7. In the caption of Figure 3, the phrase “The predicted has red and green…” has typos. Typos are also found in Figure 10. Please double check the spelling and grammar in all the captions to avoid unnecessary confusion to the readers.

8. What are the u, v in Line 494, Page 19.

There are more typos in the paper, I encourage the authors check the writing carefully in the revised version.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at figures@plos.org. Please note that Supporting Information files do not need this step.

6 Apr 2020

We have attached a response letter as a separate PDF. Copy Pasted the response below.

Reviewer I

Summary

I have read the paper, “Kernel Methods and their derivatives: Concept and perspectives for the Earth system sciences”. The paper is well organized, well written and of high academic quality. I am happy to see that the authors have made a really good effort to apply the kernel methods and their derivatives for the Earth system sciences. The majority of the work is very well, but there are some minor issues that I would recommend to address in the final version of the paper.

Thank you for the positive comments. To the authors’ knowledge, this is the first time that the derivatives of kernel methods have been showcased and summarized in such a detailed matter, collectively, and for the most widely used kernel machines and learning paradigms. In addition, we have applied these methods to toy examples, which highlight understanding, and to example applications in the very relevant field of Earth sciences. We have performed a thorough revision of the language and corrected the typos as suggested. We feel that with the additional changes, the manuscript has improved and is ready for your consideration.

Comments

Minor Issues

The citation of Figure is not consistent. Cite the figure as (Fig. xx) or (Figure xx).

At line # 61 what mean of m-th order? As it is not defined before.

At line # 83 change the text, Section 7.4 with Section 6

At line # 216 citation is missing

All of the minor issues and typos have been corrected.

The detail is missing in conclusion section, What was done in this paper related to Earth system sciences?

Following the reviewers’ suggestions, we have updated the conclusion section by emphasizing our contributions to the field of Earth sciences too. We have added the following paragraphs:

“To recap, we have provided intuitive explanations for derivatives of each kernel method through illustrative toy examples, and also highlighted the links between each of the formulations with concise expressions to showcase the similarities. We show that 1) the derivatives of kernel regression models (such as GPs) allows one to do sensitivity analysis to find relevant input features, 2) the derivatives of kernel classification models (such as SVMs) also allows one to do sensitivity analysis and relate it to the margin, 3) the derivatives of kernel density estimators (KDE) allows one to describe the ridge of the estimated multivariate densities, 4) the derivatives of kernel dependence measures (such as HSIC) allows one to visualize the magnitude and change of direction in the dependencies between two multivariate datasets. We have also given proof-of-concept examples of how they can be used in challenging applications with spatial-temporal Earth datasets. In particular, we show that 1) we can express the spatial-temporal relationships as inputs to regression algorithms and evaluate their relevance for prediction of essential climate variables, 2) we can assess the margin for classification models in drought detection, as a way to identify the most sensitive points/regions for detection, 3) the ridges can be used as indicators of potential regions of interest due to their location in the PDF, which could be related to anomalies, and 4) we can detect changes in dependence between two events during an extreme heatwave event.”

Reviewer II

Summary

This article surveys the application of kernel methods to several supervised and unsupervised machine learning tasks, including Gaussian processes for regression, SVMs for classification, density estimation, and dependence estimation, with specific focus on showing the potentials of using the derivatives of kernel functions to gain insight of the function learned and to design kernel machines. This study claims the following contributions:

• To show a way to interpret the functions learned by various kernel methods by using their derivatives;

• To provide the explicit analytic form of the 1st and 2nd derivatives of some kernel functions and generic formulas to compute arbitrary order derivatives;

• To provide illustrative toy examples as well as a real-world problem with earth system data to demonstrate the usefulness of the derivatives.

While the topic of the article is interesting and the paper is well organized, I don’t think the current version of the article is ready for publication because the paper provides neither a synergistic framework for better understanding the kernel methods nor sufficient evidences to demonstrate their effectiveness in earth system science, failing to well justify the claimed contribution. There are more typos in the paper. I encourage the authors’ check the writing carefully in the revised version.

Thank you for the comments. We want to clarify the following:

To authors’ knowledge, this is the first time the derivatives of kernel methods have been showcased and summarized in such a detailed matter: there have been no collective studies for the most widely used kernel machines or learning paradigms with illustrative toy examples. We have included four key learning paradigms in this study: classification, regression, density and dependence estimation. We have highlighted this contribution more clearly, and outlined the relations of derivatives of kernel methods with interesting features of the learned functions. For instance, we showed that the derivatives of kernel regression are useful for sensitivity analysis, derivatives of the SVM classification function are related to the margin, derivatives can be used to describe the ridge in densities, while derivatives of kernel dependence measures allow visualization and optimization of the measure.

Another novelty of our work is that we show the potential of derivatives of kernel methods in real, challenging problems. We are unaware of any collective study showing examples with real-world data, and in our case, it complements the toy examples. We demonstrated the use of the derivative in several applications involving Earth data cubes: how one can gain knowledge about the spatio-temporal relations in estimation settings, how to interpret classifiers trained for detecting extreme events, how to estimate the “interestingness” of spatial regions by looking at the joint PDF estimation of temperature and moisture data cubes, and how to detect changes in dependence by characterizing the derivatives of kernel dependence measures between Earth variables. Implications in the field of Earth sciences are obvious.

Following the reviewer's comment, we have now included a new experiment on the use of derivatives of SVM classification of the Russian heatwave to complete the 1-to-1 correspondence between the toy examples and the applications. Thanks for the suggestion.

In addition, and following the reviewer's suggestions, we have performed a thorough revision of the language use and corrected typos.

We feel that, with the included changes, the manuscript has improved substantially and is ready for your consideration.

Major Issues

As the derivations of kernel functions’ derivatives are not new, one key theoretical contribution of this paper is to give more insights into the supervised and unsupervised kernel learning methods by relating the derivatives to the associated characteristic information, such as margin, sensitivity, and leverage. However, the new development of the present article is not clear to me compared to the previous study, i.e. [28,29,51,30], instead of summarizing the findings in these papers and running new toy examples. To make the article clearer, can the authors provide more details and discussion about the differences compared to the above mentioned study?

Thank you for raising this point. We partially agree with the reviewer: while the kernel derivatives themselves are not new, their properties and their use in general kernel methods have not been reported collectively and extensively so far in the literature. We present the first paper that exclusively focuses on kernel derivatives in general from theory to practice, and shown both in simulated and complex settings. In each learning paradigm, we have identified scattered concepts in the literature: derivatives of kernel methods connect dots between distant concepts such as leverage scores, classifier margin, sensitivity analysis or densities ridges/principal curves. Many of these concepts were never treated collectively under the same umbrella, and we have expressed all of these aspects in our proposed paper.

The paper is not a mere collection of formulations; it is of practical interest with important implications in the Earth sciences. This is why we have not only provided toy illustrative examples, but also relevant applications on spatio-temporal Earth data. The latter is aimed to inspire applications in the same domain and other domains with similar data characteristics, e.g. signal processing (reference 1) and remote sensing (reference 2). We have highlighted all of these contributions in the introduction and conclusions sections, as well as included a summarizing table of the formulations and intuitions under the kernel derivatives for all methods {sensitivity, margin, ridge, leverage} at the end of the methods section.

References:

Rojo‐Álvarez et. al. - Digital Signal Processing with Kernel Methods - 2018

Camps-Valls & Bruzzone - Kernel Methods for Remote Sensing Data Analysis - 2009

The definition 1 and Theorem 1 in Page 5 seem not relevant to the main material. Are they serve any purpose?

This is a good point and we agree that this background kernel literature would be in principle unnecessary in a standard kernel methods paper. Nevertheless, we included it merely to build on the properties of kernel methods that allow to express the derivatives of the implemented function f(x) as a linear expansion of derivatives of the kernel function K(x,x'). This point, we feel, is strictly necessary for the general reader, not for experts in kernels (note that the paper will be of interest to Earth scientists too). However, we realize that it is not for all readers, so we have included a sentence at the beginning of the paragraph highlighting the main purpose of the section, and also concluded the paragraph with specific references PDF and PDF for any interested reader who would like to go further into the kernel literature.

Following last comment, the kernel classification in Section 4 is not used in the earth system problem.

Thank you for pointing this out. In the new version of the manuscript we have included a real classification experiment focused on anomaly (heatwave) classification, for which we show the usefulness of derivatives and their relation to the margin. We agree that this completes the paper nicely so we have now 1-to-1 correspondence between toy examples and ESDC applications. Thanks for the suggestion.

Section 7 looks interesting. But it is too short without sufficient evidences to demonstrate the effectiveness of kernel methods. Besides, the readers may want to know the background of spatio-temporal earth system data and the state-of-art of kernel methods in this domain, but this information is missing in the article.

Thank you for your comment. This section was included to illustrate the utility of the derivatives of kernel methods in real scenarios. The section does not pretend to be a review of kernel methods in Earth sciences, see this book for an overview of the field. Broadly speaking, kernel machines have been used in many problems in remote sensing and geosciences; from estimation of biophysical parameters to classification and anomaly detection, as well as to cluster regions and events, and to estimate dependencies between variables. However, the use of the derivatives of kernel machines are restricted to few examples, and we feel they can (and should) be exploited further, especially in spatio-temporal Earth data cubes, which is a missing part at the moment in the literature.

The selection of the spatio-temporal data for this section gives a very good opportunity to evaluate our proposal in a unique and challenging setting. Note that very few kernel methods have been used in such settings (restricted to new forms of KPCA and multi-temporal change detection with SVMs), and that the derivatives have never been exemplified before. This is why we decided to use this opportunity to do a proof-of-concept with our overview paper, and this section is essentially aimed at illustrating the use of derivatives on real, multivariate, nonlinear spatio-temporal Earth data; a challenging and important problem setting. Examples in this dataset are also illustrative and could motivate the use and adoption in other domains and/or similar problems sharing similar data characteristics, such as in neuroscience, signal processing and communications or remote sensing.

Following the reviewer’s suggestions, we have added several changes. First, we have included more contextualization about the use of kernel methods for Earth sciences at the introduction, some key references in the field, and motivated the use of kernel methods for spatio-temporal multivariate data in general and the Earth sciences in particular. We also added the new section on derivatives of kernel methods in anomaly detection in this setting using SVMs, and relating them to the concept of margin. Finally, in the conclusions section, we have included more discussion about the implications in the geosciences, as well as hypothesized on the potential use in other applications and/or domains in science and engineering.

What does the color bar in Figure 7 mean? Also, the font size of the numeric values in the plots is too small.

The colorbar represents the sensitivity; we have indicated now in the new version of the manuscript. The range is colored appropriately per variable. The font size was indeed too small; we have increased it for better visualization. Thanks.

What are the values of a and b to generate the data in Figure 10? There are typos in the caption.

Thanks for pointing out these minor issues. We have corrected the typos and added some description of the generated data in the very same caption. Thanks.

Minor Issues

In Abstract, the phrase “…various kernel methods in a much more intuitively that commonly assumed…” is strange and should be reworded.

In the last paragraph of Introduction, the sentence “Section 7.4 pays attention to …” seems to have a wrong reference to the section number. Please double check.

Please double check the symbol in Eq. (9)

The reference is missing in Section 3.1.

In the caption of Figure 2, it seems the regularized and unregularized cases are wrongly referred.

A symbol error in Line 260, Page 11

In the caption of Figure 3, the phrase “The predicted has red and green…” has typos. Typos are also found in Figure 10. Please double check the spelling and grammar in all the captions to avoid unnecessary confusion to the readers.

What are the u, v in Line 494, Page 19.

Thanks for the thorough revision. We have corrected all the typos and revised the language usage and grammar exhaustively.

Submitted filename: SAKAME - 2020 - Editor Comments.pdf

Click here for additional data file.

25 Jun 2020

Kernel Methods and their Derivatives: Concept and Perspectives for the Earth System Sciences

PONE-D-19-17584R1

Dear Dr. Johnson,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Qichun 'Kit' Zhang, PhD

University of Bradford, UK

Academic Editor, PLOS ONE

Additional Editor Comments (optional):

Two reviewers have returned the reports which show that the revised manuscript has been improved strongly in terms of contribution and quality. Therefore, I recommend to accept this paper after correcting the typos.

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: (No Response)

Reviewer #2: The manuscript was well revised and Reviewer's comments have been well addressed.

Please note that there are still few typos about the inconsistent notations: e.g., in the last paragraph of Introduction, "section 8" is not capitalized.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

17 Jul 2020

PONE-D-19-17584R1

Kernel Methods and their Derivatives: Concept and Perspectives for the Earth System Sciences

Dear Dr. Johnson:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Qichun Zhang

Academic Editor

PLOS ONE