A Comparative Study of Land Cover Mapping Based on Support Vector Machine

Hyperspectral and multispectral imagery has many applications, such as invasive species mapping or geological applications, such as mineral reconnaissance and mapping, lithologic mapping, mineral resource prospecting, mining environment monitoring, and oil and oil leak monitoring. gas,...etc. [1-3]. There are hundreds of other applications in agriculture, ecology, oceanography, and atmospheric studies where multispectral and hyperspectral remote sensing are used to understand the world around us better.

Classification of land cover images [4-6] is becoming a difficult task for many operational applications, due to number of spectral bands [7-9].

However, the availability of these maps within a reasonable time and with sufficient quality depends on the number of spectral bands and the number of training samples. These high dimensionality problems are addressed on one hand using feature reduction and extraction techniques and in another hand investigate an effective classifier [6, 10, 11] in terms of accuracy and computational time.

Support Vector Machine (SVM) [12] is suggested in this article to address the multiclass problem of remote sensing imaging. The robustness of this classifier is that it locates the optimal hyperplane between the class of interest and the rest of the classes by taking into account only the training samples that lie on the edge of the class distributions known as the support vectors.

The robustness of this classifier is also due to the use of kernel functions [13, 14], which made it more flexible and robust against outliers.

The purpose of this study is to classify hyperspectral images for mapping land occupation of Mohammadia [15, 16], mountainous region located in the west of Algeria, using Support Vector Machine (SVM) with three kernel functions: linear (LN), polynomial (PL) and radial basis function (RBF).

The region of El Mohammadia is rich in reliefs and natural wealth, such as mountains, rivers, vegetation, and forests. However, unfortunately, precise and recent ground reality mapping does not exist. The existing one is old, and its update by the concerned service is complicated and takes much time (sometimes months), which is why we thought of implementing an automatic mapping system based on multispectral images. The main advantages of this are speed, accuracy, and the possibility of updating at any time of the year to compare changes in nature and agglomerations.

Our principal contribution consists on one hand in the region of study (Mohamadia) that doesn’t have a land cover mapping from remote sensing image and on another hand in the proposed method based on the comparison of three SVM kernels to represent the land cover map closest to the ground truth.

The remainder of this paper is structured as follows: Section 2 describes the study image. Section 3 provides the principal steps of our proposed framework. Section 4 develops the experimental results and section 5 concludes the paper with discussions and future research directions.

2.png

Figure 2. Proposed system architecture

3.png

Figure 3. Image in recomposed color with three bands TM3, TM4, TM5 and Regions of Interest

where, Red=Channel 5 (TM5), Green=Channel 4 (TM4), Blue=Channel 3 (TM3).

3.2 Extraction of regions of interest

The choice of regions of interest (Figure 4) deserves a lot of attention. It is not always enough to choose only one game per class. Indeed, a class such as "vegetation" may contain several types of landscapes. In this case, we should create subclasses. The classes present in this image are (Figure 3): Class 1: water, Class 2: building, Class 3: Dense vegetation, Class 4: Sparse vegetation, Class 5: Forest Class 6: Bare land, Class 7: road.

4.png

Figure 4. Selection of interest regions

3.3 Training and tests

To recognize each area it is necessary to build a classification model. In this case, we used Wide Margin Separators (SVM) [19, 20]. Where the main objective of the SVM classifier is to obtain function f(x) which determines the optimal hyperplane that separates the different classes' zones. We obtain this hyperplane when it managed to separate two classes, or more, of input data points S [11].

1. The following steps summarize the phases of learning and construction of the classification model: Input a training set: S ={(x₁, y₁),…,(x_N, y_N)}

2. Choose a Kernel: k(․,․)

3. Training a SVM in the feature space

i.e.To find the decision function

$f(x)=\sum \alpha_i y_i k\left(x_i, x\right)$                    (1)

4. Classify any new object and test efficiency on the research of data.

There is, usually, no automatic way to choose a Kernel and to adjust the corresponding parameters. Therefore we, usually, have to try different Kernels and parameters. In our case, we have performed tests on the three following kernels:

1. Linear function [21]:

$k(x, x^{\prime})=b\left(x^T \cdot x^{\prime}\right)$                    (2)

2. RBF (Radial Basis Function) [22] kernel given by (Eq. (3)):

$K\left(x, x^{\prime}\right)=e^{-\frac{\|x-x^{\prime} \|^2}{2 \sigma^2}}$                    (3)

The parameter σ allows adjusting the width of the Gaussian. Taking a large σ, the similarity of one example to those around it will be quite high, while taking an σ tending to 0; the example will not be similar to any other.

3. Polynomial function [23, 24]:

$K(x, {x^{\prime}})=\left(\alpha x^T \cdot {x^{\prime}}+\lambda\right)^d$                    (4)

where, d is order of polynomial function; λ is Equilibrium parameter of the polynomial function.

In this study, we implement an automatic tool for producing the land cover map based on the classification of multispectral satellite images. To accomplish this task, we used the SVM algorithm known for its discrimination performances and faster learning compared to other classification methods. Obtaining the optimal SVM model required, from us, several combined tests on the rate of learning, classification, and the kernels used.

The choice of the best kernel was subject to two main criteria: the best classification rate and the lowest interclass confusion rate, so the choice of the polynomial kernel as the best to adopt for the SVM classifier is justified by its classification rate which is close to of 99% and its the percentage of max confusion obtained (4.17%) which is the lowest among the three kernels. A reduction in interclass confusion is possible by reinforcing learning on the classes subject to confusion, mainly the water/road and sparse vegetation/building classes.

The tests of our system were carried out on a multispectral image of the region of el Mohammedia (a small village located in the west of Algeria). A classification rate exceeding 98% was obtained with the polynomial kernel-based SVM model which is very satisfactory compared with the rates obtained by previous studies carried out on the same region.

A complementary study to this study is the classification of the image of the same region in 2021 and the mapping of the development of the urban fabric.