A STATISTICAL APPROACH TO CLASSIFY AGRICULTURAL SATELLITE IMAGES USING TEXTURAL FEATURES EXTRACTION

Increasing the resolution of recent satellite images has become one of the problems for image classification in scientific studies. One way to overcome this problem is to characterize the pixel being classified by parameters measuring the spatial organization of the pixels in its neighborhood. In this study, textural parameters of one-step and two-step encoded images were calculated in the form of a vector. These parameters were subject to the support vector machine classifier to decide which class image is belonging to. Results showed that the encoding method is a proper way to compress an image without losing textural information. It reduces the size of data and thereby reducing the computation time characteristics.


INTRODUCTION
With the advent of remote sensing observation, satellites regularly offer a large number of images, data and information inaccessible from the ground.The analysis and interpretation of the remote sensing images is an important field of researches and scientific studies such as deforestation monitoring, evolution of desertification, water resource development and construction and updating topographic maps [1].The classification of land cover from remote sensing imageries can be based on pixels or objects [2].
In a research, pixel-based thematic mapping accuracies produced using four different classification algorithms including Support Vector Machines (SVMs), Decision Trees (DTs), Artificial Neural Network Classifier (ANNC), and Maximum Likelihood Classifier (MLC) were compared.Results of this study showed that the accuracy of SVM-based classification generally outperformed the other three classification techniques [3].In another research, high spatial resolution airborne imagery was used and performance of a pixel-based classification based on MLC was compared with an object-based classification using K-Nearest Neighbors (K-NN).Although the 1-NN object-based classification had a better performance than the pixel-based MLC classification by 17%, average classification accuracy of pixel-based classification was 10% higher than objectbased classification for 48 vegetation classes listed in that paper [4].In a recent study, broad agricultural land cover types were classified for two time periods (1995 and 2005) using Landsat-5 Thematic Mapper (TM) imagery.It compared land cover maps produced using MLC and K-NN algorithms and it was found that the difference in overall accuracy between these classification approaches was not statistically significant [5].Another study used supervised machine learning algorithms including DTs, Random Forests (RFs), and SVMs to classify agricultural landscapes [2].It conducted a visual and statistical assessment of the classification outputs using medium spatial resolution (10 m) multi-spectral imagery from the SPOT-5 HRG sensor.For the purposes of this study, six broad land cover classes were mapped in a riparian area undergoing intensive agricultural development in western Canada.
There have been several image textural features extraction techniques that have emerged as useful methods to improve the classification accuracy of satellite imageries: image spatial co-occurrence matrix [6], local variance [7], variogram [8], fractal analysis [9], spatial autocorrelation [10], and wavelet transforms [11].Image texture analysis has become one of the applicable techniques to characterize the structural heterogeneity of classes.The texture of an image is related to the spatial distribution of the intensity values in the image, and as such contains information regarding contrast, uniformity, regularity, and so on [12].
In this study, a novel method for satellite images classification is proposed using textural features extraction.A new method for pre-treatment of satellite image textures is presented to improve the extraction of image features attributes.Supervised classification of images was done using textural properties as feature vectors.Section 2 is devoted to the presentation of a new pretreatment method textures to improve the extraction of characteristic attributes.Therefore, two encoding methods are introduced to show the importance of grey levels in the discrimination of textures.In the section 3, two textural feature extraction methods, namely the method of cooccurrence matrix and correlogram are introduced.Section 4 introduces the basic concepts of SVM in order to test the relevance of the proposed encoding.Section 5 presents the results of textures classification by using coded attributes on textures.This new approach characterizes the spatial relationships and relationships between the grey-scale pixels of an image.It has allowed us to utilize two encodings for characterizing the textures.After applying the method of coding, the co-occurrence matrix and correlogram of the images were extracted to calculate the features used for classification.Several nuclei of SVM classifier (linear, polynomial, Gaussian and sigmoid) were used to discuss the different results.

PRETREATMENT AND TEXTURE CODING
In this section, we proposed (a) the method of extracting texture in the satellite images and (b) development of the image encoding method.Most studies on the extraction of textural information act directly on the grey-scale image.In this article, a new method to encode grey-scale image is utilized.Therefore, the study will be made on the codes of grey-scale.One-step grey level encoding was utilized to encode neighboring pixels, because it is based on the use of a local neighborhood.A neighborhood V 8 , 8-neighbors of 3×3 pixels defined on an image texture was considered to observe either all pixels located just around the central pixel without favoring one direction rather than another or greater distance.Therefore, nine neighboring pixels were considered in ascending order of their grey level.Then, the processed pixel was assigned the highest rank corresponding to the grey level in the ordered sequence.To evaluate our approach, we used a set consisting of satellite images.Extracting textures attributes was achieved through the representations of encoded images.Thus, for all textures of the set, their encodings were calculated.Regarding the one-step coding, the pixels were coded 0 to 8, while the two-step was used to encode the pixels 0 to 15 (a neighborhood V 16 including 16-neighbors of 4×4 pixels).

TEXTURAL FEATURES MATRICES
Image texture analysis is a difficult problem due to the fact that there is no precise and rigorous definition to fully characterize the notion of texture [13].After one-step and two-step encoding the images, co-occurrence matrices and correlogram were used to extract various parameters characterizing the satellite images that are subsequently used in their interpretation and their classification.Because of their high texture information, cooccurrence matrices have become one of the best known methods to extract the characteristics of textures [14].They consider the properties of images relating to the second order statistics.Co-occurrence matrix measures probability of the occurrence of pixel pairs values located at certain distances in an image.It is based on the calculation of the probability P(i, j, , ) that represents the number of times a pixel grey level i appears at a relative distance of a pixel grey level j and depending given direction [15].The color coherence vector is a histogram of the refined representation in which colors are divided into two components, coherent and noncoherent [16].Coherent component comprises the number of neighboring pixels having the same color as the current pixel.The distance between the histograms of this type depends on the two components.Another more refined representation of the histogram is color correlogram which equals to the probability of occurrence of a color in a predetermined vicinity of a different color [17].The advantage of this method is that it takes good information account for spatial distribution of the grey level in the V 8 and V 16 neighborhoods of image pixels.To extract attributes characteristic for each texture, 4 parameters including homogeneity, contrast, dissimilarity and entropy moment were calculated using equations (1) to (4) [4].

CLASSIFICATION
The objective of the satellite image processing is to assign a label from a collection previously defined in order to achieve a classification of satellite images.There are two main types of image classification techniques: supervised and unsupervised.The latter does not require any prior knowledge of the user and characteristics of the classes are set automatically for classification.Supervised classification requires the user to instruct the system by designating areas of the image as representative samples of classes to extract.To instruct such this system, user must have a good knowledge of the observed field.Therefore, it is necessary to have a ground truth (all guarantees accurate data).It is a tool to achieve a good learning, and to validate classification.In this study, a supervised classification is performed.Classification methods aim to identify the classes that inputs belong to them from some descriptive traits.They can apply to many human activities and they are particularly suitable for the automated decision-making.The procedure for classification is extracted automatically from a set of examples.The procedure generated in this study was carried out to classify the satellite images into 10 agricultural land cover classes including: weed shrub, soil, grassland, cropland, riparian, water, rock, wetland and dune using SVM classifier.Learning phase of SVM consisted of 144 vectors per class or 1440 vehicles in total.The classification process is conducted as follows: the parameters of each texture were calculated in the form of a vector.These parameters were subject to the SVM classifier to decide which class texture is belonging to.

Dataset
As it is shown in Figure 1, 10 land cover textures were extracted in this study.They were obtained from 20 satellite images and represent urban areas, soils, forests, seas and roads (Figure 2).Texture objects are all composed of 256 grey levels 56×56 dimension.To make a classification, all textures were representative of each of the 10 classes and a set of tests were carried out to classify textures.Then, textures were encoded and before making rankings, attributes characteristic were extracted from each texture.For this, co-occurrence and correlogram were utilized.With these attributes, it was possible to compare the results among the classifications made from the original images and those made from encoded grey-scale.We used two encodings grey-scale on each of the 10 texture sets.

Performance of grey level encoding
The first experiment was to apply the encoding on the textures and we chose initially a number of tiles equal to 280 in total or 28 per class.Table 1 compares the results of texture classification obtained from one-step and two-step encoded with the original un-coded textures.The results are presented in this table as overall classification accuracy for each agricultural class.It can be seen clearly on the table that the two-step encoding method has allowed us to achieve good results using SVM classifier.We achieved these results using the kernel of default Gaussian.We could achieve better rates by testing several existing kernels (linear, polynomial and sigmoid).One-step encoding reduces the number of grey levels to 9 grey -scale.Although, the latter leads to higher grey-scale 16 in parallel and thereby improve the quality of the coding.

Performance of the encoded correlogram
To compare the results, the SVM classification was carried out on both the extracted features of the correlogram with quantification and those extracted from the correlogram with coding.The latter shows its effectiveness compared to the conventional method of quantification.Table 3 is much more indicative of the effectiveness and relevance of coding applied to the correlogram compared to the conventional method of quantification.For different nuclei (kernels) and values of C, the results indicated that the Gaussian kernel with C=1000 had the best performance and provided high levels of classification.The classification results base on coding were significantly higher than those obtained from the correlogram with quantification.The percentage of proposed classification method is approximately 81% while it is only 53% using quantization.

CONCLUSION
The investigative method developed in this article is coding grey-scale textures.A series of experiments was presented using two texture extraction methods namely co-occurrence matrices and correlogram methods.The classification results were significantly higher than those obtained from the correlogram without coding.The aim was to show the effectiveness of the correlogram with the new approach of coding.The SVM classification was performed on both the correlogram with quantification and co-occurrence matrix with coding, to prove the effectiveness of the latter relative to the other.The objective was to highlight the effectiveness of the cooccurrence matrix with the new coding approach compared with the correlogram.

Table 1 .
Overall accuracy of the classification by SVM (%).Another experiment was carried out to evaluate the performance of two-step encoding by varying the constant C and kernel functions (linear, polynomial, Gaussian and sigmoid) of the SVM classifier.The purpose of this experiment was to compare the kernels for one that gives a better classification rate.The results of classification of different types of classifiers built by SVM algorithm, selecting four types of kernels namely the linear kernel, polynomial, Gaussian and sigmoid, are presented in Table2.It is clear that for large values of C, high classification rate is obtained, which increases the computation time.Based on this table, it can be seen that the linear and Gaussian kernels have better performances.

Table 3 .
Overall classification rate of the different kernel classifiers based on the parameter C (%).