Classification of the Agricultural Crops Using Landsat-8 NDVI Parameters by Support Vector Machine

— Along with the data obtained from the developing remote sensing technologies, the use of machine learning techniques is widely employed in classification at a more effective and precise level. In this study, support vector machines (SVM) technique, one of the machine learning approaches, was utilized with the help of data obtained from satellite image, and it was aimed to classify agricultural products. Moreover, lentil and wheat products were employed for object detection, and Landsat-8 satellite was preferred as satellite imagery. In order to determine the plant indexes in the image, Landsat-8 image of the development period of agricultural products dated May 6, 2018 was used and 98 sample points were taken with the help of GPS on the pilot area. After that, the position of these points were transferred to Landsat-8 satellite image employing the QGIS program and NDVI values were calculated from these points, which corresponds to Landsat-8 NDVI image pixels. The obtained NDVI values were then utilized in the SVM as inputs. As a result, the accuracy of the overall system for crop classification on the pilot area was computed as 83.3%.

The most up-to-date data about the land ensured by the remote sensing technology, merged with the Geographic Information Systems (GIS) technology, ensures a great convenience to the conventional techniques because it reaches its target faster, cheaper with limited labor [2].
Remote sensing systems can be categorized into two kinds as active and passive sensors. Of these, passive sensors enroll solar radiation emitted or reflected from the surface of earth. Photo cameras, electro-optic sensors, antenna sensors and thermal IR sensors can be thought as samples of passive sensors. Active sensors employ the energy, which sent from the sensor. They ensure their own energy towards the object to be investigated. The reflected energy from this object is then recorded and detected via the sensors. Active sensors can be employed in order to view the surface at any time regardless of the season. An instance of active detection is the radar that transmits an electromagnetic wave signal towards the object as well as measures and detects the backscattered signal [1].
Object detection and classification through machine learning in satellite images has recently become widespread. Support Vector Machines (SVM), one of the machine learning methods, have been among the highest performing approaches in the studies for object detection. SVM is a concept associated with a series of supervised learning methods in statistics and computer science. It is used for pattern recognition, data analysis, classification and regression analysis [3][4]. In this study, SVM model was preferred for crop classification since its capability and performance.
Many studies have been conducted in the literature using remote sensing data and machine learning approaches for object classification and detection [5][6][7][8].
In this study, two different agricultural products (lentil and wheat) were classified using Landsat-8 data, one of the optical remote sensors, and the SVM technique, one of the machine learning approaches. Landsat-8 satellite image used in the study was obtained in May 2018, which corresponds to the development period of the products. Parcels belonging to the village of Bıçakçı in Batman province, Turkey were selected as the study area. GPS coordinate points of 98 sample points over these parcels were then transferred to the pre-processed image. Finally, thanks to calculating the NDVI values, which corresponds to these points and employing them as inputs of SVM technique, the agricultural products (wheat or lentil) were classified with a high performance.

A. Pilot Area and Data Collection
As the study area, the agricultural land where lentil and wheat agricultural products were cultivated, covering an area of 2x2 km 2 within the borders of Bıçakçı Village of Batman Center, which is under the control of Batman Agriculture Directorate, was selected. GPS data of 96 different sample points (40 lentil and 56 wheat points) determined within the boundaries of the selected agricultural pilot area were employed in this study. In Figure 1, the Google Earth view of this study area and the distribution of GPS data of 96 sample points in the pilot agricultural area are shown. In addition to the ground measurements, Landsat-8 data was utilized in order to generate NDVI parameters of the sample points.

B. Landsat-8 Image Preprocessing
The satellite image of the Landsat-8 OLI satellite used within the scope of this study was obtained from the website of the American Geological Service (USGS) with geometric and radiometric correction [9]. This image is in the Universal Transverse Mercator (UTM) coordinate system slice 35 and WGS84 datum. In the ENVI 5.0 software (Exelis Visual Information Solutions, Boulder, Colorado), band composition, band proportioning and image enhancement methods were applied to the data. Gaussian Distribution Method was used in the image enhancement process. In this method, the pixel values were spread over the 0-255 color range via histogram and the contrast values were increased. Band rationing is an image enhancement analysis method in which new pixel values are obtained by dividing the gray color value of a pixel in one band by the value of the same pixel in another band or by applying other mathematical operations [5,10]. In addition, by using this image enhancement method, the spectral differences between the bands are enriched and the effect of the terrain roughness on the images is reduced.

C. Feature Extraction
In the feature extraction stage, one of the most important vegetation indexes, Normalized Difference Vegetation Index (NDVI), was computed from Landsat-8 image in order to form feature vectors for each sample point.
With the help of vegetation indices such as NDVI, it is possible to observe the plant diversity and vegetation density on the earth with remote sensing systems. In NDVI method, visible and near infrared regions of the electromagnetic spectrum were used [10][11].
In this study, the NDVI image was obtained by applying arithmetic operations (Equation-1) to the near infrared and red bands obtained from the Landsat-8 image, and the NDVI values of the 96 sample points on this image were then calculated.
Here, NIR and RED express the reflection values of the near infrared and red (visible) bands, respectively.

III. METHOD
In our study, a machine learning approach (SVM) was preferred to classify agricultural crops (lentil and wheat) utilizing the NDVI data, which obtained from the Landsat-8 image. The detailed information about SVM classifier is given as below.

A. Support Vector Machine (SVM)
Support Vector Machines (SVM) is a concept associated with a range of supervised learning methods in statistics and computer science. It is used for pattern recognition and data analysis as well as classification and regression analysis. The mathematical models of SVM have emerged in two classes for the classification of linear data at the first stage. In later times, new studies were conducted on the classification of nonlinear multi-class data. SVM, which tries to predict the most suitable transfer function to distinguish two different classes from each other and creates a hyper plane to separate them optimally [3,[12][13] as shown in Figure 2. The aim of SVM is to ensure equal distance to both sample groups with the hyperplane. For the linear separability case, the differential hyperplane is defined in Equation-2.
Here x, b and w denote input feature vector, bias and weight vectors, respectively.

IV. RESULTS AND DISCUSSION
In this section, the pre-processed Landsat-8 image and the NDVI image obtained from Landsat-8 image bands were presented. In addition, the data obtained from the NDVI image were passed through the training-test phase in the SVM technique and the performance values of the classification result were tabulated.

A. Obtaining NDVI Image
In this part, images of red (visible) and near infrared bands (band 4, band 5) to employ for obtaining NDVI image from the Landsat-8 image, which derived in 06.05.2018 are shown in Figure 3. These images were calculated in accordance with the NDVI formula specified in Equation 1, thanks to the raster calculator in the QGIS program. As a result of these operations, the NDVI image in Figure 4 was obtained. The obtained NDVI image was then converted into a "single band pseudo color layer" and shown in Figure 5. In addition, the import of GPS data to the image was carried out as shown in Figure 6.

B. The Results of Crop Classification by Employing NDVI Parameters and SVM
At this stage, leave-one-out cross validation method was applied to the data obtained to calculate the performance value of the whole system and the results obtained were presented in Table-1 using performance metrics (Sensitivity, Specificity, Precision, Recall, F1 Score and Accuracy). Moreover, the confusion matrix of the recommended system was given in Figure 7.     Agricultural systems constitute an important economic sector in the Anatolian countryside and worldwide. In particular, the combination of remote sensing data and machine learning techniques can facilitate the detection of many agricultural products at very large distances in a short time with little cost. In this study, it has been proposed to classify two different agricultural products (lentil and wheat) by employing NDVI parameters obtained from Landsat-8 satellite images and SVM approach from machine learning techniques.
The results of the proposed system indicated that a high classification success was performed in order to classify the crops thanks to NDVI parameters. Moreover, considering the success rate, it was observed that the machine learning approach employed in the proposed crop classification system affected the overall performance significantly.
In the future, as a continuation of this study, it is planned to classify various products with different vegetation indexes and different machine learning methods. It is thought that this proposed study will contribute to the national economy, especially in agricultural states.