Determination of vineyards with support vector machine and deep learning-based Image classification

Abstract

The study aims to determine the spatial distribution of vineyards with support vector machines (SVM) and convolutional neural network (CNN) based deep learning model. Multispectral (MS) and Panchromatic (PAN) bands of the high spatial resolution Worldview-2 (WV-2) satellite image were used for the study area located in Erzincan Üzümlü district. MS and PAN bands were fused to enhance the spatial resolution of the WV-2 multispectral image, making the vineyards more distinct and visible. Then, training samples were collected for five predetermined classes (vineyard, forest, soil, road and shadow) within the boundaries of the study area to generate training and test data, and the satellite image was classified using both Support Vector Machine (SVM) and CNN algorithms. Classification results were investigated using error matrices, kappa analyzes, and Mcnemar tests. As a result of the accuracy analysis, general classification accuracies and kappa values for CNN and SVM were obtained as 86.00% (0.8536) and 63.33% (0.6077), respectively. It has been observed that the CNN classifier provides higher classification accuracy (24% higher than the SVM). In addition, it was examined whether the differences between the McNemar test and the classification results were significant or not. As a result of the McNemar test for CNN and SVM, a value of 10.298 χ^2 was calculated. The fact that the calculated χ^2 value is greater than 3.84 reveals that the CNN classifier significantly increases the classification accuracy at the 95% confidence interval.