Agricultural Area Classification of Small Region Satellite Image Using Convolutional Neural Network

Abstract

Convolutional Neural Network (CNN) is widely used to develop image classification models which are designed to recognize patterns within pixel images and CNN improves model efficiency from traditional neural network models. This paper aims to investigate several CNN structures of an image classification model on images which are agricultural areas of small polygons on transparent background. We experimented in two different directions: first, increasing the number of convolution, non-linearity and pooling layers from the basic CNN structure; and second, increasing the number of training iterations with newly randomized data as input on every iteration. From the experimental results, the latter modified structure with newly randomized data and multiple iterations shows significantly higher testing accuracies and lower loss values than those of the basic CNN structure. This structure model improved the basic CNN by relative accuracy improvements of 1.03 - 10.94%.