Classification of Croplands Using Sentinel-2 Satellite Images and a Novel Deep 3D Convolutional Neural Network (Case Study: Shahrekord)

Agriculture has been recognized as the main motive for economic growth and development in different countries of the world. In the meantime, mapping croplands through the classification of remote sensing images is one of the effective solutions in decision making and providing food security to the community. In this research, croplands are classified into different classes of agricultural products (including wheat, barley, corn, alfalfa, potatoes, and Sugar beets) using multi-temporal optical (Sentinel-2) and synthetic aperture radar (Sentinel-1) satellite images. All the steps related to the preparation of satellite images, have been conducted in the Google Earth Engine online processing platform. A novel three-dimensional deep convolutional neural network is used as the classifier. The designed network, in addition to three-dimensional kernels with the ability to extract spatial and temporal information of each pixel simultaneously, uses some escape connections of the previous layers. These connections, contrary to the feed-forward convolutional networks, feed the output of the previous layers to the new layers. After dividing the ground truth data into two categories of training and evaluation and assessing the performance of the network with 50 different training and evaluation data, the network’s overall accuracy was calculated 91.6% on average. According to the final results, the designed escape connections increased the overall accuracy of classification by 2%. The proposed network was also compared with temporal and spatial-temporal Random Forests and Support Vector Machines which showed a better performance with a difference of at least 2.4%.