Study of CNN deep learning model for temporal remote sensing data processing to map rabi crops

Abstract

Convolution Neural Network (CNN) is a deep learning approach that has become an area of interest to the researchers for solving complex problems. With the evaluation of CNN, extraction of deep features for accurate classification of remotely sensed images has gained lot of momentum. This research work uses CNN deep learning model for mapping rabi crops (mustard and wheat) using temporal remote sensing data. The mappings of mustard and wheat crops have been conducted using multispectral temporal images obtained from Sentinel 2A/2B between the dates 1st Nov 2019 and 24th Feb 2020 of Banasthali, Rajasthan region. The CNN model created in this research work uses several layers along with 5 activation functions (relu, sigmoid, tanh, elu and selu) for finding out which activation function gave the best result for the proposed study. Batch size has been examined from 1 to 50 in the multiple of 5 and epochs have been tested from 1 to 10 for a training data of 200 samples for each class. The optimal value with a batch size of 5 and epochs of 30 has been calculated as best suited in this study as the accuracy was getting constant. The implementation of CNN model for classification shows better results as compared to the traditional approach as the CNN algorithms are learning algorithms. This also helps in handling the heterogeneity within a class. A comparison has been conducted using Modified Possibilistic c-Means (MPCM) fuzzy algorithm for the classification of the same set of classes. F-Score, Kappa and Overall Accuracy have been calculated to show how the proposed approach has been outperformed and the level of classification accuracy achieved.