A Comparative Study of 1D-Convolutional Neural Networks with Modified Possibilistic c-Mean Algorithm for Mapping Transplanted Paddy Fields Using Temporal Data

Abstract

With increasing availability of satellite data of high temporal resolution, a more robust classifier is needed which can exploit the temporal information along with the spectral information of the remote sensing images. Specific fuzzy-based and learning-based algorithms are two broad categories and have the potential to perform well in spectral–temporal domain. In the present study, for mapping paddy fields as a specific class two classification algorithms, viz. fuzzy-based modified possibilistic c-mean (MPCM) algorithm and learning-based 1D-convolutional neural networks (CNN), were tested using Sentinel-2A/2B temporal data. The overall accuracy for learning-based 1D-CNN and fuzzy-based MPCM classifiers was found to be 96% and 93%, respectively. The F-measure values were found to be 0.95 and 0.92 for 1D-CNN- and MPCM-based classifier, respectively. Thus, it can be inferred from this study that the 1D-CNN classifier performed better than the traditional fuzzy-based classifier and can handle heterogeneity within class.