Comparison of Neural Networks for Binary Spatial Classification of Rice Field by Studying Temporal Pattern using Dual Polarimetric SAR Measurements

Abstract

Timely and precise information on rice cultivation plays a pivotal role in reshaping the global food and agricultural system. Synthetic aperture radar, with its capability to observe around the clock and in all weather conditions, is an invaluable tool for monitoring rice distribution. Such comprehensive cropland data at vast spatial scales not only enhances crop management but also provides critical support to governmental decision-making processes. The paper focuses on Binary classification by learning the temporal pattern of the Rice pixel. Time series curves of VV, VH, VV+VH, and VV/VH polarization and major rice varieties, MO4 and Kaje Jaya, cultivated in the area are analyzed to study the similarity of the curves the similarities in the curves, which could influence the temporal pattern recognition capacity of deep learning models. The study underscores the superior performance of RNN models, particularly BiLSTM and the proposed Dual Branch BiLSTM, over their CNN counterparts, such as 3DCNN and 3DUNET, especially for the VH and VV+VH polarizations. Specifically, the Dual Branch BiLSTM emerged as a standout, exhibiting an accuracy rate of 99.92% for combination of VH and VV+VH polarization. This model adeptly combined features from both VH and VV+VH polarizations, ensuring robust rice field mapping. Our results present a promising avenue for enhanced rice mapping, especially in tropical or subtropical zones, through the nuanced application of deep learning models. © Indian Society of Remote Sensing 2024.