Semi-supervised Classification of Paddy Fields from Dual Polarized Synthetic Aperture Radar (SAR) images using Deep Learning

Abstract

ABSTRACT Rice is the major food grain for many countries in this world, in particular South Asian countries such as Pakistan, India, Bangladesh, Sri Lanka, Myanmar, Thailand, etc. India is one of the leading producers of rice, with agriculture contributing 14% to 15% to the Gross domestic product (GDP). Detection of paddy field requires constant monitoring of a region throughout the year. Hence, we use dual-polarized Sentinel-1 Synthetic Aperture Radar (SAR) data with 20 m resolutions for our study. In this paper, we propose a semi-supervised algorithm that detects paddy fields in various seasons. The technique is divided into two parts, unsupervised and supervised. Feature extraction layers of the Visual Geometry Group (VGG16) model are used to segregate patches into five clusters using an entropy-based loss function. Depending on the availability of the ground truth, data is sampled (stratified sampling), and only a certain amount of data is further used for the supervised section. Clusters showing higher similarity (Jaccard/Tanimoto test) with the paddy or nonpaddy classes are connected to a three-layer neural network used for the supervised method. This has been tested on two study sites using 10 data sets, five from 2018 to 2019, each. The highest accuracy reported for 2018 and 2019 is 85.61% and 84.92%, respectively. The two sites are selected from different districts of West Bengal, India. Our model has also shown higher generalization capability against the conventional-supervised models.