A novel machine learning based approach to generate dynamic monsoon crop mask for small-scale farms in tropical regions using a combination of radar and optical satellite data

Crop monitoring in tropical regions, especially during monsoon/wet season, is challenging due to high cloud cover during the growing season, small field size, and highly dynamic cropping patterns. Here we present a novel method, Radar Optical cross Masking (ROM), to generate high-resolution agriculture mask suitable for monsoon crops. The ROM is a two-step synergistic approach for combining Synthetic Aperture Radar (SAR) derived image metrics with temporal aggregation of Normalized Different Vegetation Index (NDVI) derived from optical data. In this work, a random forest (RF) classifier was applied for generating ROM resulting in a binary monsoon crop/non-crop map for 2018. For the present work, we utilized a Sentinel-1 SAR images of dual polarization (VH+VV) and Sentinel-2 images on Google Earth Engine (GEE) platform. The proposed method was tested and validated across eight different agro-ecological regions in India extending from north to south representing the major diversity in agriculture, soil and climatic variations. The classification accuracy of generating the monsoon crop map using ROM on the combined SAR+optical dataset was 95.7% whereas the accuracy obtained using only SAR data was 90.8 % respectively. The ROM-derived monsoon crop estimates correlated well with the crop statistics provided by the Government of India. The proposed method for generating monsoon crop map using agriculture mask (ROM) is particularly effective in regions with cropland mixed with plantation/mixed forest typical of small-scale farms in tropical regions. The method is also well suited to support the global agricultural monitoring missions such as GEOGLAM and S2Agri for constructing a dynamic cropland mask suitable for tropical regions.