Field-Level Identification of TPR and DSR Using Sentinel-2 Imagery and Machine Learning Approaches in East India

Abstract

Accurately identifying how rice is planted is important for supporting sustainable agriculture and informed decisionmaking. In India, two main methods are used for growing rice: transplanted Rice (TPR), where seedlings are first grown in nurseries and then moved to fields, and direct seeded rice (DSR), where seeds are sowed directly in the field. While DSR has been promoted by the government and international agencies, there is limited understanding of its adoption at scale. Here we use Sentinel-2 satellite imagery and hundreds of ground data to train a random forest classifier to map DSR adoption at scale in Madhya Pradesh, India. We compared several different data processing steps: one with daily-resampled and smoothed time series data, and another using biweekly composite images. We find that our random forest model using the daily resampled dataset achieved an overall classification accuracy of 71%, while the model using the biweekly dataset reached an overall accuracy of 57%. Features from the shortwave infrared bands (B11 and B12), especially in October, contributed most to overall classification accuracy. Our results provide insights into how the adoption of improved rice management strategies can be mapped at the field scale (10 m) at the landscape level, even in heterogeneous smallholder systems.