Towards Live, Nation Wide, Farm-Level Crop Monitoring

Accurate and timely information about expected crop production is crucial for various applications including agricultural monitoring, policy making, and food security assessment. Policy makers can use near-real time crop maps to better determine crop support prices, storage infrastructure, and imports. In the context of India, absence of farm-level crop maps r the government to work with aggregate statistics based on manual surveys, and therefore are fundamentally limited in scale and accuracy. Surveys over large regions such as entire states or countries are slow and provide information only after large delays. Indian farms also go through up to three crop rotations a year necessitating continual monitoring. We put forward a nation-wide, farm-level, weekly agricultural monitoring and event detection model for the study area of India. Our model leverages remote sensing and machine learning to build a crop map that allows us to accurately monitor individual farms across large areas. 

We utilize the rich spectral and temporal information provided by Sentinel-2 satellite to provide near-real time crop monitoring, including sowing, crop type, and harvesting information. The predictions are done on an individual farm level with farm boundaries coming from a field segmentation model. Making predictions on a farm level scale helps getting more accurate yield estimates and allows monitoring individual fields for credit, insurance, resource allocation, etc. Currently, the model is able to identify major winter crops with an accuracy of up to 80% as early as 2 months after sowing. Equipped with the ability to provide weekly sowing and harvesting information makes the model near-real time for agricultural purposes. We also demonstrate the scalability of the model by showing results pan-India, across several diverse agro climatic zones. The model successfully generalizes to many unseen regions without requiring regional data. Using satellite data to provide accurate and timely crop cover information has the potential of saving millions of dollars spent by the government on manual surveys.