Project Title

Cloud-Resilient Multi-Modal Crop Classification for Smallholder Agriculture in India

Abstract

Accurate, timely crop maps are essential for food security, yield forecasting, and resource allocation in India. However, persistent cloud cover during the monsoon season often renders optical satellite imagery ineffective. To address this challenge, we propose a deep-learning framework that fuses two complementary data streams: Synthetic Aperture Radar (SAR) polarizations (VH and VV) and multispectral vegetation indices (NDVI and EVI). By processing each source through a dedicated neural network branch and then combining their learned features, our model achieves robust field-level crop identification even under heavy cloud cover.

Temporal patterns across the growing season are captured with a transformer-based time-series module, which learns the distinct phenological signatures of rice, wheat, maize, and pulses. To handle geographic variation between states and smallholder landscapes, we incorporate a meta-learning adaptation strategy that allows rapid fine-tuning with only a handful of labeled samples from a new district.

Explainability is built in via attention heatmaps and SHAP-based feature attributions, enabling judges and stakeholders to see which time windows and data types drive each classification. Our containerized pipelines run on major cloud platforms and can be optimized for edge devices, ensuring scalable, transparent deployment for extension officers and policymakers. This solution delivers reliable, cloud-proof crop maps that empower decision-makers, support smallholder farmers, and strengthen India's agricultural resilience.