AgriVision

The inspiration behind KrishiNet came from a deeply rooted challenge in Indian agriculture: while satellite-based crop identification has made great progress in large, uniform farms of the West, it struggles in India’s fragmented smallholder farming system. With over 100 million farmers relying on small plots, inaccurate crop monitoring means weaker food security, unfair crop insurance, and poor access to credit. This gap motivated us to design a solution specifically tailored for India.

While working on KrishiNet, we learned the importance of phenology—the seasonal growth cycle of crops. Instead of treating crop identification as just static image classification, we explored how temporal modeling could capture the real growth stages of crops. This shift in perspective opened up possibilities for more resilient and accurate models.

We built our system by combining Synthetic Aperture Radar (SAR) data (to overcome cloud interference) with multispectral optical data. Through an attention-based fusion mechanism and geospatial foundation models like IBM Prithvi and SatMAE, we trained a framework that adapts to India’s diverse crop cycles.

The journey was not without challenges.

Data preprocessing was complex: we had to carefully handle cloud masking, speckle noise in SAR, and align multi-temporal datasets.

Model generalization was another hurdle: models trained on one state often failed in another due to soil, climate, and practice variations.

Finally, balancing technical rigor with deployability in rural India pushed us to keep the solution lightweight yet scalable.

Through these struggles, we built a pipeline that not only classifies crops but also aims to empower farmers, policymakers, and researchers with transparent, district-level crop intelligence.

Built With

• data-sources:-sentinel-1-sar
• ml
• resourcesat-2/2a-agrifieldnet-india-dataset-preprocessing-tools:-google-earth-engine-rasterio-gdal-modeling-&-ai-frameworks:-pytorch-hugging-face-torchgeo-ibm-geospatial-foundation-models-(prithvi
• satmae)-visualization-&-deployment:-streamlit-/-dash-interactive-dashboards-for-crop-maps-core-techniques:-time-series-transformers-for-crop-growth-cycles-attention-based-sar-+-optical-fusion-lightweight
• scalable
• sentinel-2