TerraCrop

Inspiration

India’s agriculture is unique due to its small, fragmented farms and diverse cropping patterns. Unlike the large, uniform farms in the USA or Europe, Indian farmlands create challenges for crop classification models. This inspired us to use satellite intelligence + AI to build a solution that works for India’s complexity.

What it does

TerraCrop uses multi-source satellite data (SAR + multispectral imagery) to automatically classify crop types across Northern India. It ingests Sentinel and ISRO satellite observations, processes them, and predicts which crop is grown in each agricultural field polygon. 1.Converts raw SAR + optical imagery into usable indices (NDVI, EVI, backscatter). 2.Applies an AI model fine-tuned on the AgriFieldNet India dataset. 3.Predicts crop types like wheat, rice, maize, and pulses. 4.Outputs field-level crop maps that can help in food security monitoring, agricultural planning, and policy-making.

How we built it

1.Dataset – Used hackathon-provided AgriFieldNet India dataset (Radiant Earth + IBM Research) with SAR + optical imagery. 2.Preprocessing-Temporal stacking across crop seasons. -Cloud masking + SAR calibration. 3.Model-Backbone: Geospatial foundation model pretrained on Earth Observation data. -Fine-tuned for multi-class crop classification. -Evaluated using accuracy and F1-score across states.

Challenges we ran into

Cloud cover in optical data → solved with SAR fusion. Small, fragmented farms in India → made generalization harder. Handling large-scale geospatial data → required optimized pipelines. Limited ground-truth labels → needed augmentation + regularization

Accomplishments that we're proud of

Built a pipeline that successfully integrates multi-source satellite data. Developed a field-level crop classification model tailored for India. Achieved good generalization across diverse states (UP, Rajasthan, Odisha, Bihar). Created a solution that can scale to real-world agricultural monitoring.

What we learned

How SAR and multispectral data complement each other for crop classification. Importance of temporal crop signatures (growth cycles) in classification. Techniques for preprocessing and handling big geospatial datasets. How to fine-tune foundation models for regional challenges.

What's next for TerraCrop

Extend to pan-India crop classification across all states. Improve accuracy using transformer-based models for time-series satellite data. Integrate weather + soil data for richer crop predictions. Build a user dashboard where farmers and policymakers can visualize crop maps. Explore use cases in yield prediction and climate impact assessment.

Built With

• cloud
• gdal
• gee
• ibm
• jupyter
• matplotlib
• numpy
• pandas
• python
• pytorch
• rasterio
• scikit-learn
• seaborn
• tensorflow