Kshetra

Inspiration

Indian farms are small, fragmented.Global models are usually trained on large farms in the US or Europe and fail to generalize here. We were inspired to build a crop classification system tuned for India’s complexity.

What it does

Kshetra uses satellite imagery and foundation models such as IBM’s Prithvi to reliably classify crop types in small fields, even under challenges like noise and class imbalance.

How we built it

We will fine-tune IBM’s Prithvi-300M backbone with Terratorch, adding a crop-classification head. To improve robustness, we use domain adapters and self-supervised pretraining on Indian satellite imagery.

Challenges we ran into

The dataset presents major hurdles: class imbalance, noisy labels, and very large images where preserving geospatial features during resizing is difficult. Handling multiple formats (SAR and multispectral) adds further complexity.

Accomplishments that we're proud of

We aim to design a pipeline that makes Prithvi truly farm-smart for India’s diverse farmlands, while identifying strategies that help it generalize across fragmented, heterogeneous regions.

What we learned

We expect to learn how to adapt large foundation models to small fragmented farms, address imbalance with robust losses, effectively combine SAR and multispectral imagery, and work with a new framework like Terratorch.

What's next for Kshetra

We plan to develop a lightweight version of Kshetra for real-world deployment. Beyond crop classification, we aim to add explainability for farmers and expand the system to provide insights on natural calamities, crop health, and broader agricultural monitoring.

Built With

• python
• terratorch