FasalVision

Inspiration

Indiaʼs agriculture landscape is predominantly made up of small, fragmented farms with diverse cropping patterns. Existing crop classification models, mostly developed for large, homogeneous fields in the West, fail to accurately capture this complexity. Inspired by the desire to empower Indian farmers with timely, precise crop identification, this project leverages multispectral and SAR satellite data fused with ground-truth labels to build a model tailored to local conditions.

What it does

Our solution classifies crop types across fragmented smallholder farms in Northern India using multisource satellite data Sentinel-1 SAR, Sentinel-2 and ResourceSAT multispectral). It provides accurate, field-level crop maps covering multiple seasons and diverse crops, enabling better agricultural decision-making for farmers, policymakers, and extension services.

How we built it

We built the project by preprocessing the AgriFieldNet India satellite dataset—combining multispectral and SAR imagery—using advanced cloud masking, normalization, and temporal stacking. We fine-tuned geospatial foundation models (like UNet and SegFormer variants) on these features together with field-collected labels. Our model integrates multimodal fusion methods and temporal deep learning to capture crop phenology and spatial heterogeneity. Rigorous evaluation on unseen districts validated generalizable accuracy.

##Challenges we ran into Handling cloud cover and atmospheric noise in optical data, aligning different satellite datasets with varying resolutions, and accurately modeling fragmented farm plots were major challenges. Label imbalance due to rare crop types and fallow land required special addressing through data augmentation and ensemble methods. Ensuring model generalizability across different districts with varied cropping practices was also complex.

Accomplishments that we're proud of

The project successfully fused multispectral and radar data to improve classification accuracy for 13 crop classes. Our model demonstrated robustness when tested on unseen geographic regions and farming conditions. We developed an interpretable pipeline using explainability techniques to aid in stakeholder trust and understanding. The solution is built on open-source datasets and can be adapted for scalable agricultural monitoring in India.

What we learned

We deepened our understanding of multisource remote sensing data fusion and spatiotemporal modeling techniques. We learned that local ground truth data is indispensable for accurate crop classification in Indiaʼs complex farming landscape. Handling data inconsistencies and preprocessing were just as critical as the machine learning model itself. The value of explainability and stakeholder-centered design in agriculture tech became clear.

What's next for Untitled

Next, we aim to expand the model to cover additional Indian states and crop types, incorporate higher-resolution PlanetScope or Cartosat imagery for finer detail, and explore transfer learning to rapidly adapt models across regions and seasons. We also plan to develop a mobile-friendly interface for real-time crop classification accessible to extension workers and farmers on the ground.

Built With

• eos-4
• gdal
• github
• google-earth
• lstm
• python
• random-forest
• rasterio
• resourcesat-2/2a
• segformer
• sentinel-1
• sentinel-2
• u-net
• xgboost