Inspiration: Agriculture sustains millions of livelihoods in India, yet most farms are small, fragmented, and follow diverse cropping practices, making accurate crop monitoring a persistent challenge. Existing satellite-based approaches perform well in regions like the US or Europe with large, uniform farms but fail to generalize in the Indian context. This gap inspired us to build AgriVision AI, a solution that leverages geospatial foundation models fine-tuned with Indian satellite data to identify crop types more effectively. Our goal is to transform complex remote sensing data into actionable insights that can strengthen food security, promote sustainable farming, and support climate-resilient agricultural policies.
What it does: AgriVision AI uses a combination of SAR and multispectral satellite data from Sentinel-1, Sentinel-2, and ResourceSAT-2/2A to accurately classify crop types across Northern India. The system processes large-scale geospatial datasets, extracts vegetation indices like NDVI, and applies fine-tuned transformer-based models to capture crop growth patterns over time. By doing so, it provides field-level crop identification maps that can help policymakers monitor agricultural trends, assist researchers in analyzing cropping patterns, and enable farmers to make data-driven decisions. Ultimately, it transforms raw satellite observations into timely, actionable insights that support smarter farming and sustainable resource management.
How we built it: We built AgriVision AI by integrating SAR and multispectral observations from Sentinel-1, Sentinel-2, and ResourceSAT-2/2A with the enhanced AgriFieldNet India dataset. The data was preprocessed through normalization, cloud masking, and temporal harmonization to align spatial and time-series inputs. We then extracted vegetation indices such as NDVI and EVI to capture crop growth dynamics. A transformer-based geospatial foundation model was fine-tuned on Indian-specific data, leveraging temporal attention to learn seasonal patterns. The pipeline was implemented using Python, PyTorch, Hugging Face Transformers, and geospatial libraries like Rasterio and GDAL for satellite data processing.
Challenges we ran into: We faced challenges in handling huge volumes of multi-source satellite data, aligning SAR and optical imagery with different resolutions, and dealing with noisy inputs such as cloud cover and seasonal variability. Another major challenge was capturing the complexity of intercropping and irregular farm boundaries, which are common in India but not in datasets from other regions. Balancing model accuracy with computational efficiency was also critical due to resource constraints.
Accomplishments that we're proud of: We are proud of developing a working pipeline that can harmonize multiple satellite sources and fine-tune a transformer model for accurate crop classification in smallholder farm contexts. Our system not only achieved promising accuracy on validation data but also demonstrated scalability to diverse regions. Most importantly, we translated raw geospatial data into actionable agricultural insights that can benefit policymakers, researchers, and farmers alike.
What we learned: We learned the importance of data preprocessing and feature engineering when working with remote sensing data, especially in a diverse agricultural landscape like India. We gained hands-on experience in using geospatial libraries, satellite imagery processing, and transformer architectures for spatiotemporal learning. This project also taught us how to manage challenges like seasonal variation, missing data, and computational constraints, while reinforcing the value of collaboration in solving real-world problems.
What's next for AgriVision: Next, we plan to expand our dataset coverage to more Indian states and cropping systems, improve the model’s ability to handle intercropping, and optimize inference for real-time monitoring. We also aim to integrate weather data and soil information for more holistic predictions, and explore a farmer-facing platform that delivers insights directly to stakeholders in an accessible format. Ultimately, we envision AgriVision AI evolving into a scalable decision-support system for sustainable and climate-resilient agriculture.
Built With
- gdal
- google-colab-data-sources:-sentinel-1
- googlecolab
- hugging-face-transformers
- huggingfacetransformers
- jupyternotebook
- numpy
- pandas
- pandas-platforms-&-tools:-jupyter-notebook
- python
- pytorch
- rasterio
- resourcesat-2/2a
- sentinel-1
- sentinel-2
