🌱 About the Project

🔹 Inspiration

Agriculture supports nearly half of India’s workforce, yet timely and accurate crop type information remains scarce.
Fragmented farms, diverse practices, and frequent cloud cover make traditional crop surveys unreliable.
We were inspired to leverage satellite data + AI foundation models to empower policymakers and farmers with better insights for crop monitoring, insurance, and food security.

🔹 What I Learned

• The importance of data fusion: SAR complements multispectral imagery by providing structural and soil moisture information even under cloud cover.
  
• Temporal modeling is crucial: crops are best identified by their phenological signatures across growth stages.
  
• Foundation models like SatMAE and Prithvi can generalize well when fine-tuned with Indian-specific datasets.
  
• The value of open datasets such as AgriFieldNet India and IBM-enhanced SAR/optical data.
  

🔹 How I will Build It

1. Data Preparation

   • Collected SAR (Sentinel-1/EOS-4) and optical (Sentinel-2/ResourceSAT-2/2A) satellite observations.
     
   • Preprocessed and clipped imagery to field polygons from AgriFieldNet India.
     
   • Extracted vegetation indices (NDVI, EVI, red-edge features) and radar backscatter metrics.
     

2. Model Development

   • Used a transformer-based geospatial foundation model pre-trained on global Earth observation datasets.
     
   • Fine-tuned with Indian crop labels across Uttar Pradesh, Bihar, Rajasthan, and Odisha.
     
   • Implemented temporal encoders to capture crop growth cycles.
     

3. Evaluation & Visualization

   • Metrics: Accuracy, F1-score, and IoU across crop types.
     
   • Developed a prototype dashboard to display color-coded crop maps with confidence scores.
     

🔹 Math Behind the Model

We model the crop classification task as a multi-class sequence classification problem:

$$ X = {x_1, x_2, \dots, x_T}, \quad x_t \in \mathbb{R}^d $$

where \(X\) is the time-series of fused SAR + multispectral features.
A transformer encoder learns representations:

$$ h_t = \text{TransformerEncoder}(x_t) $$

The final crop type prediction is obtained by:

$$ \hat{y} = \arg\max \; \text{Softmax}(W \cdot \text{Pool}(h_1, \dots, h_T) + b) $$

Built With

• agrifieldnet
• bhuvan
• gdal
• geopandas
• huggingface
• ibm
• isro
• leaflet.js
• matplotlib
• postgresql
• prithvi
• python
• pytorch
• rasterio
• resourcesat-2
• satmae
• scikit-learn
• seaborn
• sentinel
• streamlit
• transformers