Project Title

Phenology-Aware Multi-Sensor Fusion for Crop Type Identification in Fragmented Indian Farms

Abstract

Crop identification using satellite observations is a long-standing challenge, particularly in India where farms are small, fragmented, and cropping practices highly diverse. Traditional approaches that perform well in homogeneous agricultural landscapes of the USA or Europe often fail to generalize in the Indian context.

To address this, we propose a multi-sensor, phenology-aware crop classification framework that leverages geospatial foundation models fine-tuned with Indian satellite observations and ground reference data.

Our approach fuses:

• Sentinel-1 SAR backscatter $(VV, VH)$ with coherence and texture features
  
• Sentinel-2 multispectral indices such as NDVI, EVI, SAVI, NDMI, and red-edge chlorophyll indicators $(NDRE, CI_{green})$
  
• ResourceSAT-2/2A and EOS-4 datasets for additional optical/SAR coverage
  

This combination provides complementary information: SAR captures canopy structure and moisture dynamics, while multispectral indices highlight vegetation health, water stress, and chlorophyll content.

Novelty of the Approach

1. Phenology-Aware Temporal Modeling

   • Extraction of crop growth cycle parameters: start of season (SOS), peak, and senescence (EOS).
     
   • For example, NDVI time series is smoothed and derivatives are used to estimate phenological markers.
     
   • Growth dynamics are modeled as:
     $$ \text{Phenology Features} = f\big( NDVI(t), EVI(t), VV(t), VH(t) \big) $$

2. Foundation Model Fine-Tuning

   • Adapting transformer-based geospatial foundation models using AgriFieldNet India dataset enhanced with ISRO/EPA SAR + optical data.
     
   • Cross-attention is applied between S1 and S2 sequences for multi-sensor fusion.

3. Uncertainty-Aware Classification

   • Alongside predictions, the model provides confidence maps using Monte Carlo Dropout:
     $$ \hat{y}, \ \sigma^2 = f_\theta(X), \quad \sigma^2 = \text{Var}[f_{\theta}^{(k)}(X)] $$
     
   • Critical for deployment in agricultural planning and food security systems.
     

Pipeline

1. Preprocessing

   • Cloud/shadow masking (S2), speckle filtering (S1), BRDF & topographic correction.
     
   • Temporal compositing (5–10 day) with smoothing (Savitzky–Golay).
     

2. Feature Engineering

   • Vegetation indices: $NDVI, EVI, SAVI, NDMI$
     
   • Chlorophyll indices: $NDRE, CI_{green}, REIP$
     
   • Water/stress indices: $NDWI, NDII, NBR$
     
   • Radar metrics: $VV, VH, VV/VH$, RVI, coherence, GLCM textures
     
   • Temporal phenology: SOS, POS, EOS, AUC, length of season
     

3. Modeling

   • Baselines: Random Forest, XGBoost
     
   • Sequence models: Bi-LSTM/GRU
     
   • Transformer-based geospatial foundation models for temporal + multi-sensor fusion
     

4. Evaluation

   • Generalization tested via leave-one-district-out (UP, Bihar, Odisha, Rajasthan).
     
   • Metrics: Macro-F1, per-class F1, overall accuracy.
     

Expected Impact

• Scalable, India-focused crop classification framework
  
• District-level crop maps with confidence estimates
  
• Open-source, reproducible pipeline for precision agriculture and food security monitoring
  
• Direct relevance to India’s fragmented smallholder farming systems
  

Built With

• api
• for-visualization-if-you-want-to-show-results-on-a-web-map)-frameworks-&-libraries:-tensorflow-/-pytorch-(deep-learning)-scikit-learn-(baseline-models)-rasterio
• gdal
• gee
• hub
• matplotlib
• model-training)-javascript-(optional
• notebook
• numpy
• pandas
• preprocessing
• python
• pytorch
• remote-sensing
• scikit-learn
• seaborn
• sentinel
• tensorflow