This project focuses on crop classification by fusing Sentinel-1 SAR and Sentinel-2 optical satellite imagery, using Deep Learning and Machine Learning techniques. The objective is to accurately classify crop types based on multi-source remote sensing data.
Objective
To build a robust model that can classify different types of crops (e.g., Wheat 1, Wheat 2, Onion, Garlic) using fused satellite imagery and evaluate the performance of various models including CNN, ANN, RNN, and Random Forest.
Dataset Description
Satellite Data Sources: Sentinel-1 (SAR) and Sentinel-2 (Optical) images were obtained from the Copernicus Open Access Hub.
Study Area: Indore City, India (specifically the Mhow Road, Simrol region). Crop Shapefile: Provided labeled ground truth data for four crop types.
Workflow
Data Collection Downloaded Sentinel-1 and Sentinel-2 images. Obtained crop shapefiles for ground truth labels.
Preprocessing Used SNAP and QGIS for: Radiometric and geometric corrections Geocoding and co-registration Clipping and aligning datasets
Feature Extraction and Fusion Merged SAR and optical features. Created fused multi-band images for classification. Model Training Python libraries used: TensorFlow, Keras, Scikit-learn, Geopandas, Rasterio
Models Implemented:
Random Forest Convolutional Neural Network (CNN) Artificial Neural Network (ANN) Recurrent Neural Network (RNN)
Evaluation
Accuracy of each model was measured and compared. Random Forest achieved 95%, while CNN-based data fusion yielded the highest deep learning accuracy of 61%.
Results Model Accuracy Random Forest 95% CNN (Fusion Model) 61% ANN ~57% RNN ~54%
Project Structure graphql Copy Edit Crop-Classification/ │ ├── data/ # Satellite and shapefile data ├── preprocessing/ # SNAP and QGIS preprocessing scripts ├── models/ # Trained ML/DL models ├── notebooks/ # Jupyter notebooks for training/testing ├── results/ # Accuracy reports and plots ├── requirements.txt # Python dependencies └── README.md # Project documentation
Tools and Technologies Python, TensorFlow, Keras, Scikit-learn, Rasterio, Geopandas SNAP (Sentinel Application Platform), QGIS Jupyter Notebooks, Pandas, Matplotlib
Future Work Explore Transformer-based models on fused imagery Incorporate attention mechanisms to enhance spatial feature learning Extend classification to time-series analysis for monitoring crop cycles
Author Rishika Nigam B.Tech, Mechanical Engineering, Birla Institute of Technology, Mesra Email: rishikanigam4@gmail.com LinkedIn: Rishika Nigam