Landcover Classification of Satellite Imagery Using Deep Learning

  • Tech Stack: Python (NumPy, Pandas, Seaborn), Deep Learning (Hybrid Spectral Network), GPU Acceleration, Apache Beam
  • Publication paper: Pdf

This project presents a novel approach to land cover classification of hyperspectral images using a Hybrid Spectral Convolutional 2D-3D Network (HybridSN), which effectively integrates both spatial and spectral information for accurate classification.

Key features of the project include:

  • Utilization of a 2D-3D hybrid CNN architecture to improve classification accuracy.
  • Effective handling of complex data from various hyperspectral image datasets including Indian Pines, Salinas, and Pavia University.
  • Application of Principal Component Analysis (PCA) for dimensionality reduction to enhance computational efficiency.
  • Robust testing of the model's scalability through generation of 10,000 user requests using Apache Beam.
  • Evaluation of classification performance using confusion matrices and accuracy metrics.

Experimental results demonstrated significant improvements in classification accuracy, with the HybridSN achieving accuracies of 98.37%, 99.37%, and 99.08% for the Indian Pines, Salinas, and Pavia University datasets, respectively. The model's architecture was validated against ground truth data to confirm its effectiveness.