Skip to content

kindo-tk/EuroSAT-LandCover-Ensemble-CNN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 

Repository files navigation

An Efficient Light-Weight Ensemble of CNN For EuroSAT Land Cover Classification

Remote Sensing • Deep Learning • Land Cover Mapping


Overview

This project proposes a Light-Weight Ensemble CNN for accurate and computationally efficient land cover classification using the EuroSAT Sentinel-2 dataset.
The ensemble integrates ShuffleNet V2, MobileNet V2, and EfficientNet-B1, each adapted for 4-channel RGB+NIR input.

  • 98.00% accuracy
  • 90% reduction in model size compared to VGG16
  • Suitable for edge devices (drones, mobile/embedded systems)
  • Uses transfer learning, feature fusion, and meta-classification

All experiments were performed in Kaggle Notebooks.


Dataset Summary

The project uses the EuroSAT Land Use and Land Cover Classification Dataset based on Sentinel-2 multispectral imagery.

Property Value
Total images 27,000
Image Size 64×64 pixels
Channels 13 (we use RGB + NIR = 4 channels)
Classes 10

10 EuroSAT Land Cover Classes

Annual Crop, Forest, Herbaceous Vegetation, Highway, Industrial, Pasture, Permanent Crop, Residential, River, SeaLake


Methodology

  • Z-score normalization
  • Clipping to [-5, 5]
  • Resizing to 224×224
  • Augmentations: flips, rotations, jitter
  • 5-Fold Cross Validation
  • Hyperparameter tuning (20 trials × 6 epochs)

Model Architecture

Backbones (4‑channel modified):

  • ShuffleNet V2
  • MobileNet V2
  • EfficientNet-B1

Feature Fusion → 3584 dims
Meta-classifier:
3584 → 256 → Dropout → 128 → Dropout → 10


Results

Metric Score
Accuracy 98.00%
Precision (Macro Avg) 97.90%
Recall (Macro Avg) 98.00%
F1-Score (Macro Avg) 97.94%
AUC-ROC (OvR) 0.9990

Comparison With VGG16

Metric Ensemble VGG16
Accuracy (%) 98.00 97.20
Model Size 42.5 MB 512 MB
Params 10.94M 134M
Inference Time 0.023s 0.009s

Repository Structure

notebooks/
results/
data/

Reproducibility

  • The project was executed entirely in Kaggle Notebooks.
  • No local setup required.
  • Simply upload the notebooks, attach the EuroSAT dataset, and run all cells.

Acknowledgements

  • EuroSAT dataset authors for providing the land cover dataset.
  • Sentinel-2 mission by ESA for multispectral imagery.

Contributors

This work was carried out in collaboration with:

We thank all contributors for their support throughout the project.

About

A lightweight ensemble deep learning model achieving 98% accuracy on EuroSAT land-cover classification using 4-channel Sentinel-2 data.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors