Satellite Image Classification

Transfer-learned ResNet18 trained on EuroSAT to classify satellite imagery into 10 land-cover types — for earth observation, UAV and disaster-response applications.

PyTorch ResNet18 Transfer Learning EuroSAT
View on GitHub
Satellite land-cover classification Satellite image classification sample predictions

Overview

This deep learning project trains a ResNet18 convolutional neural network, pretrained on ImageNet and fine-tuned on the EuroSAT dataset, to categorize satellite imagery across 10 land-cover types. It demonstrates practical applications in earth observation, UAV operations, disaster response and terrain analysis through an automated, reproducible pipeline.

Purpose — Why This Project Exists

Manually labeling land cover from satellite imagery does not scale to the volume of imagery captured daily — agriculture monitoring, urban planning and disaster response all need automated, fast, repeatable classification instead. This project shows how transfer learning turns a general-purpose ImageNet model into an accurate land-cover classifier with a fraction of the training data and compute a from-scratch model would need.

Animated Pipeline

1
Satellite Image
EuroSAT imagery loaded for training/inference.
2
Preprocessing
Resize, normalize, random crop and flip augmentation.
3
ResNet18 (Transfer)
ImageNet-pretrained backbone loaded.
4
Fine-Tuning
CrossEntropy loss + Adam optimizer on EuroSAT.
5
Evaluation
Confusion matrix, ROC curves, accuracy metrics.
6
Land-Cover Prediction
Real-time classification into 10 terrain classes.

Key Features

Transfer Learning — ImageNet-pretrained ResNet18 fine-tuned on EuroSAT.
Comprehensive Evaluation — confusion matrices, ROC curves, training visualizations.
Data Preprocessing — normalization, resizing, random crops and flips.
Real-Time Inference — with performance benchmarking.
Extensible Architecture — supports multispectral .tif imagery and NDVI analysis.
Reproducible Pipeline — modular train.py, evaluate.py, make_gif.py scripts.

Benefits

Rapid Mapping

Classifies land cover at a speed manual labeling can't match.

Cheaper to Train

Transfer learning cuts the data and compute needed versus training from scratch.

Extensible

Architecture supports multispectral .tif imagery and NDVI analysis.

Clearly Evaluated

Confusion matrix and ROC curves make model quality easy to verify.

Evaluation Metrics

Accuracy, Precision, Recall, F1-score, per-class ROC curves, confusion matrix, training/validation loss curves and inference-time benchmarking.

Technology Stack

Python PyTorch ResNet18 ImageNet Pretraining CrossEntropy Loss Adam Optimizer
Explore the Repository