SkyNet - Classifier

SkyNet Classifier-

This project focuses on automatically classifying land cover types from satellite imagery using deep learning. A convolutional neural network based on transfer learning with ResNet50 was trained on the EuroSAT dataset, which contains labeled satellite images across 10 land-use categories.

The goal was not only to achieve high classification accuracy, but also to explore how such models behave in real-world scenarios where satellite images often contain mixed land-use regions rather than a single dominant class.

Methodology

Dataset: EuroSAT (27,000 labeled satellite images)
Model: ResNet50 (pretrained on ImageNet, fine-tuned for 10 classes)
Input Size: 224×224 pixels
Classes: 'AnnualCrop', 'Forest', 'HerbaceousVegetation', 'Highway', 'Industrial', 'Pasture', 'PermanentCrop', 'Residential', 'River', 'SeaLake'

To address real-world complexity, the system was extended beyond standard classification:

Patch-Based Prediction: Large satellite images are divided into smaller patches, each classified independently.
Spatial Visualization: Predictions are overlaid on the original image to generate a land-use map.
Uncertainty Handling: For ambiguous regions, the model displays the top-2 predicted classes with confidence scores.