Dummy Satellite Image Classification

This project serves as a “hello world” example in computer vision, focusing on a simple image classification task using a custom ResNet architecture to classify satellite images.

Problem domain: Multi-class image classification
Data type: Remote sensing satellite images

Source Code Dataset Source

Dataset

The data used in this project consist of low-resolution Remote Sensing (RS) images stored as .jpg files. Each image is associated with exactly one of the following four classes:

green area
cloudy
desert
water

The data can be freely downloaded from the original Kaggle Dataset.

Performance

              precision    recall  f1-score   support

           0     1.0000    0.9942    0.9971       172
           1     0.9909    1.0000    0.9954       109
           2     0.9930    0.9930    0.9930       143
           3     0.9929    0.9929    0.9929       140

    accuracy                         0.9947       564
   macro avg     0.9942    0.9950    0.9946       564
weighted avg     0.9947    0.9947    0.9947       564

Training Setup

Training aspect	Details
Model architecture	Custom ResNet (trained from scratch, no transfer learning)
Splits	Stratified Holdout (80%, 10%, 10%)
Epochs	40
Batch size	16
Optimizer	Adam
LR scheduler	OneCycleLR (max_lr = 0.1)
Gradient clipping	max_norm = 0.1
Loss function	CrossEntropyLoss
Data augmentation	Random Horizontal Flip

Dogs vs. Cats Image Classification Langchain Document Processor