Satellite Image Classification
A custom ResNet50 model was designed to classify satellite images into Cloudy, Desert, Green Area, and Water categories, achieving high accuracy and effective performance on a small dataset.
Let’s Work Together for Development
Call us directly, submit a sample or email us!
The Story
Satellite imagery classification is essential for environmental monitoring, land use mapping, and climate studies. MozaicAI embarked on this project to develop a reliable model that could accurately classify satellite images into specific categories, including Cloudy, Desert, Green Area, and Water.
The aim was to create a model that maintained high accuracy even with a modest dataset, making it both effective and accessible for various applications. Here are some of the example images and corresponding categories:
What Did MozaicAI Do
MozaicAI fine-tuned a custom ResNet50 architecture initially trained on ImageNet, adapting it for this satellite image classification task. Working with a dataset of just 5,067 images (roughly 1,200 per category), we optimized the model’s performance through tailored training techniques and integrated TensorBoard to closely monitor and adjust model metrics. This approach allowed us to achieve high accuracy while using a relatively small, balanced dataset.
The Results
The model achieved an impressive 94.86% accuracy on the test dataset, demonstrating strong classification capabilities across all four categories. In-depth performance visualizations revealed key insights:
Shows rare misclassifications, primarily minor confusion between Green Area and Water.
Demonstrates robust classification performance across classes.
