Satellite Image Segmentation

Supervisor:

Built and trained an autoencoder for satellite image segmentation into color-coded land types, with self-tuned architecture and training.

Jan 2024 - Apr 2024

Segmentation ModelAutoencoder

Highlight Contributions

Optimized and iterated on model architectures, testing various techniques to improve accuracy.
Programmed and trained the model for effective performance.

Introduction

The Satellite Image Segmentation project developed and trained an autoencoder to automatically categorize and segment land covers from satellite imagery, producing color-coded outputs for applications ranging from land cover monitoring to natural disaster detection.

Interation of Model Structure

By iterating on the model structure (Figure 1), we achieved clear improvements in segmentation accuracy with a fully self-built model trained from scratch, without pretraining.

model improvements — Fig 1: Segmentation results through iteration of model structure.

We've setteled down in the final model architecture as illustrated in Figure 2, and their training accuracy and loss function can be found in the Figure 3 & 4.

model sturcture — Fig 2: Model architecture of final model.

Experimental setup — Fig 3: Final model accuracy.

Methodology flow — Fig 4: Final model loss.

Acknowledgements

This project was a team effort. Credit to my wonderful teammates - Yijie, Emma, and Kara!

Highlight Contributions

Introduction

Interation of Model Structure

Acknowledgements

Resources