Deep Learning Deforestation Detection

This repository contains all of the resources associated with my final project for the ST456 Deep Learning course at the London School of Economics, which I completed jointly with Rick Holubec and Chayenne Mosk. Our work focuses on identifying and tracking deforestation in the Amazon Rainforest using state-of-the-art deep learning models and multispectral satellite imagery.

Structure

The repository is organized as follows:

• data: Contains subfolders for different steps in the data processing pipeline.
• paper: Contains our research paper.
• scripts: Contains various scripts for data processing, model creation, and evaluation. Scripts named with a * character were written by me; the others were written by my collaborators.
• sample: Contains sample results for this document.

Findings

Our study finds that in this setting, simple models using the FCN and U-Net architectures outperformed more complex ones, such as the SegNet, Attention U-Net, and DeepLabV3+ models, in terms of accuracy. This illustrates the continued effectiveness of simple models for image segmentation tasks in the context of satellite imagery, despite recent advances in segmentation models generally. We were not able to successfully fine tune the new Segment Anything Model for our purposes.

Results table

Model	Validation data IoU	Validation data Precision	Validation data Recall	Validation data F1-Score	Test data IoU	Test data Precision	Test data Recall	Test data F1-Score
FCN	0.9553	0.9708	0.9836	0.9772	0.9447	0.9620	0.9814	0.9716
U-Net	0.9617	0.9839	0.9771	0.9805	0.9572	0.9815	0.9747	0.9781
SegNet	0.9058	0.9619	0.9395	0.9505	0.9028	0.9518	0.9461	0.9489
Attention U-Net	0.9579	0.9777	0.9793	0.9785	0.9474	0.9682	0.9778	0.9730
DeepLabV3+ ResNet50	0.9334	0.9827	0.9490	0.9656	0.9184	0.9361	0.9799	0.9575
DeepLabV3+ Xception (3 bands)	0.9135	0.9679	0.9421	0.9548	0.9014	0.9641	0.9328	0.9481
DeepLabV3+ Xception (4 bands)	0.9265	0.9649	0.9588	0.9618	0.9235	0.9618	0.9586	0.9602

Predictions

The actual and predicted deforested areas are shown below on a sample test set satellite image. The biggest differences can be seen on the right-hand-side of the picture.

•	•	•	•
Image	Ground Truth Label	FCN Prediction	SegNet Prediction
U-Net Prediction	Attention U-Net Prediction	DeepLabV3+ with ResNet50 Prediction	DeepLabV3+ with 4-band Xception Prediction

Data

For this project, we used the dataset containing 619 4-band satellite images from the Brazilian Amazon rainforest made available by Bragagnolo, Lucimara, da Silva, Roberto Valmir, & Grzybowski, José Mario Vicensi (2021).

Contributions

As stated above, this project was completed jointly with Rick Holubec and Chayenne Mosk. Scripts named with a * character were written by me; the others were written by my collaborators. Our paper includes this statement about our individual contributions:

All authors contributed equally to the research question formulation and data sourcing. All authors contributed to the determination of the analytical framework and methods. All authors completed independent exploratory data analyses. All authors contributed to the evaluation framework used. All authors contributed to the processing and augmentation script used.

• Mosk: implemented FCN, SegNet and Attention U-Net
• Eliason: implemented tiling function, data saving and loading function, CNN (discarded after preliminary analysis), U-Net without attention, SAM fine-tuning experiment
• Holubec: implemented DeepLabV3+ model with multiple backbones and input shapes, worked on model fine-tuning and optimal model selection (ModelCheckpoint, EarlyStopping)

My collaborators and I originally submitted the contents of this repository to a private, anonymized repository for the grading. I have since made the project available here for visibility and added this README file.