Deforestation Detection in Peru

The Challenge

Deforested areas cover only about 2.9% of the pixels in the images, so a model can reach high accuracy and still miss almost all forest loss. We needed a model that finds small clearings by comparing an image taken before with one taken after.

Our Solution

We built the full pipeline: Sentinel-2 download through STAC, mask creation from 9,910 official SERFOR polygons, and a U-Net with a ResNet34 encoder trained in PyTorch on 8 bands (red, green, blue and near-infrared, before and after). We tested 7 configurations, changing encoders, NDVI bands and loss functions, and built a web tool to review and correct the labels.

Key Results

• IoU 0.413 and F1 0.584 on the test set
• 1,039 curated image pairs (727 train, 156 validation, 156 test)
• Public code and a web tool to review labels