satellite_deforestation_segmenter

Overview

This model is designed for high-resolution semantic segmentation of satellite imagery (RGB) to detect changes in forest cover. It categorizes pixels into six classes, prioritizing the identification of deforested_area and sparse_vegetation to assist in real-time ecological monitoring and conservation efforts.

Model Architecture

The model utilizes the SegFormer architecture, which combines a hierarchical Transformer encoder with a lightweight All-MLP decoder.

• Encoder: Hierarchical Transformer that outputs multi-scale features. Unlike traditional ViT, it does not require positional encodings, making it robust to varying input resolutions.
• Decoder: A simple MLP-based head that aggregates features from different layers to produce the final segmentation mask.

Intended Use

• Environmental Monitoring: Automated detection of illegal logging activities.
• Carbon Credit Verification: Estimating biomass loss in specific geographical sectors.
• Urban Planning: Tracking the encroachment of urban infrastructure into protected green zones.

Limitations

• Cloud Cover: Performance significantly degrades in images with high cloud density or heavy atmospheric haze.
• Topography: Steep terrain shadows may be misclassified as water bodies or dense forest.
• Sensor Variance: Optimized for Sentinel-2 and Landsat-8 data; performance on commercial high-res imagery (e.g., Planet) may require further fine-tuning.