mohammadanwarx/Multi-Modal-Deep-Learning-and-EOFMs-for-Pixel-Level-Forest-Loss-Detection-in-Sudan

# Multi-Modal Deep Learning and EOFMs for Pixel-Level Forest Loss Detection in Sudan Benchmarking multi-modal deep learning architectures and Earth Observation Foundation Models (EOFMs) for **pixel-level forest loss detection in Sudan** using Sentinel-1 SAR and Sentinel-2 optical imagery. **Author:** Mohammad Anwar --- ## Overview A fully remote-sensing-based framework for pixel-level forest loss detection. Rather than relying on field surveys, training labels are constructed via a semi-automatic workflow (change metrics + Dynamic World) and confirmed through manual verification against high-resolution imagery. Three generations of semantic segmentation models are benchmarked: | Model | Type | Role | |-------|------|------| | **U-Net** | CNN | Baseline | | **SegFormer** | Transformer | Baseline | | **Prithvi** | EO Foundation Model | Latest generation | ## Research questions - **RQ1** — Does combining Sentinel-1 and Sentinel-2 improve forest loss detection versus Sentinel-2 alone? - **RQ2** — Can remotely generated + manually verified labels replace field observations for supervised mapping? - **RQ3** — Do EO Foundation Models outperform conventional deep learning models in Sudanese forest ecosystems? ## Data - **Sentinel-2** (L2A, 10 m): spectral bands + derived indices (NDVI, NDMI, NBR, EVI, SAVI) - **Sentinel-1** (GRD, 10 m): VV, VH, VV/VH ratio (+ optional texture / temporal stats) - **Google Dynamic World** (10 m): land-cover probabilities for Trees → Non-Trees transitions - **High-resolution imagery** (Google Earth, ESRI, Bing): manual label verification ## Repository structure ```text . ├── configs/ # Experiment configs (YAML) ├── notebooks/ # Exploratory analysis & prototyping ├── data/ │ ├── raw/ # Downloaded Sentinel-1/2, Dynamic World │ ├── processed/ # Preprocessed, co-registered stacks │ ├── labels/ # Verified forest-loss labels │ ├── patches/ # Model-ready training patc …