Thabetahmed/FDA-Swin-Crop-Classification-and-Domain-Adaption

Cross-Domain Cereal Crop Classification in Algeria using Fourier Domain Adaptation and Swin Transformers # Cross-Domain Cereal Crop Classification in Algeria ## Using Fourier Domain Adaptation and Swin Transformers > **Paper:** Cross-Domain Cereal Crop Classification in Algeria using Fourier Domain Adaptation and Swin Transformers A data-efficient deep learning framework for crop classification in Algeria using Sentinel-2 satellite imagery. We leverage the label-rich **PASTIS dataset** (France) and transfer knowledge to North Africa using **Fourier Domain Adaptation (FDA)** combined with **Swin Transformers**. ## 📌 Key Results | Method | F1-Score | Notes | |--------|----------|-------| | Target Only | 39.2% | Severe overfitting | | Direct Transfer | 58.7% | Domain shift failure | | Mixed Batch (No FDA) | 88.3% | Limited by spectral shift | | MT-ADA (Iftene & Larabi, 2024) | 80.0% | Feature-level alignment | | **Ours (FDA+Mix)** | **93.2% ± 1.9%** | **State-of-the-Art** | **Potato Classification (Generalization Test):** 85.6% ± 2.4% F1-Score --- ## 🔬 Methodology Overview ### The Problem 1. **Label Scarcity:** Large labeled datasets are scarce in Algeria compared to Europe 2. **Spectral Domain Shift:** Different atmospheric conditions, sun angles, and soil types 3. **Phenological/Temporal Shift:** Different growing seasons between Europe and North Africa ### Our Solution Our framework combines three key innovations: #### 1. Max-NDVI Compression Instead of using full time-series data (susceptible to phenological shifts), we compute the **Maximum NDVI** across the temporal dimension: $$\text{Max-NDVI}(x, y) = \max_{t \in T} \left( \frac{NIR_{t} - Red_{t}}{NIR_{t} + Red_{t}} \right)$$ This creates a time-invariant feature that captures peak vegetative vigor regardless of when it occurred. #### 2. Fourier Domain Adaptation (FDA) FDA transfers the "style" (spectral characteristics) of Algerian images to European images while preserving their "content" (field boundaries, structures): ``` Source Image (Europe) + Target Style (Algeria) → Adapted Image ``` …