FusionAttNet Framework for Hierarchical Attention Driven Sentinel 1 and Sentinel 2 Fusion for Semi Arid Land Cover Classification in Far North Cameroon

Technical Summary: FusionAttNet for Semi-Arid Land Cover Classification

Problem Statement
Accurate land cover mapping in semi-arid regions, specifically the Sudano-Sahelian ecotone of Far North Cameroon, faces significant challenges due to spectral homogeneity among sparse vegetation, high seasonal variability, and persistent cloud cover that obscures optical data. Traditional approaches, including optical-only deep learning and static feature-stacking fusion methods, often fail to distinguish spectrally similar classes (e.g., bare soil versus grassland) and are vulnerable to data gaps caused by clouds. Furthermore, existing hierarchical frameworks often rely on predefined rules rather than learned spatial relationships, while single-sensor models lack the robustness required for all-weather monitoring.

Methodology
The study proposes FusionAttNet, a novel deep learning framework designed to dynamically integrate Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical data. The methodology is structured around several key components:

• Study Area & Data: The framework was validated over a 1,280 km² area near Maroua, Cameroon, using a 12-month temporal window (2021) to capture full seasonal phenology. Data sources include Sentinel-2 Level-2A Surface Reflectance (13 bands, downsampled to 10m) and Sentinel-1 GRD SAR data (VV/VH polarizations).
• Preprocessing & Feature Engineering:
  • Optical Pipeline: Includes cloud masking (QA60), scaling, and the calculation of vegetation indices (NDVI, EVI, NDWI).
  • SAR Pipeline: Involves radiometric calibration, dB conversion, and the derivation of backscatter ratios.
  • Cross-Modal Indices: The framework generates specific indices such as the NDVI/VV ratio to merge optical and radar information, alongside topographic variables derived from SRTM DEM.
• Architecture:
  • Modality-Aware Processing: Parallel processing streams handle optical and SAR data separately before fusion, respecting the unique characteristics of each sensor.
  • Hierarchical Attention Mechanism: A three-tiered attention system operates at the pixel, patch, and landscape levels. This allows the model to capture fine-grained details while simultaneously resolving spatial ambiguities through broader contextual relationships.
  • Fusion Strategy: Instead of static concatenation, the model employs a learned, adaptive fusion mechanism that weights multi-scale features dynamically.
• Training & Optimization: The model utilizes a weighted cross-entropy loss function with label smoothing and focal loss to address class imbalance. Training is optimized using the AdamW optimizer with cosine annealing and warm restarts. Data augmentation is applied specifically to preserve SAR backscatter textures.
• Ensemble Approach: The final classification output is derived from an ensemble combining the deep learning FusionAttNet branch with traditional machine learning models (Random Forest and XGBoost) to maximize discriminative power.

Key Contributions

1. Hierarchical Attention Architecture: The introduction of a pixel-patch-landscape attention mechanism that effectively resolves spectral confusion in semi-arid ecotones, particularly at the interfaces of bare soil and grassland.
2. Modality-Aware Cross-Modal Fusion: The development of a framework that integrates SAR backscatter with optical indices (e.g., NDVI/VV) to mitigate cloud-induced data gaps without relying on complex, error-prone cloud removal pre-processing.
3. SAR-Optimized Augmentation: The implementation of augmentation techniques that preserve backscatter textures, enhancing the model's ability to generalize across seasonal variations.
4. Ensemble Integration: The successful combination of deep learning features with gradient-boosted heuristics to stabilize classification in complex, fragmented landscapes.

Results
The FusionAttNet framework was evaluated against baseline models (Random Forest, XGBoost, and a standalone MultiModal Network) on an eight-class land cover dataset.

• Overall Performance: The full Ensemble approach achieved an Overall Accuracy of 96.75% and a Weighted F1-Score of 0.9680.
• Comparison: This performance outperformed the standalone MultiModal Network (94.75% accuracy) and matched or exceeded the XGBoost baseline (96.75% accuracy).
• Class-Specific Improvements:
  • Grassland: The ensemble corrected a significant performance deficit in the standalone network, boosting the Grassland F1-score from 0.8905 to 0.9530 (a +6.25% absolute recovery).
  • Herbaceous Wetlands: While challenging, the ensemble stabilized this minority class to an F1-score of 0.7273, significantly outperforming the Random Forest baseline (0.5000).
  • Macro-Structural Classes: Classes with distinct radiometric profiles (Tree Cover, Built-up, Barren, Permanent Water) achieved perfect metrics (F1 = 1.0000) across all configurations.
• Error Reduction: The study reports a 14.7% reduction in misclassification at mixed bare soil/grassland interfaces compared to state-of-the-art methods.

Significance
The paper claims that FusionAttNet establishes a new paradigm for semi-arid land cover monitoring by shifting the focus from simple gap-filling to intelligent, multi-sensor fusion. By leveraging the all-weather reliability of Sentinel-1 SAR alongside the spectral richness of Sentinel-2, the framework overcomes the limitations of cloud dependency and spectral homogeneity. The results demonstrate that a modality-aware hierarchical attention mechanism can significantly reduce spectral confusion in Sahelian ecotones, providing a robust, scalable solution for operational land cover monitoring in data-scarce and seasonally variable environments. The study highlights that while deep learning alone offers high-quality spatial features, the integration with traditional machine learning ensembles is crucial for neutralizing local vulnerabilities and achieving maximum stability in complex transitional zones.