This model is a vision transformer designed for high-resolution semantic segmentation of multi-spectral satellite imagery. It is specifically optimized to identify land-use categories and detect environmental changes over time. By utilizing hierarchical feature extraction, it provides precise pixel-level masks for forestry monitoring and urban planning.
Model Architecture
The model is based on the SegFormer architecture (Mix Transformer).
Encoder: A hierarchical Transformer that outputs multi-scale features without the need for positional encoding, making it flexible for various input resolutions.
Decoder: A lightweight All-MLP decoder that aggregates multi-level features to generate the final segmentation map.
Loss Function: Weighted Cross-Entropy to handle class imbalance in geographical data:
$$L = -\sum_{i=1}^{C} w_i y_i \log(\hat{y}_i)$$
Intended Use
Deforestation Tracking: Identifying loss of canopy cover in tropical and temperate forests.
Urban Sprawl Analysis: Mapping the expansion of impermeable surfaces in developing metropolitan areas.
Disaster Assessment: Evaluating flood extent or fire damage by comparing pre- and post-event imagery.
Limitations
Atmospheric Conditions: Heavy cloud cover or smoke can lead to significant misclassification if the input is not pre-processed with atmospheric correction.
Sensor Variability: Performance may vary when applied to sensors with significantly different spectral responses (e.g., Sentinel-2 vs. Landsat-8) without fine-tuning.
Topography: Steep mountainous terrain can introduce shadows that the model might misidentify as water bodies.
