Gaia: HybridDeepLabV3+

Gaia is a boundary-aware semantic segmentation framework for aerial imagery. It extends DeepLabV3+ with hierarchical skip connections, attention-based feature fusion, and specialized losses so small objects and land-cover boundaries survive the path from image to mask.

A modular research framework spanning training, inference, evaluation, and cross-dataset experiments. PyTorch and TensorFlow implementations, configuration-driven notebooks, automatic MPS/CUDA/CPU fallback, and mixed batches from DeepGlobe and LandCover.ai keep experiments portable and reproducible.

An Xception encoder captures features at strides 4, 8, and 16 while atrous spatial pyramid pooling adds multi-scale context. The decoder restores detail through hierarchical skip paths and channel, spatial, and adaptive-fusion attention.

A hybrid objective combines class-balanced, focal, boundary-aware, and Dice losses. Aerial-specific augmentation ranges from light to heavy geometric, color, elastic, and spectral transforms.

Reported benchmarks improve mean IoU by 3–5% over the DeepLabV3+ baseline (0.718 to 0.75+), boundary IoU by 5–8% (0.65 to 0.70+), and small-object performance by 8–12% (0.52 to 0.60+).

Standard mean IoU, per-class IoU, and pixel accuracy are paired with boundary IoU, contour matching, mean boundary distance, and sub-class IoU. This separates broad region accuracy from the edge quality that matters most in high-resolution aerial analysis.