Hybrid GeoAI Utilizing DQN-Driven Adaptive Fusion Approach for Approximate Polygon Regularisation of Rasterised Building Footprints

Acquiring accurate building footprints from remote sensing images across a continental scale presents significant challenges due to data variability, insufficient labelling, and computational limitations. We present a GPU-accelerated hybrid GeoAI framework designed to transform rasterised U.S. building data into normalised 3-channel patches (area, count, duplicated area) accompanied by binary masks. The framework consists of an eight-step progressive enhancement pipeline: baseline Mask R-CNN segmentation, GPU acceleration, enhanced regularizers (RT: closing; RR: open-close; FER: edge-augmented), basic RL adaptive fusion, continuous actions via PPO, CNN contextual embeddings, pre-trained models (COCO initialisation), and multi-state training. The pipeline underwent testing on subsets comprising approximately 500 patches sourced from eight U.S. states, utilising the Rasterised Building Footprints for USA dataset. The framework attained a mean IoU of 74.9% (F1 = 0.802), indicating a 7.1% enhancement compared to the baseline Mask R-CNN (67.8%). Although the adaptive reinforcement learning components extend the training duration, the implementation of GPU acceleration and parallelised morphological operations significantly decreases inference time, resulting in a 17.6× speed-up (70.8 ms per patch, 326 patches per minute) when compared to the CPU baseline. Multi-state validation shows a consistent improvement of 4.86% in IoU across different geographies, such as a notable increase of 5.2% in New Hampshire. Ablation studies indicate that the largest contributions come from continuous actions, which account for a 1.4% increase in IoU, and multi-state training, contributing an additional 0.5% to IoU. The platform provides a reproducible test environment for adaptive regularisation, demonstrating improved boundary accuracy and shape retention in qualitative evaluations, while enabling statewide analysis in a matter of hours.