Resolving Scale Conflict via Synchronized Multi-Stream Learning: A Framework for High-Fidelity Geospatial Mapping

Digital mapping of rural land registries is integral to financial inclusion in the Global South, but is currently constrained by the ``scale problem'' of uncontrolled aerial imagery. Existing semantic segmentation approaches have trouble pinpointing macro-geometric features and micro-structures (sub-meter) in the same dataset, such as cropland boundaries and infrastructure. In this paper, we present a Multi-Stream Hybrid Deep Learning approach for the extraction of complex geospatial features from unmanned aerial vehicle (UAV) images without explicit multi-resolution cascading. The architecture incorporates an enhanced U-Net++ semantic segmentation stream for geometry extraction with a ResNet-driven attribute stream and a dedicated point-detection stream to attain survey-level accuracy. We test the effectiveness of our approach within a highly skewed, data-poor scenario across multiple agro-climatic landscapes, showing strong domain generalization. We also benchmark zero-shot generalization on the public WHU Building Dataset and demonstrate the spectral distribution differences between low-altitude UAVs and very high-altitude (VHA) platforms.