Obtaining pixel-level annotations over large spatial extents remains a major bottleneck for deploying machine learning in ecological applications. Here we present a multi-scale weakly supervised semantic segmentation (WSSS) framework that enables training high-resolution segmentation models from dense, classification-based outputs. Our method combines fine-scale, multi-label predictions from underwater imagery with broad-coverage aerial data. We convert these point-level classifications into coarse supervision masks that can be used to train a semantic segmentation model on Unmanned Aerial Vehicle (UAV) orthophotos. A second training step using the model's own refined predictions is then used to further improve spatial accuracy without requiring additional annotations. We demonstrate the approach on coral reef imagery, enabling large-area segmentation of coral morphotypes and illustrating its flexibility in integrating new classes. The final model achieves 86.07% pixel accuracy and 52.23% mean Intersection over Union (mIoU) on manually annotated reef zones, demonstrating that accurate large-scale coral segmentation can be obtained without pixel-level annotations. By bridging image classification and segmentation across scales and modalities, this method provides an efficient solution for deploying segmentation models in settings where annotations are unavailable and opens opportunities for scalable, efficient monitoring in ecology and beyond.