Multi-task learning for simultaneous surface water and burned area mapping from earth observation data

Kurzfassung

Multi-task learning has emerged as a promising approach for combining related tasks into a unified framework with reduced computational costs through shared representations. However, its application to hazard mapping using Earth Observation data remains underexplored. In this study, we provide the first systematic evaluation of multi-task deep learning architectures for simultaneous detection of crisis information (surface water and burned area) using Sentinel-2 and Landsat-8 optical imagerytwo societally critical hazards with strong operational demand for rapid and accurate detection. We evaluate five semantic segmentation architectures: three U-Net variants with different encoders (MobileNetV2, EfficientNet-B4, ResNet50) and two transformer-based SegFormer models (SegFormer-B0, SegFormer-B4). Each architecture was trained in both single-task learning and multi-task learning modes. For reference data, we employed the benchmark S1S2-Water dataset for water body detection and constructed a novel burned area dataset by integrating diverse global datasets with already available in-house annotations. Both reference datasets cover heterogeneous geographic regions and temporal periods. The results show that multi-task learning effectiveness is highly architecture-dependent. Burned area detection achieved F1-scores ranging from 0.624 to 0.701, with three of five architectures showing improvements under multi-task learning. SegFormer-B4 reached the best performance (F1-score: 0.67) with gains of +8.24% over its single-task baseline, while EfficientNet-B4 achieved the highest absolute accuracy (F1-score: 0.701). Conversely, ResNet50 exhibited negative transfer with an F1-score decrease of 8.24%, demonstrating that not all architectures benefit from multi-task learning. Water detection performance remained consistently high across all models (F1-score > 0.97), with SegFormer-B4 achieving an F1-score of 0.98. Importantly, the multi-task framework reduces inference time by 89–91% and parameter count by 17-48% compared to maintaining two separate single-task models, demonstrating its practical value for operational emergency mapping services where rapid hazard assessment is critical.