Mechanistic Interpretability-Guided Selective Fine-Tuning of Vision-Language Models for Centimeter-Level Flood Depth Estimation

Timely, high-resolution maps of flood extent around settlements are essential for emergency response and damage assessment. We consider airborne RGB imagery for flood mapping as it can be collected rapidly at low cost. To produce flood maps, deep learning models for water segmentation are often used. CNN based and small vision transformer models are used. However, they need much data for adaptation to a change of scenery, i.e., another flooding event. Vision foundation models or large vision transformers are known to generalize across domains. Recently, foundation models for Earth observation became available. They are pretrained on satellite data, whose spatial resolution, viewing geometry, and radiometry differ from nadir RGB imagery. Thus, adaptation is required. We investigate how a satellite-pretrained Earth observation foundation model can be adapted to centimeter-scale floodwater mapping from RGB imagery. Specifically, we fine-tune a model we call Prithvi-2.0-UPN consisting of the Prithvi-EO-2.0-600M Vision Transformer combined with a UPerNet decoder for binary water segmentation on two RGB datasets (BlessemFlood21, NeuenahrFlood). In a first experiment we observe that Prithvi-2.0-UPN reaches state-of-the-art results on BlessemFlood21 and NeuenahrFlood, when trained on their datasets. In a second experiment we show that Prithvi-2.0-UPN performs better than state-of-the-art baseline models for transfer to a new flood event (trained on BlessemFlood21, tested on NeuenahrFlood) in a zero-shot setting. However, the performance indicates room for improvement. In this respect, we investigate in a third experiment how performance improves when further fine-tuning the models with small shares of NeuenahrFlood training data: Prithvi-2.0-UPN improves the fastest and reaches almost the performance level when fully trained on NeuenahrFlood, indicating transfer capabilities.

Deep learning surrogate models trained with mean-squared-error loss produce statistically accurate but physically unconstrained flood predictions: water may flow uphill, appear spontaneously, or smooth over street-level corridors. We develop a physics-informed training framework for CNN-LSTM models that predict urban flood depths at 15 min intervals over a 128x128 spatial grid. Three differentiable penalty terms are embedded into the loss: (i) a gravity loss penalizing depth increases against the water-surface-elevation gradient, (ii) a continuity loss enforcing local mass conservation with rainfall-adaptive thresholds, and (iii) a topography-aware false-alarm penalty modulated by the topographic wetness index (TWI). We evaluate on the Norfolk, Virginia flood dataset spanning two storm events (August 2017 and September 2022, 300 samples), with all variants trained on identical splits and robustness assessed over repeated random splits and leave-one-storm-out tests. A road-proximal evaluation restricted to a TWI-derived street mask quantifies street-level skill. The physics-constrained model achieves near-zero gravity violations (order 1e-6) and the highest street-channel recall (0.77 +/- 0.09 vs 0.44 +/- 0.10 for the unconstrained baseline), the capability most relevant to traffic routing, and its advantage more than doubles on a held-out storm; a uniform false-alarm variant attains 16% lower mean absolute error but suppresses street recall to 0.25. The TWI-modulated penalty reconciles this trade-off: it improves on the uniform variant on every metric, recovering 60% higher street recall at the lowest MAE among constrained variants and the best street-level F1. These results expose a fundamental tension between aggregate pixel-level error and application-specific physical plausibility, and show that terrain-aware loss modulation offers a principled resolution.

Geospatial foundation models can provide strong flood-segmentation performance, but their size limits deployment on memory-constrained edge hardware. We distill a 300-million-parameter Prithvi-EO-2.0 teacher, fine-tuned on the 252 manually labeled Sen1Floods11 training scenes, into a 0.7-million-parameter EfficientViT-B0 student. The teacher supervises additional unlabeled Sentinel-2 imagery, allowing the student training set to grow without new manual annotations. At the matched budget of 252 scenes, teacher-supervised training is competitive with direct training and improves STURM-Flood performance across tested configurations; a geometry-matched control shows that label source alone does not explain the difference. Scaling the teacher-supervised pool to 2,500 scenes narrows the remaining student--teacher gap: the float student reaches 0.787 water intersection over union on the Sen1Floods11 test split against 0.822 for the teacher, matches the teacher on STURM-Flood under our evaluation protocol, and remains below it on WorldFloods-v2. After activation replacement and quantization-aware training, the student runs as a 1.5-megabyte 8-bit integer (INT8) TensorRT engine on a Jetson Xavier NX at 5.57 milliseconds of graphics processing unit (GPU) compute per 512-by-512 image, with approximately 14 megabytes of runtime device memory. A fixed modified normalized difference water index (MNDWI) threshold is competitive with both models on the two clean external benchmarks, so we interpret those benchmarks as generalization tests rather than as evidence of learned-model superiority over a spectral rule. The results support the conclusion: foundation-model supervision can amplify a fixed manual annotation budget into a substantially larger training set and yield a compact, deployable edge model.