<p>Accurate and timely flood monitoring is critical for disaster risk reduction, yet operational deep-learning flood mapping in resource-limited regions is constrained by high computational cost and the cloud sensitivity of optical imagery. We present a scalable, weather-independent framework that fuses Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical data for training, then deploys a lightweight KD-compressed model for basin-scale monitoring. The workflow comprises three stages: (i) sparse label generation via unsupervised clustering on fused Sentinel-1/Sentinel-2 features with threshold-based confidence filtering, retaining only high-confidence pixels as supervised targets; (ii) Knowledge Distillation (KD) from a PSPNet teacher to compact students, evaluated under redundancy-stratified five-fold cross-validation with spatial autocorrelation control; and (iii) SAR-only deployment using Sentinel-1 VV and VH for all-weather inference. A band-agnostic, single-channel architecture trained on seven spectral bands requires only SAR at inference, decoupling operational monitoring from optical availability. We evaluate the framework over a 14-month wet–dry cycle in the Mun–Chi Basin, northeastern Thailand. KD from PSPNet to PSPMixer (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({\sim }6.5{\times }\)</EquationSource> </InlineEquation> compression, 349k parameters) raises <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\text {IoU}_{\text {pos}}\)</EquationSource> </InlineEquation> from 0.100 to 0.404 while producing spatially coherent flood masks; a <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(571{\times }\)</EquationSource> </InlineEquation>-compressed U-NetLight (11.6k parameters) is also evaluated but PSPMixer KD is selected for deployment owing to superior map quality. Focal Tversky Loss consistently outperforms cross-entropy, which causes complete failure in two of five architectures. At basin scale, predicted inundation area tracks gauge water level with Pearson <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(r = 0.858\)</EquationSource> </InlineEquation> and a <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\({\sim }7\)</EquationSource> </InlineEquation>-day mean lag explained by floodplain storage and geomorphological routing across trap, pathway, and far-field domains. These results demonstrate that KD-compressed models can faithfully reproduce basin-scale flood dynamics at a fraction of the computational cost of their teachers, offering a practical pathway toward near-real-time, climate-resilient flood monitoring in data-scarce tropical basins.</p> Graphical Abstract <p>This graphical abstract presents a streamlined workflow for scalable, weather-independent flood monitoring in resource-limited regions, integrating satellite remote sensing, machine learning, and model compression. The left section illustrates the study background, highlighting the challenge of monitoring flood-prone floodplains with limited computational resources. Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical imagery are identified as key data sources, with SAR enabling consistent monitoring under all weather conditions. The central panel depicts the methodology, beginning with the acquisition of SAR and optical datasets, followed by sparse labeling via unsupervised clustering with threshold-based confidence filtering to reduce annotation requirements. A Knowledge Distillation (KD) framework is then employed, transferring predictive capabilities from a high-capacity PSPNet teacher model to a compact PSPMixer student model. This <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\({\sim }6.5{\times }\)</EquationSource> </InlineEquation> compression pipeline (349k parameters) raises <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\text {IoU}_{\text {pos}}\)</EquationSource> </InlineEquation> from 0.100 to 0.404 while producing spatially coherent flood masks, trained with Focal Tversky Loss under a redundancy-stratified five-fold cross-validation protocol. The right panel summarizes the main findings. SAR-only inference enables operational deployment in cloud-covered conditions, producing flood extent maps with strong agreement to hydrological gauge measurements (Pearson <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(r = 0.858\)</EquationSource> </InlineEquation>). The framework captures a <InlineEquation ID="IEq9"> <EquationSource Format="TEX">\({\sim }7\)</EquationSource> </InlineEquation>-day mean lag between peak flood extent and downstream gauge peaks, explained by floodplain storage and geomorphological routing across trap, pathway, and far-field domains. Overall, the visual encapsulates how data fusion, sparse labeling, and knowledge distillation synergize to deliver a cost-effective, reproducible, and operationally viable approach for flood risk monitoring in data- and resource-constrained settings, with potential for broader application in Earth systems and environmental monitoring.</p>

Distilling Intelligence from Space: Lightweight Deep Learning for SAR-Based Flood Monitoring

期刊论文
发表日期：

2026年4月29日

查看全文

Pongthep Thongsang,
Srilert Chotpantarat,
Wacharapong Saengnill,
Sukonmeth Jitmahantakul

摘要

Accurate and timely flood monitoring is critical for disaster risk reduction, yet operational deep-learning flood mapping in resource-limited regions is constrained by high computational cost and the cloud sensitivity of optical imagery. We present a scalable, weather-independent framework that fuses Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical data for training, then deploys a lightweight KD-compressed model for basin-scale monitoring. The workflow comprises three stages: (i) sparse label generation via unsupervised clustering on fused Sentinel-1/Sentinel-2 features with threshold-based confidence filtering, retaining only high-confidence pixels as supervised targets; (ii) Knowledge Distillation (KD) from a PSPNet teacher to compact students, evaluated under redundancy-stratified five-fold cross-validation with spatial autocorrelation control; and (iii) SAR-only deployment using Sentinel-1 VV and VH for all-weather inference. A band-agnostic, single-channel architecture trained on seven spectral bands requires only SAR at inference, decoupling operational monitoring from optical availability. We evaluate the framework over a 14-month wet–dry cycle in the Mun–Chi Basin, northeastern Thailand. KD from PSPNet to PSPMixer ( \({\sim }6.5{\times }\) compression, 349k parameters) raises \(\text {IoU}_{\text {pos}}\) from 0.100 to 0.404 while producing spatially coherent flood masks; a \(571{\times }\) -compressed U-NetLight (11.6k parameters) is also evaluated but PSPMixer KD is selected for deployment owing to superior map quality. Focal Tversky Loss consistently outperforms cross-entropy, which causes complete failure in two of five architectures. At basin scale, predicted inundation area tracks gauge water level with Pearson \(r = 0.858\) and a \({\sim }7\) -day mean lag explained by floodplain storage and geomorphological routing across trap, pathway, and far-field domains. These results demonstrate that KD-compressed models can faithfully reproduce basin-scale flood dynamics at a fraction of the computational cost of their teachers, offering a practical pathway toward near-real-time, climate-resilient flood monitoring in data-scarce tropical basins.

Graphical Abstract

This graphical abstract presents a streamlined workflow for scalable, weather-independent flood monitoring in resource-limited regions, integrating satellite remote sensing, machine learning, and model compression. The left section illustrates the study background, highlighting the challenge of monitoring flood-prone floodplains with limited computational resources. Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical imagery are identified as key data sources, with SAR enabling consistent monitoring under all weather conditions. The central panel depicts the methodology, beginning with the acquisition of SAR and optical datasets, followed by sparse labeling via unsupervised clustering with threshold-based confidence filtering to reduce annotation requirements. A Knowledge Distillation (KD) framework is then employed, transferring predictive capabilities from a high-capacity PSPNet teacher model to a compact PSPMixer student model. This \({\sim }6.5{\times }\) compression pipeline (349k parameters) raises \(\text {IoU}_{\text {pos}}\) from 0.100 to 0.404 while producing spatially coherent flood masks, trained with Focal Tversky Loss under a redundancy-stratified five-fold cross-validation protocol. The right panel summarizes the main findings. SAR-only inference enables operational deployment in cloud-covered conditions, producing flood extent maps with strong agreement to hydrological gauge measurements (Pearson \(r = 0.858\) ). The framework captures a \({\sim }7\) -day mean lag between peak flood extent and downstream gauge peaks, explained by floodplain storage and geomorphological routing across trap, pathway, and far-field domains. Overall, the visual encapsulates how data fusion, sparse labeling, and knowledge distillation synergize to deliver a cost-effective, reproducible, and operationally viable approach for flood risk monitoring in data- and resource-constrained settings, with potential for broader application in Earth systems and environmental monitoring.