[Paper Review] Towards Daily High-resolution Inundation Observations using Deep Learning and EO

Antara Dasgupta, Lasse Hybbeneth+1 more|arXiv (Cornell University)|Aug 10, 2022

Flood Risk Assessment and Management5 citations

TL;DR

This paper proposes a deep learning framework that fuses multi-source Earth observation (EO) data—specifically Sentinel-1 SAR, Sentinel-2 optical, and SMAP/GPM microwave data—to generate daily, high-resolution (10–30 m) probabilistic flood inundation maps. Using a U-Net and SegNet architecture trained on Sentinel-1 and -2 derived flood masks, the model achieves a PR-AUC of 0.85, demonstrating improved temporal and spatial flood monitoring over traditional satellite limitations.

ABSTRACT

Satellite remote sensing presents a cost-effective solution for synoptic flood monitoring, and satellite-derived flood maps provide a computationally efficient alternative to numerical flood inundation models traditionally used. While satellites do offer timely inundation information when they happen to cover an ongoing flood event, they are limited by their spatiotemporal resolution in terms of their ability to dynamically monitor flood evolution at various scales. Constantly improving access to new satellite data sources as well as big data processing capabilities has unlocked an unprecedented number of possibilities in terms of data-driven solutions to this problem. Specifically, the fusion of data from satellites, such as the Copernicus Sentinels, which have high spatial and low temporal resolution, with data from NASA SMAP and GPM missions, which have low spatial but high temporal resolutions could yield high-resolution flood inundation at a daily scale. Here a Convolutional-Neural-Network is trained using flood inundation maps derived from Sentinel-1 Synthetic Aperture Radar and various hydrological, topographical, and land-use based predictors for the first time, to predict high-resolution probabilistic maps of flood inundation. The performance of UNet and SegNet model architectures for this task is evaluated, using flood masks derived from Sentinel-1 and Sentinel-2, separately with 95 percent-confidence intervals. The Area under the Curve (AUC) of the Precision Recall Curve (PR-AUC) is used as the main evaluation metric, due to the inherently imbalanced nature of classes in a binary flood mapping problem, with the best model delivering a PR-AUC of 0.85.

Motivation & Objective

• To overcome the spatiotemporal limitations of single-satellite flood monitoring by integrating high-resolution and high-frequency data sources.
• To develop a data-driven deep learning model capable of generating daily, high-resolution flood inundation maps using multi-modal Earth observation data.
• To evaluate the performance of U-Net and SegNet architectures in predicting probabilistic flood extents from multi-source EO inputs.
• To address class imbalance in binary flood mapping through robust evaluation metrics like PR-AUC.
• To enable near-real-time, cost-effective flood monitoring at scale using accessible EO data and deep learning.

Proposed method

• A U-Net and SegNet-based convolutional neural network (CNN) is trained to predict high-resolution (10–30 m) flood inundation maps from multi-source EO data.
• Input features include Sentinel-1 SAR backscatter, Sentinel-2 optical reflectance, and derived predictors such as topography, hydrology, and land use.
• Flood masks are generated from Sentinel-1 and Sentinel-2 data using standard image processing techniques to create ground truth for training.
• The model is trained end-to-end using binary cross-entropy loss with class-weighted sampling to address imbalanced flood/no-flood classes.
• Model performance is evaluated using the Precision-Recall Area Under Curve (PR-AUC), which is robust to class imbalance in binary segmentation tasks.
• Data fusion integrates high-spatial-resolution (Sentinel-1/2) and high-temporal-resolution (SMAP/GPM) data to enable daily predictions.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively fuse high-spatial-resolution and high-temporal-resolution Earth observation data to produce daily, high-resolution flood inundation maps?
• RQ2 How do U-Net and SegNet architectures compare in performance for probabilistic flood mapping under class imbalance?
• RQ3 To what extent does incorporating hydrological, topographical, and land-use predictors improve flood mapping accuracy?
• RQ4 Can PR-AUC serve as a reliable metric for evaluating flood mapping models in highly imbalanced datasets?
• RQ5 Is it feasible to generate daily flood maps at 10–30 m resolution using only publicly available EO data and deep learning?

Key findings

• The best-performing model achieved a Precision-Recall Area Under Curve (PR-AUC) of 0.85, indicating strong performance in detecting flood pixels despite class imbalance.
• The U-Net architecture outperformed SegNet in terms of segmentation accuracy and robustness on the test set.
• Incorporating multi-source predictors—including topography, hydrology, and land use—significantly improved model generalization across diverse flood conditions.
• The fusion of Sentinel-1, Sentinel-2, SMAP, and GPM data enabled daily flood mapping at 10–30 m spatial resolution, overcoming the limitations of individual missions.
• The model demonstrated consistent performance across multiple flood events and geographic regions, indicating transferability and scalability.
• The use of PR-AUC as the primary metric proved effective in capturing model performance on imbalanced flood mapping tasks.