Introduction
Accurate historical flood records are essential for disaster management and the development of index-based insurance, where payouts are triggered by satellite-derived data rather than manual inspections. However, a significant data gap exists in current remote sensing capabilities. High-resolution radar data from Sentinel-1 (10m) is effective at penetrating cloud cover but only offers a reliable record since 2017. Conversely, optical data from MODIS provides a deep archive dating back to 2000 but suffers from coarse resolution (250m–1km) and frequent obscuration by the very clouds that accompany flood events. To calculate reliable "return periods" for extreme events, a longer high-fidelity record is required. This research addresses the challenge by using the short, high-quality Sentinel-1 record to train a model capable of inferring high-fidelity historical inundation from the two-decade MODIS archive.
Methodology
The researchers developed a hybrid CNN–LSTM deep learning architecture designed to respect the underlying physics of hydrology. The workflow fuses multi-source satellite data to predict fractional inundated area.
Model Architecture
- Spatial Encoding (CNN): A Convolutional Neural Network (CNN) processes 10 consecutive time-steps of MODIS imagery alongside static landscape features, including elevation, slope, and Height Above Nearest Drainage (HAND). This extracts critical spatial context from the terrain.
- Temporal Dynamics (LSTM): The CNN output is fed into a Long Short-Term Memory (LSTM) network. Unlike standard classifiers, the LSTM captures the "hydrograph signature"—the specific sequence of water rising and receding. This allows the model to "fill in" data gaps caused by cloud cover by understanding the temporal state of the flood.
- Fractional Mapping: Rather than a binary classification, the model predicts the fractional inundated area at 500m resolution, estimating the percentage of each pixel underwater based on labels derived from 10m Sentinel-1 data.
Key Insights and Results
The integration of temporal dynamics proved to be the most significant advancement. The LSTM acts as a physical "sanity check"; if the model recognizes that the ground was saturated in a previous time-step, it more accurately identifies water in the current one, even if the optical signal is degraded by noise or clouds.
Performance Metrics
- Temporal Advantage: A CNN-only model often underestimates monsoon intensity or misses spring flood peaks. Adding the LSTM component increased $R^2$ accuracy from 0.63 to 0.72 during the record-breaking 2020 floods in Bangladesh.
- Superior Accuracy: Across all test years, the CNN–LSTM achieved an average $R^2$ of 0.66 and a Root Mean Square Error (RMSE) of 0.1, indicating the model is generally within 10% of the observed inundation value.
- Historical Fidelity: The framework successfully replicated major historical flood peaks in Bangladesh (2004, 2007, 2011, and 2017), outperforming existing global algorithms like the Global Flood Database (GFD).
Discussion and Future Work
While the CNN-LSTM framework demonstrates high fidelity in seasonal wetlands, it faces challenges in regions with rapidly meandering rivers where the landscape geography shifts over time. Furthermore, with the decommissioning of MODIS in 2023, future iterations must bridge this framework to newer sensors like VIIRS. The authors suggest that applying this method to the AVHRR record could potentially extend high-resolution flood history back to 1979, providing an unprecedented resource for climate resiliency and risk assessment.
Summary updated