What if flood prediction could keep the physical model’s realism, but deliver answers fast enough for emergency decisions? This paper tests that idea with MaxFloodCast, a machine-learning surrogate for peak inundation depth. As flood occurrences become more intense and frequent, decision makers, emergency managers, infrastructure owners and operators, and first responders have struggled to expeditiously assess and respond to flood accidents.

This is partly because timely, accurate, and reliable information is lacking for complex developed systems. Physics-based hydrodynamic models are typically used to compute flood hazards, but these models can become computationally expensive and even prohibitive as spatial and temporal scale, resolution, and complexity increase.

The central challenge is minimizing computational cost while maintaining accuracy, reliability, and interpretability for near-time flood inundation estimation and prediction. MaxFloodCast is introduced as a surrogate machine-learning modeling framework with novel feature engineering.

It provides peak inundation predictions using nearby and upstream precipitation information while maintaining model interpretability. The model is trained and tested using physics-based model simulations to address the scarcity of historical inundation data. The application covers the multi-watershed region of Harris County, Texas, and validation reproduces Hurricane Harvey in 2017 and Tropical Storm Imelda in 2019.

The study domain covers most of Harris County, Texas, represented by 26,301 polygonal cells classified as channel or non-channel cells. Channel cells represent major natural or manmade streams, while non-channel cells represent overland areas outside water bodies that may drain into nearby channels or remain disconnected.

Because the dominant hydrologic processes differ between these cell types, an individual machine-learning model was created and trained for each cell. The two experiment setups use tree-based XGBoost: Experiment 1 uses within-cell hourly intensity peak and cumulative precipitation, while Experiment 2 adds watershed-wise heavy peak and heavy cumulative precipitation ratios.

MaxFloodCast is trained using physics-based simulations of 592 synthetic storm events, addressing the scarcity of historical inundation data. The synthetic storms vary in duration from one up to 33 hours, precipitation intensity, and spatial movement, and they are stored as hourly NetCDF grids with a resolution of 3,315 feet, or 1,010 meters.

Simulated flood depths come from HEC-RAS 2D, a hydrologic software developed by the Army Corps of Engineers, using a two-dimensional unsteady-flow model. The resulting mesh comprises 26,301 cells and provides a detailed representation of the study area.

The added heavy peak precipitation ratio and heavy cumulative precipitation ratio describe the spatial occurrence of intense precipitation at the watershed level, in addition to cell-level peak and cumulative precipitation. The study considers nine watershed regions, and each ratio is calculated from the area of cells where heavy precipitation occurred.

Heavy precipitation is defined with a threshold of two inches, or 50.8 millimeters, corresponding to the rain intensity for a one-hour event with a return period of at least two years in the study area. For each event, cells above the threshold receive a binary value of one, cells within the threshold receive zero, and the ratio is the area of marked cells divided by watershed area.

XGBoost was selected over support vector machine, random forest, decision tree, and multi-layer perceptron models. After further experiments, a separate model for each cell was more successful than a single universal model, allowing better control of temporal and spatial precipitation dependencies.

Training used ten-fold cross-validation and random-search hyperparameter tuning; the objective minimized squared errors with a learning rate of zero point zero one. Each XGBoost model constructed 1,000 trees with a maximum depth of five, used L1 regularization, and set the column subsample ratio to zero point three.

The study involved two experiment setups. The first incorporated peak and cumulative precipitation within each cell. The second added watershed heavy cumulative precipitation ratio and heavy peak precipitation.

Validation uses data from Hurricane Harvey in 2017 and Tropical Storm Imelda in 2019, alongside HEC-RAS 2D model results. Rainfall was modeled from records at 122 stations in the Harris County Flood Control District network. The gages record every 15 minutes and are distributed within and outside the study area, making them more reliable for this use case than other data sources.

High-water marks were considered for validation outside gauged channels, but high-water marks and gage records differed by up to thirty feet at some locations, so the study relied on stream gages only. Table 1 reports average test R squared and root mean squared error for MaxFloodCast’s two experiments, separating channel, non-channel, and overall cells.

In Experiment 1, non-channel cells have an R squared of zero point nine five three and an RMSE of zero point four six feet, while Experiment 2 reports channel-cell values of zero point nine one six and one point seven six feet. These results matter because they show how performance is documented across the two cell types and experimental setups.

Experiment 1 shows strong performance for predicting inundation depth in non-channel cells using only peak and cumulative precipitation within each cell. Its reported R2 is zero point nine five three, with an RMSE of zero point four six.

Channel-cell performance is comparatively weaker in Experiment 1, likely because that setup does not include precipitation from nearby and upstream contributing drainage areas. Adding precipitation from contributing watersheds and drainage areas in Experiment 2 raises channel-cell R2 from zero point eight seven eight to zero point nine one six.

For those channel cells, RMSE falls from 2.17 feet, or 0.66 meters, to 1.76 feet, or 0.54 meters. Watershed-wise features slightly worsen non-channel predictions, and because approximately 80 percent of cells are non-channel, Experiment 2 performs worse overall than Experiment 1.

For Tropical Storm Imelda, rainfall was non-uniform, less intense, and shorter in duration than Hurricane Harvey, and the additional Experiment 2 features slightly improved prediction compared with Experiment 1. That result highlights the value of spatially relevant precipitation information when rainfall distribution is not uniform.

During Imelda, machine-learning prediction performance was comparable to the physics-based model when gage depth was less than 15 feet, or 4.57 meters, and only slightly worse above that depth. Table three compares HEC-RAS two-D with two machine-learning experiments for Hurricane Harvey, grouping gage records into three depth ranges and reporting RMSE, MAPE, and runtime.

The authors show that the machine-learning runs take one minute thirty-two seconds and one minute forty-six seconds, compared with three minutes thirty seconds for HEC-RAS two-D, while errors vary across the depth groups. This matters because the nearby discussion notes that Hurricane Harvey’s widespread heavy rainfall did not improve Experiment Two, suggesting that added features may introduce noise.

Table four validates peak inundation depths against peak gage records for Tropical Storm Imelda, grouping the records into three depth ranges: below eight feet, eight to under fifteen feet, and at least fifteen feet. It reports RMSE, MAPE, and runtime for HEC-RAS two-D and two machine-learning experiments, with runtimes from one minute and thirty-six seconds to two minutes and forty-nine seconds.

This comparison matters because it shows how prediction error and computational cost are assessed across different observed flood-depth regimes. Machine-learning models trained with physics-based simulations can provide prediction performance comparable to hydrodynamic models while improving computational efficiency and model interpretability.

For most flood intensities considered, MaxFloodCast aligns with a standalone hydrodynamic model when predicting peak flood depths. Prediction of extreme events is less successful, especially when flood depths exceed 15 feet, or 4.57 meters, partly because extreme events were lacking in the training set.

Large and widespread runoff can also trigger operations at Addicks and Barker reservoirs, and those flow alterations were not captured for Harvey and Imelda, partly explaining inconsistent predictions upstream and downstream of the structures. Interpretability and evaluation of watershed interdependencies are presented as unique aspects of MaxFloodCast compared with existing flood prediction models.

Feature-importance analysis examines how flood exposure in one place is influenced by rainfall at that place, in its related drainage area, and in nearby watersheds. These insights can support a system-of-systems approach to flood-risk assessment and help identify flood dynamics such as pluvial flooding, bank overtopping, and backwater effects.

The model can also be retrained over time using updated hydrodynamic-model results that account for city development and flood-mitigation measures. In its current form, the machine-learning model uses only peak flood depth as its output feature.

Flood level is a proxy for flood severity and is used in most damage functions to estimate damage to buildings and infrastructure. Time-to-peak is important for emergency decisions, while inundation duration affects emergency and recovery phases and is also used in building-damage modeling.

Future developments aim to include more training information and produce other hydrograph elements in addition to flood peak. MaxFloodCast can closely match hydrodynamic-model predictions for most tested flood intensities while making predictions more computationally efficient and interpretable, although extreme depths and hydraulic-structure effects remain important limitations.