Machine learning maps where Dhaka is most likely to flood after rain

Dhaka, one of the most densely populated cities on Earth, now has a new, spatially explicit map of where water is most likely to pool and linger after heavy monsoon rain, and the picture it paints is one in which terrain, rainfall and the density of the built environment together decide which neighborhoods drown. A research team led by Hiron Khandaker and Hriday Dey, publishing in the journal Environmental Challenges, built an ensemble machine learning framework that compares three algorithms, Random Forest, XGBoost and LightGBM, to model waterlogging susceptibility across the Dhaka City Corporation area, and then used explainable AI to reveal which factors drive the predictions.

The starting point was a waterlogging inventory assembled from a 2024 hotspot map published by the Bengal Institute, georeferenced against the city boundary, and then verified and updated through field surveys during the severe waterlogging event of 10 to 13 July 2026, which followed record precipitation. The team surveyed eight sites across the Dhaka North and Dhaka South city corporations, confirming or adjusting the mapped polygons based on the extent of standing water they observed. From these polygons they generated 500 presence points and 500 pseudo-absence points, a deliberately balanced 1:1 sample of 1,000 spatially referenced observations. The authors are explicit that this artificial prevalence means the model outputs must be read as a relative susceptibility index, not as calibrated real-world probabilities of flooding.

Two variables, NDVI and distance to road, were dropped after multicollinearity screening, which found an extreme variance inflation factor of 64.15 for NDVI and a strong negative rank correlation between distance to road and road density. The remaining eleven predictors all showed variance inflation factors well below 2.

A central methodological choice sets this study apart from many earlier susceptibility maps. Because geographically close samples tend to be spatially autocorrelated, conventional random train-test splits can produce flattering performance estimates. The team instead partitioned their sample points into five geographically cohesive blocks using k-means clustering and ran five-fold cross-validation with scikit-learn’s GroupKFold, so that training and test samples in every fold came from entirely disjoint parts of the city. Minimum train-test separation distances ranged from 198 to 431 meters across folds, with median nearest-training distances between 2,066 and 2,728 meters. Hyperparameters were tuned inside a nested spatial cross-validation framework to prevent information leakage, with the outer test fold never used for optimization.

The three classifiers performed comparably and well. Fold-wise mean ROC-AUC was 0.875 plus or minus 0.078 for XGBoost, 0.859 plus or minus 0.085 for Random Forest, and 0.848 plus or minus 0.092 for LightGBM across the four evaluable spatial folds, since the fifth fold contained no test presences and its AUC was undefined. Pooled out-of-fold ROC-AUC ranged from 0.91 to 0.93. A bootstrap analysis of pairwise differences found no statistically significant gap between XGBoost and Random Forest, while LightGBM was slightly but significantly worse. XGBoost also showed the best probability calibration with a Brier score of 0.08, followed by LightGBM at 0.09 and Random Forest at 0.10. The authors stress that because training data were balanced 50:50, precision, F1-scores and predicted probabilities represent relative ordering within the sample rather than calibrated event likelihoods.

Applied pixel by pixel across the city, the models produced broadly consistent susceptibility surfaces. Under a common fixed 0.20-interval classification, the proportion of Dhaka City Corporation in the very-high susceptibility zone was 14.73 percent, or 44.86 square kilometers, for Random Forest, 17.82 percent, or 54.26 square kilometers, for XGBoost, and 17.27 percent, or 52.60 square kilometers, for LightGBM, while very-low susceptibility covered 65.66 to 72.23 percent of the area depending on the model. High and very-high zones concentrated in the central-western and southern urban belt, including Mirpur, Mohammadpur and adjoining areas, and coincided with high flow accumulation, dense settlement and population, high road density and strong spectral moisture signatures.

The explainability layer proved the most revealing. SHAP analysis, computed with TreeExplainer on models refitted to the full 1,000-sample dataset, showed a consistent importance hierarchy across all three algorithms: rainfall, elevation and road density were always the most influential predictors, with settlement density fourth. Road density and settlement density showed consistent positive relationships with susceptibility, pointing to the role of dense impervious infrastructure. Interestingly, some SHAP distributions showed higher rainfall and higher elevation values associated with positive contributions, a counterintuitive pattern the authors attribute to the coarse regional rainfall gradient and to the fact that inventory samples cluster in the urban core, where filled ground sits higher than the unfilled floodplain periphery. They are careful to note that SHAP values describe conditional model behavior, not causal mechanisms.

The study then connected the susceptibility map to the changing shape of the city itself. Landscape metrics computed from classified Landsat imagery for 2010, 2015, 2020 and 2025, with overall accuracies of 89 to 90 percent and kappa coefficients of 86 to 87 percent, documented dramatic urban growth and consolidation. The proportion of landscape occupied by urban land rose from 37.80 percent in 2010 to 55.70 percent in 2025, the largest patch index climbed from 28.11 to 50.40 percent, the aggregation index increased from 82.88 to 89.49 percent, and the number of urban patches jumped from 4,040 to 15,938, with the most pronounced changes after 2015. Mean urban proportion rose from roughly 42 to 48 percent in the very-low susceptibility class to around 71 to 77 percent across the low to very-high classes, although the relationship was not strictly monotonic.

Comparing areas that transitioned from non-urban to urban between 2010 and 2025 with the 2026 susceptibility classes showed a heterogeneous, model-dependent pattern. Under Random Forest, 82.16 percent of newly urbanized area fell within the very-low susceptibility class, with 7.64 percent in the very-high class; LightGBM gave a nearly identical 82.27 percent in the very-low class, while XGBoost placed 64.04 percent there and 18.11 percent in the low class. Recent urban expansion, in other words, was not concentrated exclusively in high-susceptibility locations, and the authors repeatedly caution that these are spatial associations, not evidence that the 2010 to 2025 morphological changes caused the current susceptibility pattern. A direct temporal or causal attribution would require longitudinal susceptibility observations and temporally matched conditioning variables.

The team also probed the robustness of their results. A pseudo-absence sensitivity analysis across exclusion buffer distances of 0 to 1,000 meters found Spearman rank correlations between susceptibility surfaces of 0.943 to 0.972, top-20 percent spatial similarity of 0.783 to 0.879, and SHAP rank correlations of 0.818 to 0.988, indicating that the zonation and variable rankings were very robust to how negative samples were placed. A threshold analysis showed the expected sensitivity-specificity trade-off as cutoffs moved from 0.70 down to 0.30, with Youden’s J statistic suggesting optimal operating thresholds of 0.35 for Random Forest, 0.27 for XGBoost and 0.16 for LightGBM. The authors acknowledge real limitations: the rainfall predictor is an annual total at roughly 5 kilometer resolution rather than event-scale data, the terrain variables derive from a 2000-era DEM that cannot capture subsequent land filling and construction, there is a temporal mismatch between 2025 predictors and the historical-plus-2026 inventory, and drainage density does not measure hydraulic capacity or blockage. They suggest future work with multi-event inventories, high-resolution LiDAR elevation data and hydrodynamic modelling, and argue that areas where high susceptibility coincides with intensive urban occupation may warrant priority for drainage and retention interventions, while less-developed susceptible areas offer opportunities to protect wetlands, open spaces and drainage corridors before they are built over.

Subject of Research: Ensemble machine learning modelling of urban waterlogging susceptibility in Dhaka, Bangladesh

Article Title: Urban expansion and waterlogging susceptibility in Dhaka city: an explainable ensemble machine learning model and urban morphology framework

Article References: Khandaker, H., Dey, H., Khuda, I. H. M. E., Ryhan, M. A., & Rahman, M. T. (2026). Urban expansion and waterlogging susceptibility in Dhaka city: an explainable ensemble machine learning model and urban morphology framework. Environmental Challenges, 25, Article 101703. https://doi.org/10.1016/j.envc.2026.101703

Image Credits: AI Generated

DOI: 10.1016/j.envc.2026.101703

Keywords: Dhaka, waterlogging, urban flooding, machine learning, XGBoost, Random Forest, LightGBM, SHAP, spatial cross-validation, urban morphology, landscape metrics, susceptibility mapping