AI Maps Exactly How Far Urban Lakes Cool City Streets

As heatwaves grow more frequent and more intense across Southeast Asia, the humble urban lake has quietly become one of a city’s most valuable pieces of thermal infrastructure. A new study of Da Nang, Vietnam, now puts hard numbers on that value, using satellite data and an interpretable machine learning framework to quantify precisely how much lakes cool their surroundings and how far that cooling reaches. The findings, published in Discover Cities, offer one of the most detailed pictures yet of how water bodies behave as cooling engines inside a rapidly urbanizing tropical city.

The research, conducted by Nhat-Duc Hoang of Duy Tan University, focuses on 25 lakes scattered across the urban core of Da Nang, a coastal city of roughly 1.3 million people that has experienced dramatic growth in both population and built-up area over recent decades. Between 1996 and 2015, the city’s built-up footprint expanded by an average of 430.9 hectares per year, and with it came the familiar signature of the urban heat island: surface temperatures in dense districts that consistently exceed those of the surrounding countryside. During the dry season of 2025, multiple heatwaves struck the region, making the timing of the study particularly relevant for understanding how lakes perform under genuine thermal stress.

To measure cooling, the study turned to Landsat 8 satellite imagery, processing the thermal infrared band to retrieve land surface temperature, or LST, for every 30-meter pixel in the study area. Rather than comparing each pixel to an arbitrary local reference, the researchers established a single citywide thermal baseline: the mean surface temperature of all built-up and bare-land pixels, which came out at 40.16 degrees Celsius. Cooling intensity at any location was then defined as the difference between that baseline and the pixel’s own temperature. A positive value means a spot is cooler than typical urban surfaces; a negative value means it is hotter. This approach yields results that are directly meaningful for planners, expressed simply as degrees cooler than the prevailing urban background.

Around each of the 25 lakes, the team delineated a 500-meter buffer zone, a distance chosen because previous empirical studies suggest lake cooling rarely extends beyond that range. Within each buffer, 300 sampling points were randomly generated, producing a dataset of 7,500 locations. Each point was described by 16 conditioning factors spanning five categories: distance to the shoreline, lake morphometric properties such as area, perimeter, shape index, compactness, and elongation ratio, land use and land cover composition, urban morphology metrics including the density of built-up surfaces, bare land, shrubland, and tree canopy, and finally topography and proximity features such as elevation, slope, aspect, distance to rivers, and distance to roads.

The predictive engine at the heart of the study is CatBoost, a gradient boosting algorithm that builds an ensemble of decision trees sequentially, with each new tree trained to correct the residual errors of its predecessors. CatBoost was a deliberate choice for this kind of geospatial problem. Its ordered boosting scheme helps mitigate overfitting, it handles both numerical and categorical predictors in a unified framework, and it trains efficiently on large datasets, including with GPU acceleration. Hyperparameters, including tree depth, learning rate, number of boosting iterations, and L2 regularization, were tuned through five-fold cross-validation within the training data.

The model’s performance was strong. On the training set, it achieved a coefficient of determination of 0.96 with a root mean square error of 0.68 degrees Celsius. On the unseen testing set, which comprised 30 percent of the samples, the model still explained 88 percent of the variance in cooling intensity, with an RMSE of 1.27 degrees. An analysis of error types showed that 64.8 percent of predictions fell within one degree of the observed value, a threshold the author adopted as negligible error given that Landsat-derived LST itself carries intrinsic uncertainties on the order of one degree. In other words, the model’s accuracy approaches the physical limit imposed by the satellite data itself.

But prediction alone was not the goal. To understand why lakes cool the way they do, the study employed SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each feature a contribution value for every individual prediction. Positive SHAP values indicate a feature pushes cooling intensity upward; negative values pull it down. The SHAP analysis revealed that distance to river, the lake’s elongation ratio, built-up density, the Polsby-Popper compactness index, and lake area were the most influential predictors, followed by moderate contributions from shrubland density, shape index, perimeter, bare land density, and distance to shoreline.

Several of the model-inferred patterns carry immediate practical weight. The partial dependence analysis for distance to shoreline showed that the lake’s direct cooling contribution is concentrated within roughly 180 meters of the water’s edge, after which the marginal effect fades toward zero. This estimated effective cooling footprint sits comfortably within ranges reported elsewhere, such as the 100 to 150 meter influence documented for lakes in Dhaka, Bangladesh. Meanwhile, the density thresholds were striking: once bare land density around a lake exceeded roughly 0.3, cooling intensity dropped sharply, and when built-up density surpassed about 0.52, the lake’s cooling contribution turned negative. Conversely, shrubland began to boost cooling once its density exceeded about 0.2, and tree canopy enhanced cooling even at low densities, with the effect rising markedly beyond 0.3.

The study is careful to frame these thresholds as model-inferred patterns rather than strict physical laws. They emerge from a specific algorithm, a specific dry season, and a specific city, and the author notes that the relationships uncovered represent statistical associations rather than proven causal mechanisms. Other limitations are acknowledged as well: the 30-meter resolution of Landsat thermal data limits the detection of fine-scale variation, the LULC classification introduces its own errors, variables such as building height, sky view factor, and wind conditions were not included, and a sample of 25 lakes may not generalize to other climates or lake types. The random train-test split also means samples from the same lake appeared in both subsets, which may inflate apparent generalization to entirely independent lakes.

Even with those caveats, the implications for urban planning are concrete. The quantified 180-meter cooling footprint gives Da Nang’s planners a data-driven reference for prioritizing interventions and designing lake-adjacent spaces, with the littoral zone emerging as the critical design focus. The results argue for protecting and expanding vegetation around lakes, capping impervious development in the riparian zone, and treating lakes not as isolated amenities but as active components of blue-green infrastructure networks. In a city where heat stress poses serious public health risks, including elevated cardiovascular and respiratory disease burdens, knowing exactly where and how lakes cool the urban fabric transforms them from scenic backdrops into quantifiable, plannable instruments of climate adaptation.

Subject of Research: Geospatial machine learning modeling of the cooling intensity of urban lakes in Da Nang, Vietnam

Article Title: Geospatial modeling of urban lake cooling based on morphometric and urban contexts using tree ensembles and SHAP interpretability

Article References: Hoang, N.-D. (2026). Geospatial modeling of urban lake cooling based on morphometric and urban contexts using tree ensembles and SHAP interpretability. Discover Cities, 3(1), Article 193. https://doi.org/10.1007/s44327-026-00373-2

Image Credits: AI Generated

DOI: 10.1007/s44327-026-00373-2

Keywords: urban lakes, urban heat island, land surface temperature, CatBoost, SHAP, machine learning, remote sensing, Da Nang, cooling intensity, blue-green infrastructure, urban planning, geospatial analysis