The scale of transformation is striking. Built-up land in Bharatpur expanded from 21.63 percent of the city’s 433 square kilometers in 2000 to 45.11 percent in 2024, a gain of roughly 102 square kilometers of concrete, asphalt, and roofing. Much of that expansion consumed the city’s open and barren land, which collapsed from nearly 32 percent of the study area to under 2 percent over the same period, before the pressure of growth shifted increasingly onto vegetated surfaces. The pattern mirrors a two-stage urbanization process documented in other rapidly growing cities of Africa and South Asia: easily developable open land is absorbed first, and agricultural and vegetated land follows. Population growth in Bharatpur has run at roughly 5 percent annually, fueled by migration from hill districts, post-earthquake displacement after 2015, and the city’s designation as a candidate smart city.

Methodologically, the study assembles an unusually thorough pipeline. The researchers used Google Earth Engine to process cloud-free Landsat composites from the pre-monsoon April-to-June window, when heat stress peaks in the Terai. A Random Forest classifier—an ensemble method that builds many decision trees on bootstrap samples and aggregates their votes—assigned each pixel to built-up, vegetation, water, or barren land, achieving overall accuracies between 89 and 95.5 percent with Kappa coefficients as high as 0.94. Land surface temperature was retrieved from the thermal infrared bands through a standard radiometric chain, converting digital counts to top-of-atmosphere radiance, then brightness temperature, and finally surface temperature after correcting for emissivity estimated from vegetation cover. Satellite-derived temperatures were bias-corrected against ground observations from a Nepal Department of Hydrology and Meteorology station, closing a validation loop that many comparable city-scale studies lack.

The thermal record is not a simple warming line. Mean surface temperature peaked at 37.72 degrees Celsius in 2000, declined to a minimum of 30.46 degrees in 2020—a year when monsoon flooding expanded water bodies and pandemic lockdowns curtailed construction and traffic—and then rebounded sharply to 35.45 degrees in 2024 as the structural drivers of heating reasserted themselves. The single hottest pixel ever recorded in the study reached 57.75 degrees Celsius in 2010, concentrated in the dense commercial core of Narayangadh. Tellingly, the standard deviation of surface temperature fell steadily across the study period, from 4.09 to 3.18, an early statistical hint that the city’s thermal landscape is flattening into a more uniformly warm surface rather than concentrating its heat in a few extreme zones.

Perhaps the most consequential finding concerns which land-surface property best predicts temperature. Across all land cover classes and years, the Normalized Difference Moisture Index, or NDMI—a measure of water content in vegetation and soil—showed the strongest negative correlation with land surface temperature, at r = −0.84. That relationship significantly exceeded the correlation for the Normalized Difference Vegetation Index, the traditional favorite of urban heat studies, which registered −0.64 in built-up areas and only −0.46 in vegetated zones. The difference was confirmed as statistically significant using Fisher’s z-transformation on stratified samples of 800 pixels per class. Meanwhile the built-up index, NDBI, correlated positively with temperature at +0.81. The implication is that in Bharatpur’s hot, monsoon-influenced climate, surface moisture—wet soils, riparian corridors, water bodies, evapotranspiring ground—does more thermal work per unit than canopy cover alone.

To look forward, the team coupled a Cellular Automata-Markov model, which estimates the quantity of land transitions, with a Random Forest-derived suitability surface that determines where new development is most likely to occur. The model was tested on a strict hold-out of the 2020-to-2024 transition, achieving a Figure of Merit of 0.302—modest in absolute terms but a substantial improvement over an earlier Random Forest-only projection that, despite high overall accuracy, fell into the classic no-change trap of predicting essentially none of the pixels that actually urbanized. A separate Random Forest regression, trained on data from 2000 to 2020 and validated on the withheld 2024 scene, predicted the standardized 2030 thermal field with an R-squared of 0.578 and a root-mean-square error of 2.07 degrees Celsius.

The projections describe a heat island that grows outward rather than upward in intensity. Built-up area is forecast to reach 56.21 percent of the city by 2030, an 11-percentage-point jump, while the statistically significant heat-hotspot area, identified through Getis-Ord Gi* spatial cluster analysis, expands from 36.69 to 49.32 percent of the study area. Under Urban Thermal Field Variance Index zoning, the land classified as ecologically heat-stressed grows from 38.84 to 48.95 percent. Yet the peak-to-baseline intensity of the heat island is projected to narrow from +3.41 to +1.62 degrees Celsius, and mean citywide warming attributable to land cover change alone is only about 0.22 degrees. The explanation is mechanistic: new development is largely allocated onto barren fringe land that is already nearly as hot as the urban core, so the city spreads its warmth across a wider area rather than raising its thermal ceiling. By 2030, some 83.4 percent of all built-up pixels fall within statistically significant hotspots—a striking agreement between a temperature-only statistic and the land cover map.

For planners, the takeaway is twofold. First, aggregate temperature statistics systematically understate the exposure problem: the number of residents living on thermally elevated ground is growing far faster than the city’s mean temperature, with direct consequences for heat-related mortality, cooling energy demand, and outdoor labor productivity in a city of nearly 370,000 people. Second, if moisture governs cooling more powerfully than canopy, then protecting the Narayani River riparian corridor, conserving water bodies, promoting permeable surfaces, and maintaining soil moisture capacity deserve first-order status in Bharatpur’s master planning rather than being treated as secondary greening measures. The authors argue this reframing is underemphasized across the South Asian heat mitigation literature.

The study is candid about its limits. Validation rests on a single weather station that cannot capture the full thermal heterogeneity of a 433-square-kilometer city; the four-class land cover scheme cannot separate thermally distinct commercial, residential, and industrial zones; predictions assume business-as-usual transitions without modeling policy or demographic shocks; and the 30-meter resolution of Landsat thermal data misses fine-scale cooling from small parks and narrow canals. Random Forest mean-reversion also smooths predicted temperature extremes, making the warm-field and mean-shift metrics more trustworthy than tail estimates. Even so, the workflow—cloud-based classification, bias-corrected thermal retrieval, stratified correlation analysis, and temporally validated prediction—offers a directly replicable template for the other fast-growing Terai cities such as Butwal, Birgunj, Biratnagar, and Janakpur, where comparable urbanization is underway but systematic thermal assessment has yet to begin.

Subject of Research: Urban heat island dynamics and land surface temperature change in Bharatpur Metropolitan City, Nepal, assessed with multi-temporal Landsat remote sensing and machine learning

Article Title: Assessing urban heat island patterns in Bharatpur metropolitan city of Nepal using multi-temporal remote sensing and machine learning

Article References: Assessing urban heat island patterns in Bharatpur metropolitan city of Nepal using multi-temporal remote sensing and machine learning. (n.d.). https://doi.org/10.1007/s44327-026-00360-7

Image Credits: AI Generated

DOI: 10.1007/s44327-026-00360-7

Keywords: urban heat island, land surface temperature, Bharatpur, Nepal, remote sensing, machine learning, Random Forest, Google Earth Engine, land use land cover, NDMI, Cellular Automata-Markov, urban planning