Abstract: Land Surface Temperature (LST) serves as a critical indicator for quantifying urban heat islands and informing climate-resilient urban planning strategies, particularly for vulnerable communities. We present a comprehensive open-source benchmark dataset for predicting monthly LST at a spatially consistent resolution of 30m. The dataset encompasses 103 U.S. cities exceeding 90 square miles, providing Digital Elevation Models (DEM), Land Cover classifications, and multiple spectral indices (NDBI, NDVI, NDWI) alongside LST measurements. We implement dual evaluation metrics: LST values and a normalized heat index (1-25) that facilitates cross-city generalization. Our transformer-based baseline achieves 2.6 RMSE for heat index prediction and 9.71F RMSE for LST forecasting across all cities. Notably, we demonstrate that accurate surface temperature prediction can be achieved through a carefully selected subset of geospatial variables when coupled with state-of-the-art vision architectures and appropriate data augmentation techniques.

https://jesseguerrero.github.io/LST-Visualization/

1 Month (F)					1 Month (P)
	Resnet18	Resnet50	B3	B5		Resnet18	Resnet50	B3	B5
Unet	20.74	12.47			Unet	5.17	3.47
DeepLabV3+	21.18	13.34			DeepLabV3+	5.18	3.75
Segformer			10.04	9.71	Segformer			2.69	2.6

3 Month (F)					3 Month (P)
	Resnet18	Resnet50	B3	B5		Resnet18	Resnet50	B3	B5
Unet	22.25	22.36			Unet	5.21	5.19
DeepLabV3+	22.29	22.2			DeepLabV3+	5.21	5.17
Segformer			11.99	11.81	Segformer			2.88	2.86

Paper: --
Code Available Here: https://github.com/JesseGuerrero/benchmark
Checkpoints: https://bit.ly/4jP7z91
Visualization: https://github.com/JesseGuerrero/LST-Visualization

Usage Code:

huggingface-cli download JesseGuerrero/LandsatTemperature --repo-type dataset --local-dir ./Data