SPATIAL DISAGGREGATION OF LANDSAT-DERIVED LAND SURFACE TEMPERATURE OVER A HETEROGENEOUS URBAN LANDSCAPE USING PLANETSCOPE IMAGE DERIVATIVES
- 1Training Center for Applied Geodesy and Photogrammetry, University of the Philippines, Diliman, Quezon City, 1101, Philippines
- 2Department of Geodetic Engineering, University of the Philippines, Diliman, Quezon City, 1101, Philippines
Keywords: Urban Heat Islands, Downscaling, Regression-Kriging, Vegetation Indices, Built-up Indices, Gray-level Co-occurrence Matrix
Abstract. Satellite-derived land surface temperature (LST) is frequently utilized to characterize the intensity of urban heat island (UHI) effect in highly urbanized and rapidly urbanizing cities. However, current spaceborne thermal sensors cannot capture temperature variations within heterogeneous urban landscapes at finer scales due to its coarse spatial resolution. This study aims to apply Regression-Kriging (RK) method to downscale a 30-meter Landsat-derived LST to 3 meters using different PlanetScope image derivatives. To avoid multicollinearity, exploratory regression was performed to reduce the initial set of 16 indices to 7 explanatory variables, namely, Enhanced Vegetation Index (EVI), Modified Soil-Adjusted Vegetation Index (MSAVI), Normalized Pigment Chlorophyll Ratio Index (NPCRI), Visible Green-based Built-up Index (VgNIR-BI), Mean, Entropy, and Homogeneity. Ordinary Least Squares (OLS) regression was applied to fit the models and the residuals of the best performing models were interpolated using Ordinary Kriging technique and added back to the downscaled LST. The model with the highest accuracy was obtained using the combination of MSAVI, EVI, and Mean, with an R2 of 0.75 and RMSE of 1.12 °C, 0.58 °C, 0.80 °C, and 1.45 °C in estimating the LST of built-up, bare soil, vegetation, and water classes, respectively. The results indicate that the inclusion of textural features in the regression could improve model accuracy without increasing the variance of coefficient estimates. Moreover, RK method (RMSE = 1.10–1.16 °C) was proven to be a reliable downscaling technique because it redistributes the spatial variability of LST that were not preserved in the OLS regression (RMSE = 1.60–1.75 °C).