Spatial Downscaling of Land Surface Temperature Based on a Multi-Factor Geographically Weighted Machine Learning Model

Abstract

Land surface temperature (LST) is a critical parameter of surface energy fluxes and has become the focus of numerous studies. LST downscaling is an effective technique for supplementing the limitations of the coarse-resolution LST data. However, the relationship between LST and other land surface parameters tends to be nonlinear and spatially nonstationary, due to spatial heterogeneity. Nonlinearity and spatial nonstationarity have not been considered simultaneously in previous studies. To address this issue, we propose a multi-factor geographically weighted machine learning (MFGWML) algorithm. MFGWML utilizes three excellent machine learning (ML) algorithms, namely extreme gradient boosting (XGBoost), multivariate adaptive regression splines (MARS), and Bayesian ridge regression (BRR), as base learners to capture the nonlinear relationships. MFGWML uses geographically weighted regression (GWR), which allows for spatial nonstationarity, to fuse the three base learners’ predictions. This paper downscales the 30 m LST data retrieved from Landsat 8 images to 10 m LST data mainly based on Sentinel-2A images. The results show that MFGWML outperforms two classic algorithms, namely thermal image sharpening (TsHARP) and the high-resolution urban thermal sharpener (HUTS). We conclude that MFGWML combines the advantages of multiple regression, ML, and GWR, to capture the local heterogeneity and obtain reliable and robust downscaled LST data.

gwverse: A Template for a New Generic Geographically Weighted R Package

“…Again this reflects a desire to explore spatial variation in model parameters or its components and to move away from global, “whole map” approaches. Examples include GW principal components analysis (PCA) (Harris, Brunsdon, and Charlton 2011), GW descriptive statistics (Brunsdon, Fotheringham, and Charlton, 2002), GW discriminant analysis (Brunsdon, Fotheringham, and Charlton, 2007; Foley and Demšar, 2013), GW correspondence matrices (Comber et al 2018), GW structural equation models (Comber et al 2017), GW evidence combination (Comber et al, 2016), GW Variograms (Harris, Charlton, and Stewart Fotheringham, 2010), GW network design (Harris et al, 2014), GW Kriging (Harris, Charlton, and Stewart Fotheringham, 2010; Harris, Brunsdon, and Stewart Fotheringham, 2011), GW visualization techniques (Dykes and Brunsdon, 2007), and more recently GW artificial neural networks (Du et al, 2020; Hagenauer and Helbich, 2022) and GW machine learning (Chen et al, 2018; Li, 2019; Quiñones, Goyal, and Ahmed, 2021; Xu et al, 2021). In each of these developments, the moving window or kernel is still used to generate local data subsets that are weighted by their distance to the kernel center, as is done in GWR, thereby providing local inputs to the model, analysis or evaluation being applied.…”

Predição do potencial radioativo de queimadas: comparando duas técnicas mediante regressão espacial

“…No tocante à região da Mata Atlântica Paraibana (uma das mesorregiões do Estado da Paraíba), não há ainda trabalhos de análises preditivas no que diz respeito a questões de incêndios florestais. Conforme o levantamento realizado para este trabalho, como descrito em Xue et al (2021) e Xu et al (2021), modelos de análise preditiva podem auxiliar na tomada de decisões e, no caso de dados geográficos, podem ser gerados a partir de métodos de Regressão Geograficamente Ponderada (RGP). RGP é um método de regressão linear que considera, para os valores de cada local de amostra, os valores do local de amostra da vizinhança, para o oferecimento de parâmetros, conforme (Dale and Fortin, 2014), que podem ser utilizados como base para a elaboração de mapas temáticos úteis aos gestores e responsáveis.…”

Reducing Scaling Effect on Downscaled Land Surface Temperature Maps in Heterogenous Urban Environments

“…The machine learning-based methods may include various artificial neural networks, support vector machines, random forest models, and partial least square models (e.g., [1,9,[17][18][19][20]). They usually perform better compared to the regression-based methods and result in a high accuracy by fitting nonlinear relationships between LST and independent variables [21].…”