Comparison of spatial and non-spatial logistic regression models for modeling the occurrence of cloud cover in north-eastern Puerto Rico

We compared a non-spatial logistic regression model (LRM) with two spatial models, i.e., logistic mixed model (LMM) and geographically weighted logistic model (GWLM), to model/predict the occurrence of orographic cloud cover at the Luquillo Experimental Forest in north-eastern Puerto Rico. The cloud cover data were derived from two Landsat images on a relatively clear day and a relatively cloudy day, respectively. In these models, the response variable was binary with 0 representing non-cloudy areas and 1 representing cloudy areas. The covariates included slope, aspect, and the difference between elevation and lifting condensation level. The results indicated that the spatial LMM did not improve the prediction of probability of cloud cover over the non-spatial LRM. One possible explanation is that the LMM was not able to account for the anisotropy in the non-spatial LRM residuals. In contrast, the GWLM with the bandwidth close to the effective range of the semivariogram of the LRM residuals showed the best model fitting among the three types of models, resulting in the lowest Akaike Information Criterion (AIC) and sum of squared errors (SSE), and the smallest spatial autocorrelation and heterogeneity in the model residuals. The GWLM model coefficients were spatially nonstationary, ranging from negative to positive depending on locations. The significance of the covariates also varied spatially. Our study demonstrated that GWLM is a useful and effective tool to account for spatial heterogeneity in modeling/predicting the occurrence or probability of cloud cover, using the spatial data derived from satellite imagery and GIS. (c) 2012 Elsevier Ltd. All rights reserved.