Area-to-point geographically weighted regression kriging for downscaling MODIS images
No fieldwork is required
For years, MODIS images have been instrumental in divers environmental and ecological studies such as monitoring deforestation, land-us/land-cover (LULC) changes, agriculture and epidemiology. Monitoring of certain environmental processes induced by human activities require in-depth/detailed information spectrally and spatially than that originally provided by MODIS. When bands 1&2 are upscaled from 250 m resolution to 500 m and fused together with the 500 m resolution of bands 3-7, a 7 band 500 m resolution is obtained. This has been the common case for data fusion. While spectral advantage is gained, spatial detail is lost. Conversely, downscaling bands 3-7 from 500m to 250 m gains spectral advantage of 7 bands and spatial detail of 250 m.
To gain both spectral and spatial advantage, this study develops a framework of area-to-point geographically weighted regression kriging to downscale MODIS images.
The methodology combines deterministic and stochastic modelling. The deterministic modelling will seek to derive linear association between target images (bands 3-7; 500m) and the explanatory images (bands 1&2, 500 m). Here, unlike simple linear regression which assumes global parameters, it is the aim to adopt geographically weighted regression to estimate local parameters. Next is to downscale the residuals from geographically weighted regression from 500 m to 250 m resolution. This is not straight forward since the semivariance is a function of the spatial support. Hence area-to area and area-to-point semivariograms should be deconvolved from the variogram model before applying ordinary kriging on the extracted residuals. The method should be tested on simulated datasets and validated before applying on real data.