Abstract
Ecological footprint (EF) plays an important role in ecological and geographical analysis, but it can only be calculated based on statistical data in a region like a country or a city. High-resolution mapping of ecological footprint is in urgent need for fine-grained analysis of carbon emission and resource consumption. However, current downscaling methods, based on classical statistical models, cannot achieve satisfactory results, because most of them neglect the huge gap of scale between training samples and predicting ones, trying to estimate gridded ecological footprint with only one black-box model. To solve this problem, this paper proposes an innovative AI method for high-resolution mapping of ecological footprint, namely GeoUNet, which can accomplish an end-to-end multi-scale prediction using multi-source datasets. Our experiments indicate that GeoUNet can surpass mainstream downscaling methods in mean square error (MSE) and upscaling mean square error (UMSE). High-resolution gridded EF results obtained by GeoUNet can reveal the spatial heterogeneity of ecological footprint and can be used in fine-scale spatio-temporal analysis.
