Predicting the spectral information of future land cover using machine learning

Application of machine learning models to study land-cover change is typically restricted to the change detection of categorical, i.e. classified, land-cover data. In this study, our aim is to extend the utility of such models to predict the spectral band information of satellite images. A Random Forests (RF)-based machine learning model is trained using topographic and historical climatic variables as inputs to predict the spectral band values of high-resolution satellite imagery across two large sites in the western United States, New Mexico (10,570 km²), and Washington (9400 km²). The model output is used to obtain a true colour photorealistic image and an image showing the normalized difference vegetation index values. We then use the trained model to explore what the land cover might look like for a climate change scenario during the 2061–2080 period. The RF model achieves high validation accuracy for both sites during the training phase, with the coefficient of determination (R²) = 0.79 for New Mexico site and R² = 0.73 for Washington site. For the climate change scenario, prominent land-cover changes are characterized by an increase in the vegetation cover at the New Mexico site and a decrease in the perennial snow cover at the Washington site. Our results suggest that direct prediction of spectral band information is highly beneficial due to the ability it provides for deriving ecologically relevant products, which can be used to analyse land-cover change scenarios from multiple perspectives.