Projecting Australia's forest cover dynamics and exploring influential factors using deep learning

This study presents the first application of deep learning techniques in capturing long-term, time-continuous forest cover dynamics at a continental scale. We developed a spatially-explicit ensemble model for projecting Australia's forest cover change using Long Short-Term Memory (LSTM) deep learning neural networks applied to a multi-dimensional, high-resolution spatiotemporal dataset and run on a high-performance computing cluster. We further quantified the influence of explanatory variables on the spatiotemporal dynamics of continental forest cover. Deep learning greatly outperformed a state-of-the-art spatial-econometric model at continental, state, and grid-cell scales. For example, at the continental scale, compared to the spatial-econometric model, the deep learning model improved projection performance by 44% (root-mean-square error) and 12% (pseudo R-squared). The results illustrate the robustness and effectiveness of the LSTM model. This work provides a reliable tool for projecting forest cover and agricultural production under given future scenarios, supporting decision-making in sustainable land development, management, and conservation.