Introduction

Background and Rationale

Human development has altered natural landscapes, leading to changes in land cover across the globe. Agricultural and urban development replaces natural ecosystems with anthropogenic land cover types such as urban areas, roads, and croplands. As climate change continues to reshape environmental conditions, there is growing interest in identifying potential future ecosystem states—and restoring landscapes to more climate-appropriate, natural conditions where feasible (Harris et al., 2006; Hobbs et al., 2013).

Machine learning approaches offer powerful tools for modeling complex environmental relationships and making predictions at scale. Random Forest (RF) and Deep Neural Network (DNN) algorithms have emerged as robust classifiers for ecological and land cover applications (Belgiu & Drăguţ, 2016; Hengl et al. 2018). RF is known for its ability to handle nonlinear relationships and noisy datasets with minimal parameter tuning, while DNNs can capture complex, high-dimensional interactions between variables (LeCun et al., 2015). In this study, we develop RF and DNN models to reclassify human-developed land cover areas based on climatic variables. By training our models on observed climate and land cover relationships in undeveloped areas, we can then apply these models to developed areas to predict the corresponding natural land cover classification.

Reclassifying human-modified landscapes based on climate will allow for a clearer understanding of the ecological potential of these areas under current climate conditions and facilitate making land cover predictions based on projected climates. These models and predictions are relevant for conservation planning, habitat restoration, and developing climate adaptation strategies (Millar et al., 2007; Seddon et al., 2016). Comparing these predicted natural land cover types with existing human-developed classes can provide insight into the degree of ecological transformation and help prioritize regions for ecological recovery or protection (Strassburg et al., 2020).

This work contributes to a growing body of research using machine learning to inform landscape management and restoration ecology. By applying climate-based predictive models to human-altered landscapes, we aim to provide spatially explicit insights into the natural land cover patterns that would be expected in the absence of development.