Differentiating between natural and agricultural trees using remote sensing is essential for assessing ecosystem services, commodity-driven deforestation, and restoration progress. Existing approaches focus on identifying a single tree commodity, rather than a system classification that is agnostic to species. This study presents a transfer learning approach to classify tree-based systems, leveraging extracted spatial embeddings from a high-performing neural network to improve classification accuracy in label-scarce environments. We applied a CatBoost classifier to a combination of Sentinel imagery, gray-level co-occurrence matrix texture features, and extracted spatial embeddings to classify four land use classes: natural, agroforestry, monoculture, and other (background). Through comparative modeling and feature selection exercises, we validate performance gains resulting from transfer learning and texture features. Building on previous efforts to model tree extent across the tropics (Brandt et al. 2023), we explore whether the spatial features extracted from Brandt et al.’s (2023) convolutional neural network can be repurposed to help classify tree-based systems. This method is demonstrated for 26 priority districts in Ghana, resulting in a 10-meter resolution land use map for 2020. Our findings suggest the spatial embeddings extracted from Brandt et al.’s (2023) tree cover model offer value beyond their original task and represent a scalable path forward for broader monitoring efforts.

Datasets: