Mapping forest-agroforest frontiers in the Peruvian Amazon with deep learning and PlanetScope satellite data

Abstract

Monitoring complex and dynamic land systems such as tropical agroforests using remote sensing presents a significant challenge in ecological research. Traditional mapping methods are hindered not only by spectral similarity between agroforests and forests, but also by the spatial heterogeneity of forest-agroforest frontiers and the high data demand at large scales is an additional challenge. In this study, we aim to develop a modeling framework to distinguish between forests, secondary forests, agroforests (e.g. shade-grown perennials), and non-tree agricultural classes (e.g. active cropland, grassland, young fallow) in the Peruvian Amazon. To achieve this, we combine deep learning and remote sensing data, including 3-m PlanetScope satellite imagery, a Digital Elevation Model (DEM), and temporal data from the Landtrendr change detection algorithm. We conducted a sequence of modeling experiments involving different complexity of the data inputs and output classes, with overall accuracies ranging from 28.6 % to 82.9 %. Integrating a DEM as an additional helped the generalization of models across different geographical sites but did not improve the overall accuracy, whereas adding temporal information did not improve generalization or accuracy. Challenges arise in accurately identifying successional land cover types, particularly young fallow, which exhibits spectral similarity to other classes. Reducing the target classes from seven to four was found to considerably improve the accuracy of the predictions. Our findings contribute to distinguishing agroforests from forests at a large scale, providing insights into previously undetected tree-covered land uses and thus informing on sustainable ecosystem management. Yet, our results underscore the limitations of remote sensing in heterogeneous forest-agriculture landscapes and emphasize the need for further research to address persistent challenges and improve classification accuracy for monitoring global environmental change.