Exploring EnMAP hyperspectral images and ensemble deep learning for classifying forest land-cover types in Brazil

Spatially explicit information on forest land-cover types is critical for effectively managing and conserving forest ecosystems and monitoring restoration initiatives. This study delves into the potential of the Environmental Mapping and Analysis Program (EnMAP) hyperspectral satellite for distinguishing between tropical forest land-cover types in São Paulo, Brazil. We developed and evaluated a novel ensemble deep learning approach, integrating a 1D Convolutional Neural Network (CNN), an autoencoder, and a Recurrent Neural Network (RNN), for pixel-wise classification. We also assessed the impact of spectral resolution by comparing EnMAP data with simulated Landsat 8 and Sentinel-2A data, and investigated the influence of band selection on classification accuracy. The study focused on mapping forest land-cover types, including conserved remnants, natural regeneration, restoration plantations, and monocultures. Our findings reveal significant distinctions in spectral responses across these forest land-cover types. The ensemble deep learning model significantly outperformed individual deep learning models and traditional Support Vector Machine (SVM) classifiers, achieving the highest weighted F1-Score of 0.700. EnMAP data further achieved significantly higher F1-scores than Sentinel-2A and Landsat-8 (p < 0.05), with the greatest gains observed for natural regeneration and restoration plantations. Interestingly, band selection did not universally improve classification performance for the deep learning models, and in some cases, led to a decrease in accuracy. While monocultures were classified with high accuracy, misclassifications were observed between natural regeneration and restoration plantations, and between conserved remnants and natural regeneration, reflecting their spectral similarities. This study provides valuable insights into the effective use of space-borne hyperspectral imagery and ensemble deep learning for robust forest land-cover type mapping in tropical regions, emphasizing the critical role of high spectral resolution.