Hierarchical mapping of Brazilian Savanna (Cerrado) physiognomies based on deep learning

A groundbreaking study achieved unprecedented accuracy in mapping the detailed physiognomies of the Brazilian Cerrado, a global biodiversity hotspot. Using deep learning techniques with an adapted U-net architecture, researchers classified 25 vegetation types into three major ecosystem groups—forest, savanna, and grassland—with overall accuracies exceeding 90%. The best performance was obtained using RGB bands combined with the enhanced vegetation index (EVI2), achieving 92.8% accuracy for the first-level classification and over 85% for detailed physiognomies in savanna and grassland. This marks a significant advancement over traditional machine learning methods, offering a robust tool for conservation efforts in protected areas like Chapada dos Veadeiros and Emas National Parks.

Kasahara Neves, A., Körting, T. S., Fonseca, L. M. G., Soares, A. R., Girolamo-Neto, C. D., & Heipke, C. (2021). Hierarchical mapping of Brazilian Savanna (Cerrado) physiognomies based on deep learning. Journal of Applied Remote Sensing, 15(04), 1–23. https://doi.org/10.1117/1.jrs.15.044504