Machine learning algorithms for land cover mapping in a Protected Natural Area

Abstract

Protected natural areas contribute to biodiversity conservation and to climate change mitigation, and provide ecosystem services. Accuracy assessment information on land cover and land use distribution is essential for managing these areas, and Sentinel-2 mission data are well-suited for monitoring them. Therefore, the objective of the study was to compare the performance of four machine learning algorithms —Support Vector Machine (SVM), Random Forests (RF), Gradient-Boosted Decision Trees (GBDT), and Classification and Regression Trees (CART)—, integrating spectral indices and topographic variables. The Sentinel-2 collection and a stratified sample set were used for validation ( n = 641 ). Accuracy was assessed using area-weighted confusion matrices. A two-proportion Z -test was used to compare the algorithms globally, and a McNemar chi-square test was used to compare predictions for each class. The results showed that SVM and GBT had the highest overall accuracy, of 88 and 86, respectively. Comparison of the Z -test algorithms showed that half of the algorithm pairings were statistically different. McNemar's chi-square test showed that 46 of the comparisons by class between paired algorithms were statistically significant ( p ≤ 0.05 ). In conclusion, machine learning algorithms enable the generation of accurate land cover and land use (LCLU) maps. Its implementation in decision-making is recommended due to its ability to recognize complex patterns. Todos los textos publicados por la Revista Mexicana de Ciencias Forestales –sin excepción– se distribuyen amparados bajo la licencia Creative Commons 4.0 Atribución-No Comercial (CC BY-NC 4.0 Internacional), que permite a terceros utilizar lo publicado siempre que mencionen la autoría del trabajo y a la primera publicación en esta revista. https://creativecommons.org/licenses/by/4.0/