Algoritmos de Clasificación Automática para el Ordenamiento Territorial de los Bosques Nativos de la Provincia de Buenos Aires

Abstract

The objective of this work is to update and improve the existing cartography on the spatial distribution of native forests in the Buenos Aires province, Argentina. It was developed within the framework of a project executed by researchers of the national scientific and technological system for the Dirección de Bosques of the Ministerio de Ambiente de la Provincia de Buenos Aires, focusing on the design and execution of a uniform methodology for the identification and delineation of native forest covers for the entire province. The cartographic update used a combination of field-collected information, complemented with visual interpretation of multitemporal series of high-resolution optical images, generalized using machine learning algorithms fed with information derived from Sentinel-2 multispectral satellite image series. The products obtained were evaluated using metrics derived from the contingency matrix, calculated from field-labeled data. The methodology used for automatic classification is detailed, including the methodology for labeling training points, the spectral information used to feed the classifiers, the selection of the classification methodology itself, as well as the details of the post-processing procedure applied to each specific forest formation and the evaluation of the final products obtained. The delineation obtained excludes 235182 ha from the 968397 ha of the current map, which we consider to not correspond to native forests, and incorporates 187512 ha of native forests that had not been previously mapped, reducing the total mapped area of native forests in the province by 4.9%. The evaluation, carried out with 719 field-labeled points, assigns an overall accuracy of 0.89 and a kappa index of 0.85 to the classification obtained, indicating that the proposed methodology is suitable for the delineation of native forests in the province.