Classificação orientada a regiões na discriminação de tipologias da floresta ombrófila mista usando imagens orbitais ikonos

Abstract

Since the 70 ́s remote sensing techniques have been used for surveying of natural resources. From the 90 ́s on, the launching of new satellites carrying high-resolution sensors has led to the implementation of new approaches in digital image processing. Forest mapping is one of the basic tools to accurately assess forest conditions in forest remnants, thus allowing for the establishment of strategies, which aim both to nature conservancy as to the economic development of a real state or region. This study evaluated the possibility of identifying and discriminating forest types in remnants of Araucarian forests, aiming to develop a methodology for mapping the remnants of this biome in a fast, inexpensive and accurate way. The research was developed at the Forest Reserve EMBRAPA/EPAGRI, located in Caçador-SC. Forest types were defined by surveying target areas on the ground. Segmentation and region-oriented classification algorithms were tested on an Ikonos image in order to describe the forest conditions by the time of image acquisition. Thematic accuracy was evaluated by comparing the classification results with a reference map obtained through on-screen visual interpretation of the same Ikonos imagery. The mapping classes were based on the presence of species indicating the successional phases of the woody vegetation (trees and shrubs) in canopy cover. The two-level classification scheme considered the successional phases as well as the forest types in a more detailed manner. Thirteen thematic classes were defined and mapped by visual interpretation. Eight of then referred to forest types. Qualitative and quantitative analyses were performed in order to define the best minimum area and similarity thresholds in the segmentation process. The quantitative analysis included the development of a modified IAVAS index. This index allowed for the comparison between different area and similarity thresholds thus eliminating the subjectiveness of a qualitative analysis in defining the best combinations. Among the tested threshold pairs, the best one was the 35 (similarity) and 1200 (area). The regions generated by this pair of thresholds were submitted to a classification process using the algorithms “Isoseg” and “Bhattacharyya”, available in software SPRING. In the classification scheme the number of classes was reduced to 11 due to the non-discrimination of a class refering to a forest type and the grouping of two classes refering to land use. The supervised digital classification was efficient in determining the forest type “Predominance of Araucaria”. For the other classes the Bhattacharyya classifier didn ́t perform well, generating low values for the overall accuracy (51.73%) and for the kappa index (0.43).