Change Detection in Forests and Savannas Using Statistical Analysis Based on Geographical Objects

Abstract

Abstract:The aim of this work was to assess techniques of land cover change detection in areas of Brazilian Forest and Savanna, using Landsat 5/TM images, and two iterative statistical methodologies based on geographical objects. The sensitivity of the methodologies was assessed in relation to the heterogeneity of the input data, the use of reflectance data and vegetation indices, and the use of different levels of confidence. The periods analyzed were from 2000 to 2006, and from 2006 to 2010. After the segmentation of images, the descriptive statistics average and standard deviation of each object were extracted. The determination of change objects was realized in an iterative way based on the Mahalanobis Distance and the chi-square distribution. The results were validated with an early visual detection and analyzed according to Receiver Operating Characteristic (ROC) Curve. Significant gains were obtained by using vegetation masks and bands 3 and 4 for both areas tested with 94,67% and 95,02% of the objects correctly detected as changes, respectively for the areas of Forest and Savanna. The use of the NDVI and different images were not satisfactory in this study.