Forest change detection by statistical object-based method

Abstract

Forest monitoring requires more automated systems to analyse the large amount of remote sensing data. A new method of change detection is proposed for identifying forest land cover change using high spatial resolution satellite images. Combining the advantages of image segmentation, image differencing and stochastic analysis of the multispectral signal, this OB-Reflectance method is object-based and statistically driven. From a multidate image, a single segmentation using region-merging technique delineates multidate objects characterised by their reflectance differences statistics. Objects considered as outliers from multitemporal point of view are successfully discriminated thanks to a statistical procedure, i.e., the iterative trimming. Based on a chi-square test of hypothesis, abnormal values of reflectance differences statistics are identified and the corresponding objects are labelled as change. The object-based method performances were assessed using two sources of reference data, including one independent forest inventory, and were compared to a pixel-based method using the RGB-NDVI technique. High detection accuracy (> 90%) and overall Kappa (> 0.80) were achieved by OB-Reflectance method in temperate forests using three SPOT-HRV images covering a 10-year period.