Image-to-map Conflict Detection Using Iterative Trimming : Application to Forest Change

Large scale vector databases are valued tools for forest management. It is therefore important to keep these databases up to date and various change detection methods have been designed in this aim. Recently, object-based iterative trimming was successfully used to detect change in temperate and tropical forests. The goal of the present study is to transfer this image-to-image method in an imageto-map application. This study focuses on the detection of clear cuts and forest regeneration areas in a multi-spectral Quickbird image. Various steps were necessary to bridge the gap between this image and the vector database. In order to reduce the effects of residual parallax, the vector database was modified along forest boundaries using the viewing parameters of the satellite. Besides, the image was segmented with a large homogeneity constraint in order to produce ”pure” image-objects. Eventually, the resulting image-objects were automatically labeled using the information from the modified vector database, and the trimming algorithm was run for each forest class. The hypothesis behind iterative trimming is that objects belonging to the same class share similar characteristics (e.g. spectral reflectance). In other words, they belong to the same distribution. The class distribution was estimated using a non parametric method in order to fit to the data even with complex distributions. The chosen method used kernel density estimates to build the probability density function. Outliers were excluded based on a density threshold and the new parameters of the distribution were reprocessed until the all objects are above the new threshold. The resulting outliers included the majority of the discrepancies between the image and the map in the forest areas. About 50 % of the forest regeneration and 100 % of the clear cuts were properly detected. It is a promising way to improve semi-automated map updating because the training dataset is the vector database itself. However, further work is needed to test the method on other land cover types and to move from the detection toward the classification of the discrepancies.