Abstract: [Objective] Collecting forest-change information accurately, quickly and efficiently to update forest resource databases in time is critically important for scientific decision-making of forest management and forestry sustainable development, and has long been one of the major technical challenges in forest resource management.[Method]In this paper, an object-based statistical method was applied to detect forest changes in a study site located in Shangsi County, Guangxi Zhuang Autonomous Region where the forest cover has experienced frequent changes over time associated with a large number of parcels with relatively small sizes. High spatial resolution satellite images of ZY-3, GF-1 and the vector sub-compartment map of forest distribution were served as the data sources. Object features were extracted by using multi-resolution segmentation of satellite images accompanying with the thematic map.As the frequencies of the difference of the objects’ mean gray value and standard deviation of the two-temporal images were approximate the normal distributions, chi-square distribution statistic was using with the difference between object mean value and standard deviation. Therefore the change objects with abnormal statistics were flagged one by one with a statistical procedure of iterative trimming.[Result]The results indicated that for a given confidence level, the detected number of change objects decreased rapidly as the iteration number increased. When the confidence level was set at 0.95, 0.98, 0.99 and 0.999, the corresponding changed objects were labeled through 25, 23, 20 and 15 iterations, respectively. Along with changes in the confidence levels from 0.95 to 0.99, the omission rate increased,as a result, the commission rate decreased and the overall accuracy increased.The best result was achieved with the confidence level was 0.99 as the overall accuracy was 92.6% and the Kappa was 0.764 8.[Conclusion] The advantage of the method proposed in this paper was object-based and statistically driven, directly detected forest changes from the dataset generated from the image segmentation, did not rely on external information and human intervention, could be executed automatically in the entire procedure, and generated satisfactory results.Furthermore,this approach was potentially an ideal method for forest change detection.

Key words: forest change detection, image segmentation, statistical test, chi square distribution