Research On Forest Change Detection Method Based On GF-2 Images

Forest resources inventory is an important part of forestry management.Accurate collection of forest change information plays an important role in the scientific decision-making of forestry sustainable development and the improvement of forest quality.The characteristics of plantation in south China,especially eucalyptus plantation,such as rapid growth,short rotation period and rapid regeneration,call for a method that can quickly detect the changes of deforestation and regeneration in a short period of time to realize annual monitoring.The development of remote sensing technology provides more rich data sources with high spatial resolution and high temporal resolution,and also provides the possibility to identify the change information more accurately and quickly.The research object was set in the plantation of Shangsi County,Guangxi Zhuang Autonomous Region,where the plantation area changed frequently and rapidly and was under integrated management.The GF-2 remote sensing images of two phases were used as data sources.Based on the change information of forest ground types,three forest change detection methods were successively adopted to study,and three types and regions of change,namely,vegetation changed to bare ground,bare ground changed to vegetation and unchanged type,were extracted:(1)Pixel-level image difference change detection method based on distribution function:Taking the pixel as the detection unit,the image differences of red band,near-infrared band and normalized difference vegetation index(NDVI)were used.The threshold value was determined based on the distribution function.Rapid change detection was carried out and different results were compared and analyzed.(2)Object-level NDVI difference change detection method based on distribution function:Taking the object as the detection unit,the segmentation results of two-phase image were obtained by using multi-scale segmentation and spectral difference segmentation.Then the object-oriented NDVI difference method and the distribution function were adopted to extract the change information.The result was compared with pixel-level NDVI difference method.(3)Object-level based on Partial least square transformation change detection methods:Taking the object as the detection unit,the segmentation results of two-phase image were also obtained by using multi-scale segmentation and spectral difference segmentation.Based on Partial Least Squares(PLS),all the components were extracted,and the difference images were constructed by the component difference.The comprehensive variable Z＿PLS constructed by standard deviation weighting method and all difference variables PLS＿i as input feature were classified by Support Vector Machines(SVM)algorithm respectively.In order to explore the applicability of PLS transformation method in forest change detection,it was compared with the commonly used multivariate change detection(MAD)method.The main conclusions are as follows:(1)In the change detection methods based on the distribution function,the accuracy from the best to the worst are: object-oriented NDVI difference method,pixel-level NDVI difference method,pixel-level red-band difference method and pixel-level near-infrared difference method.Among them,the overall accuracy of two NDVI difference methods is over 85%.(2)Compared with the pixel-level method,the object-level change detection method has more complete change pattern extraction.The 4 m spatial resolution of GF-2 image can well adapt to the change characteristics of the plantations in south China with more and smaller plots.(3)In the detection method of pixel-level and object-level image difference change based on distribution function,the features and algorithm used were simple,which greatly reduced the computational complexity and were easy to implement;Moreover,the method of determining the threshold value by using the distribution function didn’t need to select the training sample,which improved the efficiency.This provides an effective way for rapid forest change detection.(4)After PLS transformation,the correlation between components,variance of difference variables,and interpretation ability of difference variables to the two phases images were completely different in order,and the distribution of change information is uncertain.In addition,information duplication exists between the difference variables due to the fact that the former components T and the latter components U are not orthogonal and the latter components U themselfs are not orthogonal.Therefore,the change information cannot be extracted effectively either by single variable or by the comprehensive difference variable Z＿PLS.While the variables of MAD transformation method are orthogonal to each other,the difference variable MAD＿4 and comprehensive variable Z＿MAD could effectively distinguish the change information.(5)The detection accuracy of object-oriented PLS transformation method using all difference variables PLS＿i directly reached 90.26% and Kappa coefficient was 0.85.It proves that the PLS transformation method is effective.However,the accuracy was slightly lower than that of object-oriented MAD transformation method.