Study Of Forest Cover Change Detection Methods Using Multi-Polarization Space-borne SAR

Remote sensing images are a kind of important data source for land cover/forest cover change detection application because it can be used to observe the earth in large scale and in short repetition time.In the cloudy forest cover area,synthetic aperture radar(SAR)can compensate for the limitation of optical image effectively by observing earth all-day and all-weather.Studies on space-borne SAR remote sensing observation and application technology have been developed rapidly in China.The in-orbit GF-3 C band SAR satellite is of high repeating frequency and can acquire multi-polarization SAR data in different spatial resolution,showed unique advantage for regionally forest resources dynamic monitoring.However,the method of multi-polarized SAR change detection is mostly developed for land surface cover changes.Although a few of studies on "forest cover" change detection have been reported,all of them apply the method of masking "land cover" change detection result using archived forest/non-forest cover map,lacking of systematic and in-depth research on "forest cover" change detection.It is not only imminent to carry out the relevant research,but also can provide technical support for dynamic monitoringand operational application of national forest resources using space-borne SAR,being of important application value.This study took the site covered by two scenes of dual polarization(HH,HV)ALOS PALSAR data in different imaging dates as research area.The bi-temporal Landsat-5 and Google Earth images,whose imaging dates are as close to the corresponding PALSAR data as possible,were visually interpreted to produce the forest cover change reference map for validating the method developed in this paper.Five “land cover” change detection methods,including four automatic threshold determining methods,such as OTSU,LN-GKIT,WR-GKIT and NR-GKIT,and another method based on maximum likelihood classification(MLC),were investigated.Furthermore,the idea for extracting “forest cover” changes map from the “land cover” change map were implemented with the non-forest cover map through forest and non-forest classification of the bi-temporal HV polarization SAR image.The performance of the above five methods were evaluated,accuracy evaluation results show that these methods are of low accuracy,and it is difficult to keep both of False-Alarm rate and Missed-Alarm rate to low level.Among them,the two threshold segmentation methods,LN-GKIT and WR-GKIT,can achieve relative better results.One forest cover change detection method using multi-polarization SAR data(EM-MRF-FC)was developed by combining the “Change detection method based on bi-temporal Forest cover Classification(CBFC)” and the “Bayesian Maximum Expectation-Markov Random Field Classification(EM-MRF)change detection method”.Firstly,each of the bi-temporal multi-polarization SAR images was separately classified into forest and non-forest map by threshold segmentation method,so as to produce the initial forest cover changed map(IFCM).Secondly,the IFCM was input into the Fisher transformation of multi-polarization SAR ratio image and EM-MRF classification algorithm as training data.To compose the Markov energy function,the pixel class conditional probability was combined with the potential function in the Markov random field by a certain weight.Then,the posterior probability of the forest cover “changed” and “unchanged” pixel can be calculated.The class attribute was determined according to the maximum posterior probability criterion,and the iteration stopped when the condition was met.Experiment results show that: EM-MRF-FC utilizes the dual-polarization information and reduce the impact of speckle noise on the forest cover change detection by the spatial context information extraction.Its detection performance was better than EM-MRF-LC-NFM by masking EM-MRF-LC change detection result using non-forest cover map(NFM)and forest-non-forest post-classification change detection method(CBFC);The coverage area and boundaries of the pixels detected as forest cover changedwere most similar to the reference forest cover changed map,and the False-Alarm rate(0.87%)and Missed-Alarm rate(12.44%)can both be kept to low level;In addition,the method developed only needs the initial forest cover changed map as input data,owning the advantages of running semi-automatically.