A Study Of Monitoring Forest Disturbance Using Deep Autoencoder And Landsat Time Series

Accurate detection of forest disturbance can provide basic data support for global carbon sink estimation and sustainable forest management.For the lacking efficient extraction method of unchanged forest cover in forest disturbance research,our study constructed a multitemporal automatic land cover classification method which can realize the automatic extraction of training samples based on the sample-landcover knowledge included in the land cover classification products.Meanwhile,since it is limited to use the existing forest disturbance monitoring method in areas where historical forest disturbance data is scarce,this study converted forest disturbance detection into sequence anomaly detection in time series data set,and proposed a forest disturbance detection model based on sequence to sequence autoencoder.The main work and conclusions of the research are as follows:(1)The extraction of “anytime” forest is the premise of forest disturbance monitoring and the core approach is multi-temporal land cover classification.In view of the high cost of training sample collection in multi-temporal land cover classification,this study implemented the cross-period migration of sample-ground category relationship and automatic screening of high-confidence samples with the help of sample relationship transfer theory,IR-MAD change detection and local reachable density method.In which,the sample-ground category relationship was derived from existing land cover products.Then we realized the automatic mapping of land cover classification using random forest algorithm in four periods.The overall accuracy of the classification results obtained by this method in all four periods is more than 85%,and the user and producer accuracies of forest cover are greater than 90%.(2)Aiming at the problem that the existing disturbance monitoring algorithm is difficult to apply in the area lacking historical disturbance record,our study tried to solved this problem from the perspective of time series anomaly detection which regarded disturbed forest time series as abnormal sequence in all forest time series dataset.First,we build a semi-supervised learning model based on sequence to sequence autoencoder and attention mechanism.Then this model was trained by undisturbed forest time series samples.Finally,the automatic forest disturbance model was constructed based on the semi-supervised learning model and anomaly detection approach.The overall accuracy of the proposed method is 92.7% and it has strong generalization ability.In the comparative analysis of different algorithms,the proposed model achieved higher detection accuracy of disturbance forest.(3)Based on the proposed forest disturbance detection model,the spatial-temporal changes of forest disturbances in Yangming Mountain area of Hunan Province were analyzed,and the relationship between disturbance occurrence and environmental factors was explored.Forest disturbance occurred in about 25% of the area with constant forest cover in Yangming Mountain,with an area of about 454.88km2.Most of the disturbances were mild and moderate,and the duration lasted for 4-6 years.Combined with the terrain,residential location and road basic data in this area,we found that the lower the elevation(<500m),the closer to the residential area(<9km)and road(<800m),the greater the probability of disturbance.