Research On Rapid Monitoring Method Of Forest Disturbance Based On Sentinel-1 SAR Time Series

Forest is one of the important terrestrial ecosystems,and its disturbance by natural and human factors profoundly impacts the stable development of forest ecosystems and the normal functioning of ecological services.Therefore,it is of great significance to explore a large-scale and high-timeliness method for the rapid acquisition of forest disturbance information.Compared with optical remote sensing technology,which has been widely used in forestry monitoring,synthetic aperture radar(SAR),as an active remote sensing technology that can operate in all weather conditions at any time of day or night,can make up for the shortage of optical remote sensing imaging restricted by cloudy and rainy weather,and its advantages are especially obvious in the tropical areas where it is wet and rainy all year round.However,the amount of space-borne SAR remote sensing is relatively small.Its backscattering signal will be affected by many complex factors such as topography,substrate composition,canopy,and soil water content.There are still a series of problems to be solved urgently in the rapid monitoring of forest disturbance.In this study,two common types of forest disturbances,forest fires and deforestation,are studied,and the scale of forest fires is further distinguished.Three areas with cloudy and rainy weather,Xichang Lushan,China,Jambi Province,Indonesia,and Riau Province,Indonesia,are selected as the study area.This study explores the applicability of the backscattering and spatial features of Sentinel-1 SAR time series images in rapid forest disturbance monitoring,and proposes a rapid forest disturbance monitoring model based on machine learning and Sentinel-1 SAR time series data,and conducts an in-depth analysis of the model’s timeliness,sensitivity,stability,and generalization performance,which provides a reference for the application and promotion of the model in the rapid monitoring of forest disturbance.The specific research content and results are as follows:(1)This study reveals the response characteristics of time series SAR backscattering features and spatial features in two types of forest disturbance monitoring,forest fire,and deforestation,and constructs a forest disturbance monitoring database in the study area.The study shows that the SAR backscatter coefficients change abnormally when forest fire and deforestation disturbances occur,in which the VH polarization mode changes more significantly than the VV polarization mode,and the mean and angular second-order moments(ASM)of the spatial features respond to the disturbances more significantly than other texture indicators.The study shows that the SAR backscatter coefficients change abnormally when forest fire and deforestation disturbances occur,in which the VH mode changes more significantly than the VV mode,and the mean feature and ASM feature respond to the disturbances more significantly than other texture indicators.(2)Four forest disturbance monitoring models,random forest,GBDT,XGBoost,and LightGBM,were constructed and compared in each research area,and the optimal model was selected.The research shows that the disturbance monitoring accuracy of the four models in different research areas are all above 80%,among which the LightGBM model performs best,with accuracy rates of 87.36%,93.71%,and 97.66%in the three research areas respectively.The LightGBM model was selected as the optimal model for further studies.(3)In this study,the contributions of SAR spatial features and terrain features to the model were evaluated,and the LightGBM-based forest disturbance monitoring model was analyzed in terms of timeliness,sensitivity,stability,and generalization.The study shows that(i)the addition of both SAR spatial features and terrain features helps to improve the model accuracy,and the improvement of spatial features is higher than that of terrain features,and there are differences in the three regions,and the addition of the two features in Lushan,Xichang leads to a higher improvement of the model accuracy,reaching 12.12% and 2.27%,respectively;(ii)the timeliness of deforestation disturbance monitoring in Riau province,Indonesia is high,and the accuracy of the first observation after disturbance can reach 97.22%,while the timeliness of forest fire disturbance monitoring in Xichang Lushan and Jambi Province,Indonesia is lower,but still can reach 85.87% and 97.14% respectively in the third observation after the disturbance;(iii)The sensitivity is evaluated by the accuracy of different backscattering variation magnitudes at the first observation after the disturbance occurs.The greater the change in the backscatter coefficient,the higher the sensitivity of the model,and the monitoring accuracy is higher than 80% when the change is greater than 3 db;(iv)For multiple observations after the disturbance occurs,the model gives stable monitoring results,and the correct rate of the next detection after the first detection is higher than 89%;(v)The model has good generalization performance,and its accuracy reduction can be controlled within 10%～20% if the local data set is not included in the training.