Extraction Of Forest Burned Area Based On Stacked ConvLSTM And Remote Sensing Image Time Series

As the largest land type,forest plays a key role in human living environment,biological habitat and global carbon cycle.The damage and loss of forest caused by forest fire is the first in many types of forest natural disasters.Monitoring the time point and spatial range of the burned area in the remote sensing image time series for the first time after the occurrence of forest fire is very important for forest damage assessment,management,carbon accounting and forest restoration management.There is a certain continuity in the spatial distribution of forest burned area in remote sensing image.The existing detection methods based on pixel level are studied more.The spatial information existing in remote sensing image time series is not fully utilized,and the detection results contain many false alarm pixels,so complex post-processing is needed to suppress the influence of false alarm pixels.In this study,the spatiotemporal prediction model Stacked ConvLSTM is used to detect the time point and spatial range of forest burning sites for the first time in the time series of remote sensing images.On the basis of maintaining the good spatial continuity of the results,the subjective postprocessing operation is avoided,and the extraction accuracy of forest burning sites is improved.The main research results of this study are as follows:(1)In this study,the learning ability and prediction accuracy of different Stacked ConvLSTM network structures in forest remote sensing vegetation correlation index image time series was quantitatively analyzed and evaluated.The effects of stacking layers,the number of units in each layer and the size of convolution kernel on the prediction accuracy are tested and analyzed,and the optimal network structure for the prediction of forest remote sensing image vegetation correlation index time series is obtained,which provides experimental verification and basis for the performance of different stacked convlstm network structures in learning the change law of forest remote sensing image time series.(2)Based on different vegetation related index time series,the forest fire monitoring results of Stacked ConvLSTM model,Stacked LSTM model and bfast algorithm are compared,and the global standard fire product is added to analyze and evaluate qualitatively and quantitatively.The results show that: compared with Stacked LSTM model and bfast algorithm,the detection results of Stacked ConvLSTM model are more complete and continuous in space,and the accuracy indexs of Stacked ConvLSTM model,Stacked LSTM model,bfast algorithm and fire algorithm are compared and analyzed with Fire＿CCI5.1 product in different vegetation related indexes of the two regions,results show that precision,accuracy and F1 index of Stacked ConvLSTM model are relatively higher,and the overall performance is better,and can reduce the false alarm information of the burned area to a certain extent.(3)Compared and analyzed the extraction effect of Stacked ConvLSTM in four forest burned areas with different vegetation related indexes.Results show that the detection effect of BAI and NBR was better than that of EVI and NDVI.