基于VGG‒Unet模型的黄土地震滑坡自动识别方法OA

Objective Loess areas are highly prone to frequent earthquake-induced landslides.Rapid and accurate identification in loess regions is essential to obtain a detailed spatial distribution of loess earthquake landslides.The identification results of landslides reflect their development status and spatial distribution,serving as a theoretical foundation for research on landslide safety,vulnerability,and management.Field investigations and re-mote sensing technology are two primary approaches for landslide identification.However,relying solely on field surveys consumes substantial human,material,and financial resources.Similarly,depending exclusively on remote sensing technology for landslide identification does not en-sure the authenticity and reliability of the database.This research applies the VGG‒Unet network for the automatic identification of loess earth-quake landslides based on landslide data obtained from field surveys. Methods Firstly,based on loess earthquake landslides investigated by the earthquake landslide research team of the College of Disaster Preven-tion Science and Technology during field surveys in Gansu Province and Ningxia Hui Autonomous Region,the quantitative and qualitative pa-rameters for extracting loess earthquake landslides on the satellite imagery platform through visual interpretation were systematically summa-rized.The quantitative parameters included stratigraphy,topography,slope,slide length,and slide width characteristics,and the qualitative param-eters included planar morphology,profile,and color tone characteristics.These features supported the visual interpretation of remote sensing satel-lite imagery and facilitated the extraction of a database of loess landslides triggered by seismic events.Ultimately,a total of 494 original images were extracted,covering 1 052 loess landslides induced by seismic events,which served as input images to improve the model's generalization ability.Secondly,due to the exceptional feature extraction capability of the VGG16 network,its simple yet effective structural design,and its ex-tensive pre-trained weights,it facilitated the construction of a Unet network model using it as the primary backbone feature extraction network,enhancing performance.Therefore,this study applied the VGG‒Unet network model using the Python programming language based on the Py-Torch framework under the Windows 10 environment,with a GPU of NVIDIA Quadro P2000 20.9 GB,to train and validate loess earthquake landslides on the satellite imagery platform and to achieve automatic segmentation of loess earthquake landslides in the validation area,improv-ing recognition accuracy.Three databases were established to optimize the segmentation prediction performance of the model.Database 1 con-tained 494 original loess earthquake landslide maps,Database 2 contained 920 loess earthquake landslide maps obtained after expansion through rotation,and Database 3 contained 613 maps obtained after expansion through cropping.The database 1 was input into the original Unet model and the VGG‒Unet model,corresponding to Experiment 1 and Experiment 2,respectively.Then,Databases 2 and 3 were input into the model that produced better results in the first two tests,namely the VGG‒Unet model,corresponding to Experiment 3 and 4,respectively.The four sets of tests were compared and analyzed to determine the optimal model performance metrics.Landslides within the study area were extracted from satellite imagery with coverage areas of 0.5～1.0 km2,ensuring that each image contained an appropriate number of landslide patches and effec-tive training data,providing the model with sufficient landslide pixels for feature learning.A total of 494 raw images were obtained,covering 1 052 landslides,designated as Database 1.A portion of the images was expanded through rotation,yielding 920 image panels,designated as Database 2,while another portion was expanded through cropping,yielding 613 image panels,designated as Database 3.Database 1 was input into models utilizing ResNet and VGG16 as backbone features,namely the original Unet model and the VGG‒Unet model,corresponding to Experiments 1 and 2,respectively.Then,Databases 2 and 3 were input into the model that yielded superior results in the preceding two experiments,namely the VGG‒Unet model,corresponding to Experiment 3 and 4,respectively.Finally,the VGG‒Unet model,which demonstrated the best performance among the four tests,was utilized to generate prediction results for typical loess earthquake landslide areas within the validation area.The re-searchers compared the predicted identification results of the model with the field-investigated landslide locations one by one to ensure experi-mental accuracy. Results and Discussions The results showed that the VGG‒Unet model demonstrated the best training and validation performance on the dataset that was cropped and expanded to 613 frames,and performed particularly well on the unknown dataset.The accuracy was 89.57%,the MIoU was 70.13%,the F1 value was 81.11%,and the Recall reached 80.53%,and all model performance indices were high.In contrast,on the original 494-frame dataset,the performance of the original Unet model was the lowest,followed by the VGG‒Unet model,and both were lower than the results of Test 3.However,on the dataset expanded through rotation to 920 frames,although the training performance metrics of the VGG‒Unet model reached the highest values among the three tests,the validation performance metrics decreased to the lowest.This method identified loess land-slides triggered by earthquakes.The model showed higher accuracy in identifying loess earthquake landslides with well-developed topographic and profile features,whereas it tended to miss landslides with ambiguous tonal and profile features,as well as small-scale landslides,and tended to misinterpret complex terrains such as uneven strata.Overall,the VGG‒Unet model effectively predicted areas prone to loess landslides.The clearer and more complete the characteristic markers of loess landslides were,the more effectively the VGG‒Unet model performed in segmenta-tion tasks.Its high precision and accuracy enabled computers to rapidly and accurately identify target areas and determine their locations.For landslides with blurred or incomplete characteristic markers,although the model produced a limited number of misclassifications and omissions during segmentation,most areas were still accurately segmented and identified. Conclusions Therefore,the method proposed in this study can efficiently segment and recognize the same type of landslide areas and can auto-matically segment and recognize the locations of landslides before field investigations.This approach assists field teams in locating loess earthquake-induced landslides and provides technical support for large-scale landslide disaster investigations.