A high-resolution image landslide detection and segmentation method
By establishing a Mask R-CNN landslide detection and segmentation network model and performing transfer learning, the problems of low efficiency and high cost in identifying and detecting old landslides were solved, and efficient detection and segmentation of old landslides were achieved.
Patent Information
- Application Number
- CN202211258519.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-14
- Publication Date
- 2025-12-16
- Estimated Expiration
- 2042-10-14
AI Technical Summary
Existing technologies for identifying and detecting old landslides are inefficient and costly, and the limited number of old landslide image samples makes accurate identification and segmentation difficult.
A high-resolution image detection and segmentation method was adopted to establish a Mask R-CNN landslide detection and segmentation network model. The TL-Mask R-CNN landslide detection and segmentation network model was trained by transfer learning and old landslide samples in the optical remote sensing dataset was used for training to improve detection efficiency and accuracy.
It enables rapid detection and segmentation of old landslides, determines the old and new types and specific distribution of landslides, improves detection efficiency, and reduces costs.
Smart Images
Figure CN115601544B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of image detection and segmentation, and particularly relates to a high-resolution image landslide detection and segmentation method. BACKGROUND
[0002] Landslides are usually caused by heavy rain, earthquakes and human activities. The Qinghai-Tibet Plateau has complex geological structures, and landslides are extremely developed, which brings potential risks to the engineering construction and operation safety of the Sichuan-Tibet traffic project. Therefore, the investigation and detection of landslide disasters are crucial.
[0003] Currently, the research on landslides mainly focuses on identifying newly occurring landslides, and there is less research on old landslides that occurred earlier, resulting in fewer image samples of old landslides. The surface accumulation body of old landslides is unstable, and combined with geological movements and human activities, it is likely to occur again, causing great damage.
[0004] Traditional identification and detection of old landslides still mainly rely on expert experience and expensive laser radar data, which is low in efficiency and high in cost. Therefore, a recognition method is needed to identify and detect old landslides, which overcomes the defects of low efficiency, high cost and fewer image samples of old landslides of traditional detection methods, and realizes accurate identification of old landslides. SUMMARY
[0005] In view of the problems in the prior art, the purpose of the present application is to provide a high-resolution image landslide detection and segmentation method.
[0006] In order to achieve the above purpose, the technical scheme is adopted as follows.
[0007] A high-resolution image landslide detection and segmentation method, comprising the following steps:
[0008] Step 1, establishing a landslide optical remote sensing image sample set;
[0009] Step 2, establishing a Mask R-CNN landslide detection and segmentation network model;
[0010] Step 3, using the new landslide sample in the landslide optical remote sensing data set to train the Mask R-CNN landslide detection and segmentation network model, and obtaining the trained Mask R-CNN landslide detection and segmentation network model;
[0011] Step 4, constructing a TL-Mask R-CNN landslide detection and segmentation network model according to the trained Mask R-CNN landslide detection and segmentation network model;
[0012] Step 5: Use old landslide samples from the landslide optical remote sensing dataset to train the TL-Mask R-CNN landslide detection and segmentation network model to obtain the trained TL-Mask R-CNN landslide detection and segmentation network model.
[0013] Step 6: Use the trained TL-Mask R-CNN landslide detection and segmentation network model to detect and segment the landslide image.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: a landslide detection and segmentation network model based on optical images is established, and the transfer learning method is used to enable the model to detect and segment old landslides with fewer optical images, thereby quickly completing the detection and segmentation of landslides, determining the old and new types and specific distribution of landslides, improving detection efficiency and reducing costs. Attached Figure Description
[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0016] Figure 1 This is a flowchart of the method of the present invention;
[0017] Figure 2 This is a schematic diagram of the Mask R-CNN landslide detection and segmentation network model of the present invention;
[0018] Figure 3 This is a schematic diagram of the FPN feature pyramid network structure of the method of the present invention;
[0019] Figure 4 This is a schematic diagram of the structure of the RPN region candidate network in the method of the present invention;
[0020] Figure 5 This is a schematic diagram of the structure of the TL-Mask R-CNN landslide detection and segmentation network model of the present invention;
[0021] Figure 6 Images of old landslides;
[0022] Figure 7 To Figure 6 Experts' visual assessment of the landslide;
[0023] Figure 8 To Figure 6 The method of this invention is used to detect and segment the results of landslides. Detailed Implementation
[0024] The embodiments of the present invention will be described in detail below with reference to examples. However, those skilled in the art will understand that the following examples are only for illustrating the present invention and should not be regarded as limiting the scope of the present invention.
[0025] Reference Figure 1 A high-resolution image landslide detection and segmentation method, comprising the following steps:
[0026] Step 1, establish a landslide optical remote sensing image sample set;
[0027] Specifically, a plurality of optical remote sensing images containing landslides are obtained, including new landslide optical remote sensing images and old landslide optical remote sensing images; each optical remote sensing image, the corresponding landslide type of each image, and the mask coordinate position corresponding to the landslide range in each image are taken as a group of sample data.
[0028] Step 2, establish a Mask R-CNN landslide detection and segmentation network model;
[0029] Reference Figure 2 The Mask R-CNN landslide detection and segmentation network model includes a ResNet feature extraction network, an FPN feature pyramid network, a region candidate network RPN, an ROI alignment and a function network;
[0030] The ResNet feature extraction network is used to extract features and protect the integrity of information to avoid gradient disappearance or explosion;
[0031] Reference Figure 3 The FPN feature pyramid network with bottom-up and top-down structure is used to pass down higher-level features to supplement lower-level semantics in order to retrieve high-resolution and strong semantic features and promote small target detection;
[0032] The RPN region candidate network is used to generate accurate region proposals; Reference Figure 4 The RPN region candidate network includes two branches, one branch uses softmax to classify anchor points to obtain positive classification and negative classification, and the other branch calculates the boundary box regression offset of the anchor points to obtain accurate region proposals; finally, the positive anchor and the corresponding boundary offset are synthesized through the proposal layer to eliminate too small or out-of-bound proposals.
[0033] The ROI alignment is used to convert features into the same size mapping by bilinear interpolation and pooling method to avoid errors caused by two coordinate quantization rounding and improve the accuracy of coordinate system regression;
[0034] The function network is used to generate 7x7x1024 ROI features, then sample the ROI features to 2048 channels, and input the sampled features into two double-branch networks, one double-branch network is used to classify the landslide type of the image and generate coordinate frame in the landslide area, and the other double-branch network is used to generate the corresponding mask.
[0035] Step 3, training the Mask R-CNN landslide detection and segmentation network model using new landslide samples in the landslide optical remote sensing data set, obtaining the trained Mask R-CNN landslide detection and segmentation network model;
[0036] Step 4, constructing a TL-Mask R-CNN landslide detection and segmentation network model according to the trained Mask R-CNN landslide detection and segmentation network model;
[0037] Specifically, referring to Figure 5 , first, a new functional network is constructed, which contains two large branches. In one large branch, the features are generated into masks through a full convolution network. In the other large branch, the features are divided into two small branches after passing through a fully connected layer. In one small branch, the features pass through a fully connected layer and a softmax function in sequence for classification. In the other small branch, the features pass through a fully connected layer and Bbox reg in sequence to determine the coordinates of the coordinate frame. Finally, the masks, classification results and coordinate frames are added to the input image and output, obtaining the landslide image after detection and segmentation is completed.
[0038] Then, the functional network in the trained Mask R-CNN landslide detection and segmentation network model is replaced with the new functional network, obtaining the TL-Mask R-CNN landslide detection and segmentation network model.
[0039] Step 5, training the TL-Mask R-CNN landslide detection and segmentation network model using old landslide samples in the landslide optical remote sensing data set, obtaining the trained TL-Mask R-CNN landslide detection and segmentation network model;
[0040] Specifically, when training, the ResNet feature extraction network, FPN feature pyramid network, RPN region candidate network and ROI alignment in the TL-Mask R-CNN landslide detection and segmentation network model are frozen.
[0041] In view of the situation that old landslide samples are less, combined with transfer learning, all parameters except the functional network are frozen, and only the functional network is trained, effectively reducing the number of training parameters and retaining more common landslide features in the trained model. Common features extracted from new landslides are used for effective training.
[0042] Step 6, using the trained TL-Mask R-CNN landslide detection and segmentation network model to detect and segment the landslide image.
[0043] Referring to Figure 6 , for old landslide images, it can be seen that it is difficult for humans to identify old landslides without enough landslide interpretation experience. Referring to Figure 7 , for Figure 6Expert visual assessment of the landslide. (Reference) Figure 8 In order to Figure 6 The method of this invention is used to detect and segment the results of landslides.
[0044] contrast Figure 7 , Figure 8 As can be seen, the method of this invention can detect and identify old landslides with some visible features, but it is not very accurate in landslide shape segmentation. This is because manually drawing the landslide shape during the labeling process is a systematic and comprehensive task, requiring a comprehensive analysis of all geological conditions, including pore water, slope structure, location, crack properties, excavation, physical exploration, and ground surveys. Different data need to be carefully cross-validated to draw a reasonable landslide surface, not just the sliding portion. Therefore, it is difficult to accurately segment the landslide shape.
[0045] Simulation test results
[0046] Typical deep learning models used for semantic segmentation, including Unet, Unet++, and Deeplabv3+, as well as Mask R-CNN without transfer learning, are compared with TL-Mask R-CNN of the present invention for old landslide inspection segmentation.
[0047] The experimental parameters were set as follows: epoch 100, batch size 4, and learning rate 0.0001. The Adam (Adaptive Moment Estimator) function was used for optimization, and the images were cropped to 512×512 pixels for input model training.
[0048] The detection and segmentation results are shown in Table 1.
[0049] Table 1
[0050]
[0051] As shown in Table 1, TL-Mask R-CNN improves all metrics by more than 10% compared to Unet, Unet++, Deeplabv3+, and MaskR-CNN, indicating that the TL-Mask R-CNN network model of this invention can better detect and segment old landslides in images.
[0052] Although the present invention has been described in detail in this specification with general description and specific embodiments, some modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention are within the scope of protection claimed by the present invention.
Claims
1. A high resolution image landslide detection and segmentation method, characterized in that, The method comprises the following steps: Step 1, establishing a landslide optical remote sensing image sample set; Step 2, establishing a Mask R-CNN landslide detection and segmentation network model; Step 3, using new landslide samples in the landslide optical remote sensing data set to train the Mask R-CNN landslide detection and segmentation network model, to obtain a trained Mask R-CNN landslide detection and segmentation network model; Step 4, constructing a TL-Mask R-CNN landslide detection and segmentation network model according to the trained Mask R-CNN landslide detection and segmentation network model; Specifically, first, a new functional network is constructed, which contains two large branches; in one large branch, the features are generated into masks through a full convolution network; in the other large branch, the features are divided into two small branches after passing through a full connection layer; in one small branch, the features pass through a full connection layer and a Softmax function in sequence for classification; in the other small branch, the features pass through a full connection layer and a Bbox reg in sequence for determining the coordinates of the coordinate frame; finally, the masks, the classification results and the coordinate frame are added to the input image and output, to obtain a landslide image whose detection and segmentation are completed; Then, the functional network in the trained Mask R-CNN landslide detection and segmentation network model is replaced by the new functional network, to obtain a TL-Mask R-CNN landslide detection and segmentation network model; Step 5, using old landslide samples in the landslide optical remote sensing data set to train the TL-Mask R-CNN landslide detection and segmentation network model, to obtain a trained TL-Mask R-CNN landslide detection and segmentation network model; Step 6, using the trained TL-Mask R-CNN landslide detection and segmentation network model to detect and segment the landslide image.
2. The high resolution image landslip detection and segmentation method of claim 1, wherein, Specifically, in step 1, a plurality of optical remote sensing images containing landslides are obtained, including new landslide optical remote sensing images and old landslide optical remote sensing images; each optical remote sensing image, the corresponding landslide type of each image and the mask coordinate position corresponding to the landslide range in each image are taken as a group of sample data.
3. The high resolution image landslip detection and segmentation method of claim 1, wherein, Specifically, the Mask R-CNN landslide detection and segmentation network model comprises a ResNet feature extraction network, an FPN feature pyramid network, a region proposal network RPN, an ROI alignment and a functional network; The ResNet feature extraction network is used for extracting features and protecting the integrity of information to avoid gradient disappearance or explosion; the FPN feature pyramid network is used for downwardly passing higher-level features to supplement lower-level semantics and promote small target detection; the RPN region proposal network is used for generating accurate region proposals; and the ROI alignment is used for improving the accuracy of coordinate system regression; The functional network is used for generating 7x7x1024 ROI features, then sampling the ROI features to 2048 channels, and inputting the sampled features into two double-branch networks; one double-branch network is used for classifying the landslide types of the image and generating coordinate frames in the landslide area; and the other double-branch network is used for generating corresponding masks.
4. The high resolution image landslip detection and segmentation method of claim 1, wherein, During training, the ResNet feature extraction network, the FPN feature pyramid network, the RPN region candidate network and the ROI alignment in the TL-Mask R-CNN landslide detection segmentation network model are frozen.