An intelligent identification method for flood water bodies based on improved deep learning model

Abstract

The present invention discloses a method for intelligently identifying flood water bodies based on an improved deep learning model, which belongs to the technical field of geological disaster information processing, and includes the following steps: data collection and preprocessing to generate a data set; based on the generated data set, integrating the Canny edge detection operator, the normalized water body index, and the local binary pattern LBP operator texture to form a multi-feature fusion graph data set; constructing a DeepWFNet network model; inputting the multi-feature fusion graph data set into the DeepWFNet network model to identify the flood inundation range, and obtaining the flood identification result; and visualizing and optimizing the flood identification result. The present invention combines edge features, spectral features, and texture features with the DeepWFNet model, and can enhance the model's ability to understand global information while maintaining the local feature extraction capability, thereby significantly improving the segmentation accuracy of the flood inundation area.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster information processing, and in particular to a method for intelligently identifying flood water bodies based on an improved deep learning model. Background Art

[0002] Floods are a common natural disaster that often cause serious damage. According to statistics, the economic losses caused by floods around the world amount to tens of billions of dollars each year, and cause a large number of casualties. In order to effectively reduce the losses caused by floods, timely and accurate acquisition of flood inundation range information is an important part of emergency response and post-disaster recovery. However, traditional flood inundation range identification methods have great limitations in data acquisition and processing efficiency, resulting in unsatisfactory results in practical applications.

[0003] Existing flood monitoring and assessment technologies mainly rely on remote sensing technology, ground monitoring, hydrological models and other means. Remote sensing technology can provide timely image data for a large area and has become an important tool for flood monitoring. With the rapid development of satellite and drone technology, remote sensing images from different carriers have become an important supplement to flood monitoring. Therefore, more and more researchers have begun to try to apply deep learning technology to flood monitoring. As an emerging technology, deep learning has achieved remarkable results in many fields such as image recognition and speech processing. Deep learning can effectively improve the accuracy of image segmentation and classification by constructing a multi-layer convolutional network model and automatically extracting data features.

[0004] In recent years, deep learning-based image segmentation technology has been gradually applied to the identification of flood inundation areas, and has attracted much attention due to its powerful image segmentation capabilities. In particular, convolutional neural networks (CNNs) have become the mainstream choice for many image segmentation tasks due to their powerful feature extraction capabilities. Among them, the DeepLabV3+ model performs well in image segmentation tasks, but there are still some limitations, especially when dealing with large-scale images and complex scenes, which are easily disturbed by noise and local features. The Transformer type model models feature relationships in the global scope of the image through the self-attention mechanism, allowing the model to better capture complex contextual information and better handle these problems in such scenarios. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides an intelligent flood water body identification method based on an improved deep learning model with simple algorithm and high recognition accuracy.

[0006] The technical solution of the present invention to solve the above technical problems is: a method for intelligent identification of flood water bodies based on an improved deep learning model, comprising the following steps:

[0007] S1: data collection and preprocessing, generating datasets;

[0008] S2: Based on the data set generated in step S1, the Canny edge detection operator, the normalized water index, and the local binary pattern LBP operator texture are integrated to form a multi-feature fusion image data set;

[0009] S3: Build the DeepWFNet network model based on the DeepLabV3+ network and the Vision Transformer network;

[0010] S4: Input the multi-feature fusion graph dataset into the DeepWFNet network model to identify the flood inundation range and obtain the flood identification result;

[0011] S5: Visualize and evaluate flood identification results.

[0012] In the above-mentioned intelligent identification method for flood water bodies based on the improved deep learning model, the specific steps of step S1 are:

[0013] S11: Collect remote sensing images of flood events within one year as raw data. The remote sensing images include at least three bands: red, green and blue.

[0014] S12: preprocess the original data to obtain a labeled sample data set;

[0015] S13: Perform image enhancement on the labeled sample data set to generate a data set.

[0016] In the above-mentioned intelligent identification method of flood water bodies based on the improved deep learning model, in step S12, the preprocessing process is: first, the collected raw data is subjected to denoising using adaptive filtering, and then radiation correction and geometric correction are performed using ENVI software, and then the processed remote sensing images are cropped to a size of 512×512 pixels and blurred remote sensing images are removed, and then all remote sensing image formats are converted to .tif, and finally labelme is used to mark the flood coverage area to generate the corresponding true value label, thereby generating a binary classification label sample data set.

[0017] In the above-mentioned intelligent identification method of flood water bodies based on the improved deep learning model, in step S13, the image enhancement process is: the remote sensing image obtained in step S12 and the corresponding true value label are rotated, scaled, translated, and brightness adjusted to expand the data, thereby generating a data set.

[0018] In the above-mentioned intelligent identification method for flood water bodies based on the improved deep learning model, the specific steps of step S2 are:

[0019] S21: applying the Canny edge detection operator to the data set generated in step S13 to obtain an edge feature map of the remote sensing image;

[0020] S22: applying the normalized water index to the data set generated in step S13 to calculate the spectral characteristic map of the remote sensing image;

[0021] S23: applying the LBP operator to the data set generated in step S13 to generate a texture feature map of the remote sensing image;

[0022] S24: scaling the pixel values ​​of the edge feature map, the spectral feature map, and the texture feature map in geometric proportions;

[0023] S25: The geometrically scaled edge feature map, spectral feature map, and texture feature map are concatenated with the corresponding RGB image in a band combination manner to generate a final multi-feature fusion image dataset.

[0024] In the above-mentioned intelligent flood water body identification method based on the improved deep learning model, in step S22, when the remote sensing image has only three bands of red, green and blue, the ; If there is a near-infrared band, use ; If there is a short-wave infrared band, use , where Band B Indicates the blue band, Band G Indicates the green band, Band R Indicates the red band, Band NIR Indicates near infrared band, Band SWIR1 and Band SWIR2 They respectively represent the Band 6 and Band 7 values ​​on the Landsat 8 satellite.

[0025] In the above-mentioned intelligent identification method of flood water bodies based on the improved deep learning model, in step S24, the Min-Max normalization method is applied to linearly map the original grayscale value to the range of [0,255], and the formula is:

[0026] ;

[0027] in X new is the scaled grayscale value, X is the original grayscale value, X min is the minimum value of the original grayscale value, X max is the maximum value of the original grayscale value.

[0028] In the above-mentioned intelligent identification method for flood water bodies based on the improved deep learning model, the specific process of step S3 is as follows:

[0029] S31: Design of DeepWFNet model: Based on the DeepLabV3+ framework, the encoder-decoder structure is adopted. The MobileNetV3 network and the Vision Transformer network are introduced in a dual-branch parallel mode as the backbone network for feature extraction. The CAFM module is used to fuse two different types of features in the two networks. The CA attention module is introduced and the ASPP module is improved through dense connections, thereby forming an improved deep learning model, namely the DeepWFNet model.

[0030] S32: Configure the training environment and training parameters required for the DeepWFNet model;

[0031] S33: The DeepWFNet model is trained using a remote sensing image dataset with flood-inundated area annotations, i.e., the final multi-feature fusion image dataset. The DeepWFNet model training adopts a multi-feature fusion strategy, taking edge features, spectral features, texture features and RBG images as input, and the output result is a grayscale image with a binary distribution of [0,1].

[0032] In the above-mentioned intelligent flood water body identification method based on the improved deep learning model, the specific steps of step S4 are:

[0033] S41: obtaining remote sensing images of the flood coverage area to be identified, including at least three bands of red, green and blue, and preprocessing the obtained remote sensing image data to be identified, including: denoising, radiation correction, geometric correction, cropping to 512×512 pixel size, and converting to .tif format;

[0034] S42: Calculate edge features, spectral features, and texture features of the remote sensing image to be identified, generate corresponding feature maps, and splice them with the original remote sensing image to be identified to form a data set to be identified;

[0035] S43: Apply the trained DeepWFNet model to identify the dataset to be identified, classify the image at the pixel level, output the submerged area prediction map, and stitch the results back to the size before cropping.

[0036] In the above-mentioned intelligent identification method for flood water bodies based on the improved deep learning model, the specific steps of step S5 are:

[0037] S51: Post-processing the flood inundation range results output by the DeepWFNet model, including: corrosion and expansion, region merging and segmentation, and finally generating vector data of the flood inundation range;

[0038] S52: Overlay the vector data of flood inundation area into the geographic information system to facilitate visualization analysis and further disaster assessment.

[0039] The beneficial effects of the present invention are as follows: the present invention constructs a new deep learning model DeepWFNet, and combines the selection of features in traditional methods to form a multi-feature fusion strategy. By adding edge features, spectral features, and texture features in combination with the DeepWFNet model, the model's ability to understand global information can be enhanced while maintaining local feature extraction capabilities, thereby significantly improving the segmentation accuracy of flood-inundated areas, providing strong data support for subsequent emergency response and post-disaster recovery, effectively improving the accuracy and robustness of flood-affected area identification, and enhancing the practicability of the algorithm, which is helpful to promote flood relief and governance work. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the overall flow chart of the present invention.

[0041] Figure 2 Schematic diagram of the structure of the DeepWFNet model of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] like Figure 1 As shown, a flood water body intelligent identification method based on an improved deep learning model includes the following steps:

[0044] S1: Data collection and preprocessing, generating datasets.

[0045] The specific steps of step S1 are:

[0046] S11: Collect remote sensing images of flood events occurring within one year as raw data. The remote sensing images include at least three bands: red, green, and blue.

[0047] To obtain the floods that occurred in recent years, remote sensing images were obtained using satellite images or drone images covering the area, with complete images and less than 10% cloud cover.

[0048] S12: Preprocess the original data to obtain a labeled sample data set.

[0049] The preprocessing process is as follows: first, the collected raw data is denoised using adaptive filtering, and then radiation correction and geometric correction are performed using ENVI software. Next, the processed remote sensing images are cropped to a size of 512×512 pixels and blurred remote sensing images are removed. All remote sensing image formats are converted to .tif, and finally labelme is used to mark the flood coverage area to generate the corresponding true value label, thereby generating a binary classification label sample data set.

[0050] S13: Perform image enhancement on the labeled sample data set to generate a data set.

[0051] The image enhancement process is as follows: the remote sensing image obtained in step S12 and the corresponding true value label are rotated (rotated 180°), scaled (magnified 1.2 times), translated (translated 20 pixels to the left and filled with black blank areas), and brightness adjusted (the brightness value of each pixel is fixedly increased by 5) to expand the data and generate a data set.

[0052] S2: Based on the data set generated in step S1, the Canny edge detection operator, the normalized water index, and the local binary pattern LBP operator texture are integrated to form a multi-feature fusion image data set.

[0053] The specific steps of step S2 are:

[0054] S21: Apply the Canny edge detection operator to the data set generated in step S13 to obtain an edge feature map of the remote sensing image.

[0055] Use the Canny algorithm in OpenCV to batch calculate images, and set different parameters for satellite images or drone images. You can refer to (50,100) and (30,60) respectively.

[0056] S22: Apply the normalized water index to the data set generated in step S13 to calculate the spectral characteristic map of the remote sensing image.

[0057] When the remote sensing image has only three bands of red, green and blue, use ; If there is a near-infrared band, use ; If there is a short-wave infrared band, use , where Band B Indicates the blue band, Band G Indicates the green band, Band R Indicates the red band, Band NIR Indicates near infrared band, Band SWIR1 and Band SWIR2 They respectively represent the Band 6 and Band 7 values ​​on the Landsat 8 satellite.

[0058] S23: Apply the LBP operator to the data set generated in step S13 to generate a texture feature map of the remote sensing image.

[0059] Use the LBP algorithm provided by OpenCV to batch calculate images, and set different parameters for satellite images and drone images. Satellite images with higher resolution can be selected (R=2, P=24), and satellite images with lower resolution can be selected (R=1, P=16), where R is the radius and P is the number of sampling points.

[0060] S24: Scale the pixel values ​​of the edge feature map, the spectral feature map, and the texture feature map to [0, 255] in a geometric manner.

[0061] Apply the Min-Max normalization method to linearly map the original grayscale value to the range of [0,255], the formula is:

[0062] ;

[0063] in X new is the scaled grayscale value, X is the original grayscale value, X min is the minimum value of the original grayscale value, X max is the maximum value of the original grayscale value.

[0064] S25: The geometrically scaled edge feature map, spectral feature map, texture feature map and the corresponding RGB image are spliced ​​in a band combination manner to form a six-band image map, and the final multi-feature fusion image dataset is generated together with the corresponding true value labels.

[0065] S3: Build the DeepWFNet network model based on the DeepLabV3+ network and the Vision Transformer network.

[0066] The specific process of step S3 is:

[0067] S31: Design of DeepWFNet model: Adopting the encoder-decoder structure, based on the DeepLabV3+ framework, the MobileNetV3 network and the Vision Transformer network are introduced in a dual-branch parallel mode as the backbone network for feature extraction, and the CAFM module is used to fuse the two different types of features in the two networks. The CA attention module is introduced and the ASPP module is improved through dense connections, thus forming an improved deep learning model, namely the DeepWFNet model.

[0068] DeepWFNet is a dual-branch parallel feature extraction structure model. Its input is composed of a data set of RGB images, edge feature maps, spectral feature maps, and texture feature maps. This multi-feature fusion strategy can enhance the contrast of flood-covered areas, increase the amount of feature information, make it easier for the model to learn features, and improve the final recognition accuracy.

[0069] Different from previous models, DeepWFNet uses two different types of backbone networks, convolutional neural networks (C branches) and Transformer (T branches), to learn features. The C branch uses MobileNetV3, which has the advantages of lightweight network parameters and fast computing speed. The T branch uses Vision Transformer, which has a simple structure, outstanding results and strong scalability. The feature information contained in the feature layers of the two architectures of CNN and Transformer has different emphases. The receptive field and spatial hierarchical characteristics of the CNN branch make its features have obvious local correlation, while the sequence modeling characteristics of the Transformer branch allow the model to establish a direct dependency relationship between any two positions in the image, making its features have global dependency.

[0070] The CAFM module is introduced to fuse the two types of features. The features from the C branch are additionally added with a multi-layer perceptron (MLP) for low dimensionality to reduce computational complexity and storage requirements. The local features of the Transformer are re-weighted with the global information of the CNN, and the local information of the Transformer is re-weighted with the global features of the CNN. This helps the DeepWFNet model establish a good connection between the CNN and the Transformer, and promotes the fusion of the local features of the two branches with the global features. For the deep features of the C branch, an improved ASPP module is added, which can effectively improve the receptive field of the convolution layer and the number of times the feature points are used, and it plays a role in balancing the attention between different areas and improving the utilization rate of the feature layer information. The improved ASPP module reconnects the convolution layer in a densely connected manner: the results of the dilated convolution outputs of the first three layers with different dilation rates are merged into the input of the next layer; the CA attention module is introduced after each convolution layer: the CA attention module can not only capture cross-channel information, but also capture direction perception and position perception information, helping the model to more accurately locate and identify targets of interest, and upsample the output of the deep features and splice it with the output of the dual-branch feature CAFM, and then upsample again to restore it to the original input size after convolution fusion.

[0071] S32: Configure the training environment and training parameters required for the DeepWFNet model.

[0072] The specific parameters are: the activation function is ReLU, the batch size is set to 4, the epoch is set to 100, the initial learning rate is 0.001, and the optimization method is stochastic gradient descent SGD.

[0073] S33: The DeepWFNet model is trained using a remote sensing image dataset with flood-inundated area annotations, i.e., the final multi-feature fusion image dataset. The DeepWFNet model training adopts a multi-feature fusion strategy, taking edge features, spectral features, texture features and RBG images as input, and the output result is a grayscale image with a binary distribution of [0,1].

[0074] S4: Input the multi-feature fusion graph dataset into the DeepWFNet network model to identify the flood inundation range and obtain the flood identification results.

[0075] The specific steps of step S4 are:

[0076] S41: Acquire remote sensing images of the flood coverage area to be identified, including at least three bands: red, green and blue. Preprocess the acquired remote sensing image data to be identified, including: denoising, radiation correction, geometric correction, cropping to 512×512 pixel size, and converting to .tif format. The .tif format can store multi-band information.

[0077] S42: Calculate the edge features, spectral features, and texture features of the remote sensing image to be identified, generate a corresponding feature map, and splice it with the original remote sensing image to be identified to form a data set to be identified.

[0078] S43: Apply the trained DeepWFNet model to identify the dataset to be identified, classify the image at the pixel level, output the submerged area prediction map, and stitch the results back to the size before cropping.

[0079] S5: Visualize and evaluate flood identification results.

[0080] The specific steps of step S5 are:

[0081] S51: Post-processing the flood inundation range results output by the DeepWFNet model, including: corrosion and expansion (calculation window is 5×5, number of iterations is 2), region merging and segmentation (optional, small regions are merged when there are too many broken regions, and segmentation when the range is too large), and finally generating vector data of the flood inundation range.

[0082] The final prediction result will contain a small amount of errors such as misclassification, omissions, holes, and scattered points. The erosion and dilation operations in morphology can be used to remove noise and small misclassified areas, fill omissions and holes, correct errors in small areas, and optimize the final overall visualization effect.

[0083] S52: Overlay the vector data of flood inundation area into the geographic information system to facilitate visualization analysis and further disaster assessment.

Claims

1. A method for intelligent identification of flood water bodies based on an improved deep learning model, characterized in that: The following steps are involved: S1: data collection and preprocessing, generating datasets; S2: Based on the data set generated in step S1, the Canny edge detection operator, the normalized water index, and the local binary pattern LBP operator texture are integrated to form a multi-feature fusion image data set; S3: Build the DeepWFNet network model based on the DeepLabV3+ network and the Vision Transformer network; The specific process of step S3 is: S31: Design of DeepWFNet model: Based on the DeepLabV3+ framework, the encoder-decoder structure is adopted. The MobileNetV3 network and the Vision Transformer network are introduced in a dual-branch parallel mode as the backbone network for feature extraction. The CAFM module is used to fuse two different types of features in the two networks. The CA attention module is introduced and the ASPP module is improved through dense connections, thereby forming an improved deep learning model, namely the DeepWFNet model. S32: Configure the training environment and training parameters required for the DeepWFNet model; S33: The DeepWFNet model is trained using a remote sensing image dataset with flood-inundated area annotations, i.e., the final multi-feature fusion image dataset. The DeepWFNet model training adopts a multi-feature fusion strategy, taking edge features, spectral features, texture features and RBG images as input, and the output result is a grayscale image with a binary distribution of [0,1]. S4: Input the multi-feature fusion graph dataset into the DeepWFNet network model to identify the flood inundation range and obtain the flood identification result; S5: Visualize and evaluate flood identification results.

2. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 1 is characterized in that: The specific steps of step S1 are: S11: Collect remote sensing images of flood events within one year as raw data. The remote sensing images include at least three bands: red, green and blue. S12: preprocess the original data to obtain a labeled sample data set; S13: Perform image enhancement on the labeled sample data set to generate a data set.

3. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 2 is characterized in that: In step S12, the preprocessing process is as follows: first, the collected raw data is subjected to denoising using adaptive filtering, and then radiation correction and geometric correction are performed using ENVI software. Then, the processed remote sensing images are cropped to a size of 512×512 pixels and blurred remote sensing images are removed. Then, all remote sensing image formats are converted to .tif. Finally, labelme is used to mark the flood coverage area to generate the corresponding true value label, thereby generating a binary classification label sample data set.

4. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 3 is characterized in that: In step S13, the image enhancement process is: rotating, scaling, translating, and adjusting brightness of the remote sensing image obtained in step S12 and the corresponding true value label to expand the data, thereby generating a data set.

5. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 4 is characterized in that: The specific steps of step S2 are: S21: applying the Canny edge detection operator to the data set generated in step S13 to obtain an edge feature map of the remote sensing image; S22: applying the normalized water index to the data set generated in step S13 to calculate the spectral characteristic map of the remote sensing image; S23: applying the LBP operator to the data set generated in step S13 to generate a texture feature map of the remote sensing image; S24: scaling the pixel values ​​of the edge feature map, the spectral feature map, and the texture feature map in geometric proportions; S25: The geometrically scaled edge feature map, spectral feature map, and texture feature map are concatenated with the corresponding RGB image in a band combination manner to generate a final multi-feature fusion image dataset.

6. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 5 is characterized in that: In step S22, when the remote sensing image has only three bands of red, green and blue, If there is a near-infrared band, use If there is a short-wave infrared band, use Band B +2.5Band G -1.5×(Band NIR +Band SWIR1 )-0.25Band SWIR2 , where Band B Indicates the blue band, Band G Indicates the green band, Band R Indicates the red band, Band NIR Indicates near infrared band, Band SWIR1 and Band SWIR2 They respectively represent the Band 6 and Band 7 values ​​on the Landsat 8 satellite.

7. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 6 is characterized in that: In step S24, the Min-Max normalization method is applied to linearly map the original grayscale value to the range of [0, 255], and the formula is: Where X new is the scaled grayscale value, X is the original grayscale value, and X min is the minimum value of the original grayscale value, X max is the maximum value of the original grayscale value.

8. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 1, characterized in that: The specific steps of step S4 are: S41: obtaining remote sensing images of the flood coverage area to be identified, including at least three bands of red, green and blue, and preprocessing the obtained remote sensing image data to be identified, including: denoising, radiation correction, geometric correction, cropping to 512×512 pixel size, and converting to .tif format; S42: Calculate edge features, spectral features, and texture features of the remote sensing image to be identified, generate corresponding feature maps, and splice them with the original remote sensing image to be identified to form a data set to be identified; S43: Apply the trained DeepWFNet model to identify the dataset to be identified, classify the image at the pixel level, output the submerged area prediction map, and stitch the results back to the size before cropping.

9. The method for intelligent identification of flood water bodies based on an improved deep learning model according to claim 8, characterized in that: The specific steps of step S5 are: S51: Post-processing the flood inundation range results output by the DeepWFNet model, including: corrosion and expansion, region merging and segmentation, and finally generating vector data of the flood inundation range; S52: Overlay the vector data of flood inundation area into the geographic information system to facilitate visualization analysis and further disaster assessment.

Images