A DSM Vegetation Automatic Height Reduction Method Based on a Deep Learning Model

Abstract

The present invention provides a method for automatically reducing the height of DSM vegetation based on a deep learning model, comprising the following steps: constructing a vegetation recognition model and a vegetation automatic height reduction model; respectively obtaining vegetation classification results through the vegetation recognition model and obtaining DSM full-map height reduction results through inference by the vegetation automatic height reduction model; segmenting the original DSM and the corrected DSM full-map height reduction results according to the range of the vegetation classification results, and using the non-forest area of the original DSM to replace the non-forest area of the corrected DSM full-map height reduction results to finally form a height replacement result; setting forest and non-forest buffer zones based on the height replacement result and performing smoothing processing using Gaussian smoothing to form the final height-reduced result. The present invention uses the RFRNET deep learning model to perform overall regional height reduction, and then by adding vegetation and non-vegetation classification information, replaces the non-vegetation area, and optimizes the height reduction result to form the final DSM height reduction result.

Description

Technical Field

[0001] The present invention belongs to the field of power grids, and specifically relates to a method for automatically reducing the height of DSM vegetation based on a deep learning model. Background Art

[0002] A digital elevation model (DEM) is an elevation model processed from a digital surface model (DSM) and is basic data for many scientific researches. It has become direct supporting data in fields such as global mapping, ecological assessment, disaster warning, and strategic decision-making, and its quality directly affects the data analysis ability and the application accuracy of scientific achievements. The main production method of DEM is to adopt a full stereo manual editing method for DSM. Through manual interpretation, the area to be reduced in height is drawn, and the surface elevation is reduced to the ground by giving a height reduction value. Since vegetation usually occupies a large area on the map, is scattered and has various forms, it requires a large amount of data processing time, which becomes the main factor restricting the production efficiency of DEM. Currently, there are mainly the following two methods for reducing the height of vegetation:

[0003] (1) Manual method: First, manually identify the vegetation-covered area, and then subtract the uniform vegetation height deviation irrelevant to the vegetation type from the elevation of DSM.

[0004] (2) Vegetation height model method: With the help of a vegetation height model, a certain percentage of the vegetation height is subtracted from DSM.

[0005] When using the manual method to reduce the height of vegetation, it is necessary to manually collect forested areas one by one and give height references for each area to reduce the height. The work is time-consuming and laborious, and for places lacking height references, only vegetation references can be collected nearby. At the same time, for areas with large-scale and high-density vegetation coverage, this method assumes a unified spatial error, but the height of vegetation is not completely uniform, which may cause some areas of the generated DEM to be too high or too low after processing, and the reliability cannot be guaranteed.

[0006] When using the vegetation height model method to reduce the height of vegetation, it depends on the vegetation height model. However, most of the currently released vegetation height models are too rough, with problems such as low resolution and data holes, and cannot be used as an effective input model for reducing the height of vegetation. Summary of the Invention

[0007] In view of the above problems, a method for automatically reducing the height of DSM vegetation based on a deep learning model mentioned in the present invention is to use the RFRNET deep learning model to perform overall regional height reduction for DSM containing vegetation, and then by adding vegetation and non-vegetation classification information, replace the non-vegetation area, and optimize the height reduction results to form the final DSM height reduction results.

[0008] The purpose of the present invention is to provide:

[0009] A method for automatically reducing the height of DSM vegetation based on a deep learning model, comprising the following steps:

[0010] Construct a vegetation recognition model and a vegetation automatic height reduction model;

[0011] Obtain the vegetation classification results through the vegetation recognition model and the DSM full-map height reduction results through reasoning by the vegetation automatic height reduction model respectively;

[0012] Segment the original DSM and the corrected DSM full-map height reduction results according to the range of the vegetation classification results, and use the non-forest area of the original DSM to replace the non-forest area of the corrected DSM full-map height reduction results, finally forming the height replacement results;

[0013] Set forest and non-forest buffer zones on the basis of the height replacement results, and perform smoothing processing using Gaussian smoothing to form the final height-reduced results.

[0014] In a preferred embodiment of the present invention, the DSM full-map height reduction results need to be corrected by an operator to determine the local correction value for height reduction.

[0015] In a preferred embodiment of the present invention, the trained vegetation recognition model and vegetation automatic height reduction model are respectively used to perform reasoning on the DOM and DSM at the same location in a specific area.

[0016] In a preferred embodiment of the present invention, the construction of the vegetation recognition model includes:

[0017] Construct a vegetation recognition training dataset, and then use the vegetation recognition training dataset to train the HROCR model to form a DOM vegetation recognition model. Then use the DOM vegetation recognition model to perform model reasoning on the dataset to be recognized DOM to form the DOM vegetation / non-vegetation classification results.

[0018] In a preferred embodiment of the present invention, the construction of the vegetation recognition model mainly has three links: data division, data cropping, and model training:

[0019] In the data division process, the regional vegetation recognition dataset of the original data is divided into a training set and a validation set according to a ratio of 8:2;

[0020] In the data cropping process, the "sliding window overlapping method" is used to crop the original image. Using this method can largely reduce the obvious traces of the small image edges on the large image due to inaccurate recognition of the small image edge area during stitching. For the prediction results of each sliding window, only the middle valid part of the results is taken to stitch the large image;

[0021] The model training process uses a model trained on the general COCO dataset as the pre-trained model, sets the number of iterations, and for the model after each iteration, uses the recall rate and precision rate commonly used in semantic segmentation to statistically analyze the metrics of the model on the validation set.

[0022] In a preferred embodiment of the present invention, the components of the DSM full-map height reduction result specifically include:

[0023] First, a vegetation height reduction training dataset is constructed, and then the RFRNET model is trained using the vegetation height reduction training dataset to form a DSM vegetation automatic height reduction model. Then, the DSM vegetation automatic height reduction model is used to perform model inference on the DSM dataset to be height-reduced, forming the DSM full-map height reduction result.

[0024] In a preferred embodiment of the present invention, constructing the DSM vegetation automatic height reduction model specifically includes dataset division, data cropping, data filtering, and model training:

[0025] The dataset is divided into a training set and a validation set according to the ratio of 8:2 for the prepared vegetation height reduction training dataset;

[0026] After the data is divided, the training data and validation data are cropped;

[0027] After cropping, the samples are filtered, and data containing both vegetation and non-vegetation areas is used for model training;

[0028] The model training process uses a model trained on the general COCO dataset as the pre-trained model, sets the number of iterations, and after each model iteration is completed, the commonly used metric of mean absolute error in error measurement is used to statistically analyze the metrics of the model on the validation set.

[0029] In a preferred embodiment of the present invention, the "sliding window overlapping method" is used to crop the DSM data, DEM data, and the corresponding DOM data of vegetation classification.

[0030] In a preferred embodiment of the present invention, in order to increase the proportion of subgraphs containing both vegetation and non-vegetation in the cropped subgraphs, the window size of the "sliding window overlapping method" is set to 512, and the step size is set to 128.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] The present invention uses the RFRNET deep learning model to perform regional overall height reduction on the DSM containing vegetation, and then by adding vegetation and non-vegetation classification information, replacing the non-vegetation areas, and optimizing the height reduction result, the final DSM height reduction result is formed.

[0033] Specifically, DSM automatically realizes vegetation height reduction through a deep learning model. Operators can abandon the traditional manual operation method, without the need to manually extract the vegetation range and give height reduction reference values for each area for vegetation height reduction, reducing the labor intensity of operators and improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a simplified block diagram of the working principle of the present invention;

[0035] Figure 2 It is a working principle diagram of the overall height reduction result of the DSM full map of the present invention;

[0036] Figure 3 It is a working principle diagram of the DOM vegetation / non-vegetation classification result formed by the present invention;

[0037] Figure 4 It is a working principle diagram of the elevation replacement result formed by the replacement of the present invention;

[0038] Figure 5 It is a working principle diagram of obtaining the final height reduction result of the present invention;

[0039] Figure 6 It is a detailed block diagram of the working principle of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The present invention will be described in detail below in conjunction with the embodiments shown in the drawings. However, it should be noted that these embodiments are not limitations on the present invention. Any equivalent transformation or substitution in terms of function, method, or structure made by those of ordinary skill in the art based on these embodiments shall fall within the protection scope of the present invention.

[0041] Through a deep learning algorithm, the present invention forms an identification model and a height reduction model through model training, realizes the automatic extraction of the DSM vegetation range and the automatic height reduction of the vegetation range, assists in the height reduction process of the DSM vegetation area. Operators do not need to manually extract the vegetation range and give height reduction reference values for each area for vegetation height reduction, reducing the labor intensity of operators and improving the operation efficiency. Specific Embodiment

[0042] Referring to Figure 1 , a method for automatically reducing the height of DSM vegetation based on a deep learning model. The basic idea is to use the RFRNET deep learning model to perform regional overall height reduction on the DSM containing vegetation, and then by adding vegetation and non-vegetation classification information, replacing the non-vegetation area, and optimizing the height reduction result to form the final DSM height reduction result.

[0043] 1. Overall height reduction of DSM area based on RFRNET

[0044] RFRNET (Recurrent Feature Reasoning Net) is a deep learning model. This model takes an image and the area to be filled (also called the hole area), and calculates the repaired image. Introducing the image repair network into DSM height reduction, the area to be reduced in height can be regarded as the hole area. After the reduction area is repaired, it can be regarded as a height reduction process.

[0045] The Recurrent Feature Reasoning module (RFR) is used to progressively reduce the area to be filled, and it consists of the following three modules:

[0046] 1) Area Identification Module: It is used to calculate the area that should be filled in the current round;

[0047] 2) Feature Reasoning Module: For the area to be filled in this round output by the Area Identification Module, use the network to fill it;

[0048] 3) Feature Merging Operation: It is used to merge intermediate features.

[0049] The process of overall height reduction of DSM area based on RFRNET:

[0050] First, construct the training dataset for vegetation height reduction. Then, use the training dataset for vegetation height reduction to train the RFRNET model to form the DSM vegetation automatic height reduction model. Then, use the DSM vegetation automatic height reduction model to perform model inference on the DSM dataset to be reduced in height to form the height reduction result of the entire DSM map. The process is as follows Figure 2 as shown.

[0051] 1) The process of constructing the training dataset for vegetation height reduction

[0052] To train the vegetation automatic height reduction model, three types of data need to be prepared: DSM, DEM, and vegetation classification information raster data.

[0053] The input of the height reduction model is DSM data and vegetation classification raster information, and DEM is the supervision information for model training. The pixels of DSM, DEM, and vegetation analysis raster information correspond one by one, and they have the same height and width.

[0054] 2) The process of constructing the vegetation automatic height reduction model

[0055] The construction of the vegetation automatic height reduction model mainly includes four steps: dataset division, data cropping, data filtering, and model training.

[0056] The dataset is divided into a training set and a validation set at a ratio of 8:2. The prepared vegetation height reduction training dataset is divided into a training set and a validation set at a ratio of 8:2.

[0057] After the data is divided, the training data and validation data are cropped. The "sliding window overlapping method" is used to crop the DSM data, DEM data, and the corresponding DOM data of vegetation classification. To increase the proportion of subgraph samples that contain both vegetation and non-vegetation subgraphs after cropping, the window size of the "sliding window overlapping method" is set to 512, and the step size is set to 128.

[0058] After cropping, the samples are filtered. If all pixels in the subgraph sample are non-vegetation, then the corresponding DSM data are all non-height reduction areas; if all pixels in the subgraph sample are vegetation, then the corresponding DSM data are all height reduction areas. These two situations are not conducive to the height reduction model learning the elevation value differences between vegetation and non-vegetation areas. Therefore, these two types of samples are filtered out from the samples, and data containing both vegetation and non-vegetation areas are used for model training.

[0059] During the model training process, a model trained on the general COCO dataset is used as the pre-trained model. The number of iterations is set to 300,000 times, and the training time is determined according to the specific hardware and software configurations. After each model iteration is completed, the commonly used error metric mean absolute error (MAE) is used to statistically analyze the metrics of the model on the validation set, and the mean absolute error (MAE) is required to be less than 6 meters.

[0060] 2. Vegetation / non-vegetation classification based on HROCR

[0061] Construct a deep learning model for HROCR vegetation recognition based on the OCR module to extract the vegetation areas in the data to be height reduced.

[0062] The backbone network of the model is planned to use the HRNet model. HRNet maintains a high-resolution feature representation throughout the process and repeatedly exchanges information between parallel multi-resolution subnets for multi-scale fusion. The high-resolution features and low-resolution features enhance each other, giving it strong performance; the OCR module represents objects through pixels closely related to the object, solving the problem of missing context semantic information.

[0063] The vegetation / non-vegetation classification process based on the HROCR model is as follows: First, a training dataset for vegetation recognition is constructed, then the HROCR model is trained using the training dataset for vegetation recognition to form a DOM vegetation recognition model, and finally, the DOM vegetation / non-vegetation classification result is obtained by using the DOM vegetation recognition model to perform model inference on the DOM dataset to be recognized. The process is as follows Figure 3 as shown

[0064] 1) Construction of the training dataset for vegetation recognition

[0065] To train the vegetation recognition model, vegetation images and corresponding classification label data are required. Among them, the images are the input for training the model, and the image classification label data is the supervision information for training the model. In the process of constructing the vegetation recognition dataset, geographical national conditions monitoring data is used to merge vegetation land types to form vegetation areas, and non-vegetation areas are merged to form non-vegetation areas. Vector refinement correction is performed on areas where the vector and image texture are significantly inconsistent. Finally, the extracted vegetation vectors are rasterized, and after rasterization, binary raster classification data of vegetation and non-vegetation is obtained. This raster is required to have the same height and width as the corresponding image. The pixels in the vegetation area of the image are marked with the number "1", and all other non-vegetation area pixels are marked with the number "0".

[0066] 2) Construction of the vegetation recognition model

[0067] The construction of the vegetation recognition model mainly has three links: data division, data cropping, and model training.

[0068] In the data division process, the regional vegetation recognition dataset is divided into a training set and a validation set according to an 8:2 ratio of the original data. When dividing the dataset, it is necessary to ensure the consistency of the data distribution between the training set and the validation set as much as possible. Specifically, all vegetation types should be covered in the training set, and the validation set should also correspondingly contain the vegetation types that appear in the training set. Such a training set and validation set can improve the robustness of the model to different vegetation types from the data level and avoid the situation of excessive data distribution deviation between the training set and the validation set.

[0069] In the data cropping process, the "sliding window overlapping method" is used to crop the original image. Using this method can largely reduce the obvious traces of the small image edges on the large image due to inaccurate recognition of the small image edge areas during the stitching of large images. For the prediction results of each sliding window, only the middle valid part of the results is taken to stitch the large image.

[0070] The model training process uses a model trained on the general COCO dataset as the pre-trained model, sets the number of iterations to 320,000 times, and determines the training time according to the specific hardware and software configurations. For the model after each iteration, the commonly used recall and precision in semantic segmentation are used to statistically analyze the indicators of the model on the validation set, and it is required that both the recall and precision can meet the index requirement of 90%.

[0071] 3. Obtain the DSM results containing the original non-vegetation information

[0072] Based on the obtained DSM full-map height reduction results, the operator adds the height reduction results to the operation software, makes a preliminary evaluation of the height reduction results, and determines the local correction area and correction value by sampling. Combining with the DSM full-map height reduction results forms the height reduction corrected results. After correction, according to the DOM vegetation / non-vegetation classification results, the elevation values in the non-vegetation areas are replaced to form the elevation replacement results. The process is as follows Figure 4 as shown

[0073] 4. Optimization of the height reduction results

[0074] The optimization of the height reduction results is to set a buffer zone at the edge of the vegetation and non-vegetation areas on the elevation replacement results in the non-vegetation areas, and perform Gaussian smoothing in the buffer zone to obtain the final height reduction results. The process is as follows Figure 5 as shown Specific embodiments

[0075] Refer to Figure 6 As shown, a method for automatically generating DEM results by reducing the height of vegetation areas in DSM data is proposed. This method applies two deep learning models, HROCR and the Recursive Feature Reasoning Network (RFRNET), to train the models for specific operation areas respectively, constituting a vegetation classification model and a vegetation automatic height reduction model. Through data reasoning, vegetation classification results and vegetation automatic height reduction results are automatically generated. The vegetation classification results are added to the vegetation automatic height reduction results. After a series of result optimization work, the final DSM vegetation automatic height reduction results are formed.

[0076] 1. First, make model training datasets in a specific operation area, including a vegetation recognition training dataset and a vegetation automatic height reduction training dataset.

[0077] The vegetation recognition training dataset includes several scenes of images and vegetation classification marker information within a specific area, and the quantity should be sufficient to meet the model accuracy requirements. The images are required to have uniform hues, clear textures, and moderate brightness; the vegetation marker information can be obtained from collecting national geographical conditions monitoring data and merging vegetation and non-vegetation, or from collecting vegetation information in images. When classifying and marking, the vegetation areas are marked as "1", and the non-vegetation areas are marked as "0" as the background. After classification and marking, it is necessary to rasterize the classification and marking vector, and the rasterized classification and marking information should have the same height and width as the original image in the same area and the same resolution.

[0078] The vegetation automatic height reduction training dataset includes several scenes of DSM, DEM, and vegetation classification information within a specific area, and the quantity should be sufficient to meet the model accuracy requirements. DSM and DEM are obtained through the method of collecting results, and the vegetation classification information can be obtained from collecting national geographical conditions monitoring data and merging vegetation and non-vegetation, or from collecting vegetation information in images. DSM, DEM, and vegetation classification information should have the same height and width and the same resolution.

[0079] 2. Then, add the vegetation recognition training dataset to the HROCR deep learning model for model training. Divide the input vegetation recognition training dataset into a training set and a validation set according to a ratio of 8:2. Then, use the "sliding window overlapping method" to crop the images. Finally, set the model parameters for iterative training, calculate the recall rate and precision rate indicators for each iteration of the model. After the model training is completed, select the model with the best indicators as the final vegetation recognition model, and the recall rate and precision rate indicators should both reach over 90%.

[0080] Add the vegetation automatic height reduction training dataset to the RFRNET deep learning model for model training. Divide the input vegetation automatic height reduction training dataset into a training set and a validation set according to a ratio of 8:2. Then, use the "sliding window overlapping method" to crop the images. Then, filter out the cropped images that are all vegetation and all non-vegetation. Finally, set the model parameters for iterative training, calculate the mean absolute error indicator for each iteration of the model, and the average error should be less than 6 meters.

[0081] 3. Then, use the trained vegetation recognition model and vegetation automatic height reduction model to perform inferences on the DOM and DSM at the same location within the specific area respectively, to obtain the DSM full-map height reduction results and vegetation classification results. Among them, the DSM full-map height reduction results need to be determined by the operator for local correction values for height reduction correction. Segment the original DSM and the corrected DSM full-map height reduction results according to the range of the vegetation classification results, and use the non-forest area of the original DSM to replace the non-forest area of the corrected DSM full-map height reduction results, and finally form the elevation replacement results.

[0082] 4 Then, based on the elevation replacement result, set the forest land and non-forest land buffers and perform smoothing processing using Gaussian smoothing to form the final result after height reduction.

[0083] Through the present invention, a brand-new DEM production method is proposed. The vegetation height reduction is automatically realized by the DSM through a deep learning model. Operators can abandon the traditional manual operation method, without the need to manually extract the vegetation range and without giving height reduction reference values for each area for vegetation height reduction, reducing the labor intensity of the operators and improving the production efficiency.

[0084] The series of detailed descriptions listed above are only specific descriptions of the feasible implementation manners of the present invention, and they are not used to limit the protection scope of the present invention. Any equivalent implementation manners or changes made without departing from the technical spirit of the present invention should be included in the protection scope of the present invention.

[0085] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0086] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for automatically reducing the height of DSM vegetation based on a deep learning model, characterized in that, It includes the following steps: Construct a vegetation recognition model and a vegetation automatic height reduction model; Obtain the vegetation classification results through the vegetation recognition model and the DSM full-map height reduction results through inference using the vegetation automatic height reduction model respectively; Segment the original DSM and the corrected DSM full-map height reduction results according to the scope of the vegetation classification results, and use the non-forest area of the original DSM to replace the non-forest area of the corrected DSM full-map height reduction results to finally form the elevation replacement results; Set the forest and non-forest buffers based on the elevation replacement results and perform smoothing processing using Gaussian smoothing to form the final height reduction results; The composition of the DSM full-map height reduction results specifically includes: First, construct the vegetation height reduction training dataset, then use the vegetation height reduction training dataset to train the RFRNET model to form the DSM vegetation automatic height reduction model, and then use the DSM vegetation automatic height reduction model to perform model inference on the DSM dataset to be height-reduced to form the DSM full-map height reduction results.

2. The automatic vegetation height reduction method for DSM based on a deep learning model according to claim 1, characterized in that, Among them, the DSM full-map height reduction results need to be corrected by the operator to determine the local correction value.

3. A DSM vegetation automatic height reduction method based on a deep learning model according to claim 1, characterized in that Perform inference on the DOM and DSM at the same location in a specific area using the trained vegetation recognition model and vegetation automatic height reduction model respectively.

4. A DSM vegetation automatic height reduction method based on a deep learning model according to claim 1, characterized in that Constructing the vegetation recognition model includes: Construct the vegetation recognition training dataset, then use the vegetation recognition training dataset to train the HROCR model to form the DOM vegetation recognition model, and then use the DOM vegetation recognition model to perform model inference on the DOM dataset to be recognized to form the DOM vegetation / non-vegetation classification results.

5. A DSM vegetation automatic height reduction method based on a deep learning model according to claim 4, characterized in that, The construction of the vegetation recognition model mainly has three links: data division, data cropping, and model training: In the data division process, the regional vegetation recognition dataset is divided into a training set and a validation set according to the ratio of 8:2 for the original data; In the data cropping process, the "sliding window overlapping method" is used to crop the original image. Using this method can largely reduce the obvious traces of the small image edges on the large image due to inaccurate recognition of the small image edge areas during the stitching of the large image. For the prediction results of each sliding window, only the middle valid part of the results is taken to stitch the large image; In the model training process, the model trained on the general COCO dataset is used as the pre-training model, the number of iterations is set, and for the model after each iteration, the recall rate and precision rate commonly used in semantic segmentation are used to statistically analyze the indicators of the model on the validation set.

6. A DSM vegetation automatic height reduction method based on a deep learning model according to claim 1, characterized in that, Constructing the vegetation automatic height reduction model specifically includes dataset division, data cropping, data filtering, and model training: In the dataset division, the prepared vegetation height reduction training dataset is divided into a training set and a validation set according to the ratio of 8:2; After the data is divided, crop the training data and validation data; After the cropping is completed, filter the samples, and use the data that includes both vegetation and non-vegetation areas for model training; The model training process uses a model trained on the general COCO dataset as the pre-training model, sets the number of iterations, and after each model iteration is completed, uses the commonly used mean absolute error in error metrics to statistically analyze the metrics of the model on the validation set.

7. A DSM vegetation automatic height reduction method based on a deep learning model according to claim 6, characterized in that, The cropping method uses the "sliding window overlapping method" to crop DSM data, DEM data, and the corresponding DOM data of vegetation classification.

8. A DSM vegetation automatic height reduction method based on a deep learning model according to claim 7, characterized in that, To increase the sample proportion of subgraphs that contain both vegetation and non-vegetation subgraphs after cropping, the window size of the "sliding window overlapping method" is set to 512, and the step size is set to 128.

Images