A method, device, equipment and storage medium for identifying pine wood discoloration standing trees

Abstract

The present application discloses a method, device, equipment and storage medium for identifying pine discoloration standing trees, which relates to the field of artificial intelligence technology, including: determining characteristic indexes for reflecting the target attributes or conditions of pine discoloration standing trees based on the collected historical pine spectrum data and the basic theory of remote sensing, and merging the characteristic indexes; performing non-overlapping grid clipping, data augmentation and proportional division on the obtained target feature subset to obtain a model training data set; optimizing the initial deep learning model based on the preset model optimization rules, combining the multi-layer target separation model to obtain the initial multi-spectral and deep learning fusion model, and using the model training data set to train the initial multi-spectral and deep learning fusion model; using the obtained target multi-spectral and deep learning fusion model to identify pine discoloration standing trees with the pine spectrum data to be detected to obtain the identification result. It can reduce the influence of factors such as data source and environment to improve the accuracy of pine discoloration standing tree identification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for identifying discolored standing pine trees. Background Art

[0002] Pine wood nematode disease usually requires professional laboratory testing for diagnosis, but sampling and testing one by one in the vast forest is time-consuming and costly, and it is impossible to timely and effectively detect the diseased pine wood in the forest area. Since the most obvious symptom of the early disease is the discoloration of the pine trunk, crown and other parts, the identification of discolored crowns can be used as an early monitoring method to screen potential diseases. However, the existing technical means of identifying discolored standing pine wood are restricted by multiple factors such as limited data sources, complex distribution of forest objects, changeable geographical environment and strict time windows, resulting in significant instability in the overall identification accuracy.

[0003] From the above, it can be seen that how to reduce the influence of factors such as data source, time, space, and geographical environment to improve the accuracy of identifying pine wood discoloration standing trees is an urgent problem that needs to be solved. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a method, device, equipment and storage medium for identifying pine wood discoloration standing trees, which can reduce the influence of factors such as data source, time, space, geographical environment, etc. to improve the accuracy of pine wood discoloration standing tree identification. The specific scheme is as follows:

[0005] In a first aspect, the present application provides a method for identifying pine wood discoloration standing trees, comprising:

[0006] Based on the collected historical pine wood spectral data and basic remote sensing theory, characteristic indexes for reflecting the target attributes or conditions of pine wood discoloration standing trees are determined, and the characteristic indexes are combined to obtain a target characteristic subset;

[0007] By performing non-overlapping grid cropping, data augmentation and proportional division on the target feature subset, a corresponding model training data set is obtained;

[0008] The target deep learning model is obtained by optimizing the initial deep learning model based on the preset model optimization rules, the initial multispectral and deep learning fusion model is obtained by fusing the target deep learning model with the multi-layer target separation model, and the target multispectral and deep learning fusion model is obtained by training the initial multispectral and deep learning fusion model using the model training data set;

[0009] The target multi-spectral and deep learning fusion model is used to identify the discolored standing trees of pine wood on the input pine wood spectrum data to be detected, so as to obtain the corresponding identification results.

[0010] Optionally, the characteristic index used to reflect the target attribute or condition of the discolored standing trees of pine wood is determined based on the collected historical pine wood spectral data and the basic theory of remote sensing, and the characteristic index is combined to obtain a target feature subset, including:

[0011] Collecting historical pine wood spectral data based on a preset imaging device, and analyzing the historical pine wood spectral data to construct a target enhancement index formula;

[0012] The characteristic index corresponding to each of the historical pine wood spectrum data is obtained by using the band data and the target enhancement index formula, and the characteristic indexes corresponding to the historical pine wood spectrum data are merged to obtain a target characteristic subset.

[0013] Optionally, performing non-overlapping grid clipping, data augmentation and proportional division on the target feature subset to obtain a corresponding model training data set includes:

[0014] Performing non-overlapping grid cropping on the target feature subset by using a preset window size and a preset window sliding mode to obtain each image data;

[0015] Based on a preset sample ratio, respectively acquiring images including images of standing trees with discolored pine wood and images not including images of standing trees with discolored pine wood from each of the image data to obtain a pine wood image set and a background image set;

[0016] The pine wood image set and the background image set are rotated and flipped using a preset rotation angle and a preset flipping method to obtain a flipped pine wood image set and a flipped background image set;

[0017] The flipped pine wood image set and the flipped background image set are divided by a preset division ratio to obtain corresponding training sets, verification sets and test sets, and the model training data set is obtained based on the training sets, the verification sets and the test sets.

[0018] Optionally, the optimizing the initial deep learning model based on the preset model optimization rule to obtain the target deep learning model, and fusing the target deep learning model with the multi-layer target separation model to obtain the initial multispectral and deep learning fusion model, including:

[0019] An initial deep learning model is established based on a deep convolutional network, an initial convolutional layer and a pooling layer of the initial deep learning model are modified, and corresponding preset attention mechanism configurations, feature pyramid network configurations, region extraction network configurations, overall loss function configurations, and optimizer configurations are performed to obtain a target deep learning model; the overall loss function includes a classification loss function and a bounding box regression loss function;

[0020] An initial multi-spectral and deep learning fusion model is obtained by fusing the target deep learning model and the multi-layer target separation model.

[0021] Optionally, the using the model training data set to train the initial multispectral and deep learning fusion model to obtain a target multispectral and deep learning fusion model includes:

[0022] The model training data set is sequentially input into the initial multispectral and deep learning fusion model, so as to extract features in the model training data set by using the deep convolutional network in the initial multispectral and deep learning fusion model to obtain a feature map;

[0023] Using the feature pyramid network configured by the initial multi-spectral and deep learning fusion model and the preset attention mechanism to perform feature fusion on the feature map to obtain a pyramid feature map;

[0024] Generate candidate boxes corresponding to each pyramid level based on the configured region extraction network and the pyramid feature map, then optimize the candidate boxes using a binary classification loss function, a bounding box regression loss function, and a stochastic gradient descent algorithm, and remove redundant candidate boxes in each of the candidate boxes using a non-maximum suppression technique to obtain a target candidate box;

[0025] The target candidate frame is aligned using a region of interest alignment technique and a bilinear interpolation technique to obtain an aligned candidate frame; the alignment process includes performing region cropping and feature pooling operations on the target candidate frame;

[0026] Based on the configured binary classification loss function and the bounding box regression loss function, the objects in the aligned candidate frame are respectively classified and the boundaries of the aligned candidate frame are adjusted to obtain the classification results and bounding box coordinates corresponding to the aligned candidate frame, so as to complete the training of the initial multispectral and deep learning fusion model and obtain the trained multispectral and deep learning fusion model;

[0027] The trained multispectral and deep learning fusion model is optimized using a learning rate scheduler and a preset optimization algorithm to obtain a target multispectral and deep learning fusion model.

[0028] Optionally, the target multi-spectral and deep learning fusion model is used to identify the input pine wood spectrum data to be detected as discolored standing trees of pine wood to obtain corresponding identification results, including:

[0029] Matching the input data corresponding to the target multi-spectral and deep learning fusion model based on the data input by the user end, and preprocessing the input pine wood spectral data to be detected by using the size and cropped row and column numbers of the input data to obtain processed spectral data;

[0030] The processed spectral data is detected by the target multi-spectral and deep learning fusion model to obtain output frame coordinates, and the frame coordinates are affine transformed to obtain a vector detection frame;

[0031] Binarizing the vector detection frame to obtain a binary image, using a preset exponential formula combined with a multi-layer target separation algorithm to identify and eliminate areas in the binary image whose pixel values ​​meet preset conditions, so as to obtain an image after elimination;

[0032] Counting pixel values ​​in the eliminated image using Moore voting algorithm, and performing circumscribed rectangle processing and morphological processing on the eliminated image to obtain a processed image; the morphological processing includes erosion and expansion processing on the eliminated image;

[0033] The processed image is extracted to obtain the image centroid, and the target position of the pine wood discolored standing timber is obtained based on the image centroid.

[0034] Optionally, before the target multi-spectral and deep learning fusion model is used to perform pine wood discoloration standing timber identification on the input pine wood spectral data to be detected to obtain a corresponding identification result, the method further includes:

[0035] Using a preset model inference tool to convert the target multispectral and deep learning fusion model into a model format to obtain a converted model;

[0036] Based on a preset integrated development environment, the prediction code and the virtual environment corresponding to the converted model are encapsulated into a dynamic link library, so that the dynamic link library can be called through a software platform.

[0037] In a second aspect, the present application provides a pine wood discoloration standing tree identification device, comprising:

[0038] A feature subset construction module is used to determine a feature index reflecting the target property or condition of pine wood discoloration standing trees based on the collected historical pine wood spectrum data and basic remote sensing theory, and to merge the feature indexes to obtain a target feature subset;

[0039] A data set acquisition module, used for performing non-overlapping grid cropping, data augmentation and proportional division on the target feature subset to obtain a corresponding model training data set;

[0040] A model training module, used to optimize the initial deep learning model based on a preset model optimization rule to obtain a target deep learning model, obtain an initial multispectral and deep learning fusion model by fusing the target deep learning model with a multi-layer target separation model, and train the initial multispectral and deep learning fusion model using the model training data set to obtain a target multispectral and deep learning fusion model;

[0041] The color-changed standing tree recognition module is used to use the target multi-spectral and deep learning fusion model to identify the color-changed standing trees of pine wood on the input pine wood spectral data to be detected, so as to obtain the corresponding recognition results.

[0042] In a third aspect, the present application provides an electronic device, including:

[0043] Memory, used to store computer programs;

[0044] The processor is used to execute the computer program to implement the aforementioned method for identifying pine wood discoloration standing trees.

[0045] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned method for identifying standing trees with discolored pine wood.

[0046] The present application first determines the characteristic index used to reflect the target attributes or conditions of pine wood discoloration standing trees based on the collected historical pine wood spectral data and the basic theory of remote sensing, and merges the characteristic indexes to obtain a target feature subset; the image data corresponding to the target feature subset is subjected to non-overlapping grid cropping, data augmentation and proportional division in sequence to obtain a corresponding model training data set; the initial deep learning model is optimized based on a preset model optimization rule to obtain a target deep learning model, the target deep learning model and a multi-layer target separation model are fused to obtain an initial multispectral and deep learning fusion model, and the initial multispectral and deep learning fusion model is trained using the model training data set to obtain a target multispectral and deep learning fusion model; the target multispectral and deep learning fusion model is used to identify pine wood discoloration standing trees on the input pine wood spectral data to be detected to obtain corresponding identification results.

[0047] As can be seen from the above, this application obtains a characteristic index through historical pine wood spectral data and basic remote sensing theory, and constructs a target feature subset based on the characteristic index. The model training data set is obtained by preprocessing the image data corresponding to the target feature subset, and then the initial deep learning model is optimized based on the preset model optimization rules, and a precise, efficient and adaptive target multi-spectral and deep learning fusion model is constructed in combination with the multi-layer target separation model. In this way, the target multi-spectral and deep learning fusion model is used to identify the pine wood discoloration standing timber input to the pine wood spectral data to be detected, and the automatic, efficient and accurate identification and positioning of the pine wood discoloration standing timber is realized, breaking the data, time and space limitations of the pine wood discoloration standing timber identification, and bringing unprecedented convenience and benefits to forestry management. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0049] Figure 1 A flow chart of a method for identifying discolored standing pine trees disclosed in this application;

[0050] Figure 2 A flowchart of a specific method for identifying discolored standing pine trees disclosed in this application;

[0051] Figure 3 A schematic diagram of the construction and packaging of an initial multispectral and deep learning fusion model provided for this application;

[0052] Figure 4 A flowchart for constructing an initial multispectral and deep learning fusion model provided for this application;

[0053] Figure 5 This is a schematic diagram of the structure of a pine wood color-changing standing tree identification device disclosed in this application;

[0054] Figure 6 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] At present, the identification of pine wood discoloration standing trees is restricted by multiple factors such as limited data sources, complex distribution of objects in forest areas, changeable geographical environment and strict time windows, resulting in significant instability in the comprehensive identification accuracy. To this end, the present application provides a method for identifying pine wood discoloration standing trees, which uses the target multi-spectral and deep learning fusion model to identify pine wood discoloration standing trees with input pine wood spectral data to be detected, and realizes automatic, efficient and accurate identification and positioning of pine wood discoloration standing trees, breaking the data, time and space limitations of pine wood discoloration standing tree identification, and bringing unprecedented convenience and benefits to forestry management.

[0057] See also Figure 1 As shown, the embodiment of the present invention discloses a method for identifying pine wood discoloration standing trees, comprising:

[0058] Step S11, determining characteristic indexes for reflecting target properties or conditions of pine wood discoloration standing trees based on the collected historical pine wood spectrum data and basic remote sensing theory, and merging the characteristic indexes to obtain a target characteristic subset.

[0059] In this embodiment, historical pine wood spectral data are collected based on drone aerial images or satellite remote sensing images. The historical pine wood spectral data can be processed spectral data after preprocessing. The target enhancement index formula is constructed by analyzing the historical pine wood spectral data. The characteristic index used to reflect the target attribute or condition of pine wood discoloration standing trees is determined by the historical pine wood spectral data, the target enhancement index formula and the basic theory of remote sensing, and the characteristic index is merged to obtain a target feature subset. Specifically, the characteristic index used to reflect the target attribute or condition of pine wood discoloration standing trees based on the collected historical pine wood spectral data and the basic theory of remote sensing, and the characteristic index is merged to obtain a target feature subset, including: collecting historical pine wood spectral data based on a preset imaging device, and analyzing the historical pine wood spectral data to construct a target enhancement index formula; determining the characteristic index used to reflect the target attribute or condition of pine wood discoloration standing trees through the historical pine wood spectral data, the target enhancement index formula and the basic theory of remote sensing, and merging the characteristic indexes corresponding to the historical pine wood spectral data to obtain a target feature subset.

[0060] It is understandable that a large amount of pine wood spectral data is collected as historical pine wood spectral data, and multiple spectral curves of various types of objects are drawn based on the historical pine wood spectral data. The difference between different spectral curves is magnified by using attributes such as gradient, trend, and peak to construct an enhancement index, and a feature index subset is constructed based on this to distinguish pine wood discolored standing trees from background objects. The target enhancement index formula is as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] Among them, Blue represents the reflectance of the blue band, Green represents the reflectance of the green band, Red represents the reflectance of the red band, Rededge represents the reflectance of the red edge band, and Nir represents the reflectance of the near-infrared band. α, β, and γ represent different multiplication factors, and δ represents the difference factor. SDSEI_1, SDSEI_2, and SDSEI_3 are all self-developed spectral enhancement indexes, which are used to improve the contrast of discolored standing trees, filter out red objects, and filter out light yellow objects, respectively. NDVI is the traditional normalized vegetation index. The characteristic index of the 4-channel TIFF (Tagged Image File Format) data obtained using the target enhancement index formula is merged into a target feature subset.

[0066] Step S12: performing non-overlapping grid cropping, data augmentation and proportional division on the target feature subset to obtain a corresponding model training data set.

[0067] In this embodiment, the target feature subset is subjected to non-overlapping grid cropping by using a preset window size and a preset window sliding mode to obtain multi-channel image data, and image data containing images of standing trees with discolored pine wood are obtained from the image data based on a preset sample ratio to obtain a pine wood image set, and image data not containing images of standing trees with discolored pine wood are obtained from the image data to obtain a background image set, and the pine wood image set and the background image set are respectively rotated by preset angles and flipped horizontally and vertically, and the flipped pine wood image set and the flipped background image set obtained are divided based on a preset division ratio to obtain corresponding training sets, verification sets and test sets, and the model training data set is obtained through the training set, the verification set and the test set.

[0068] Specifically, the target feature subset is subjected to non-overlapping grid cropping, data augmentation and proportional division to obtain a corresponding model training data set, including: non-overlapping grid cropping of the target feature subset is performed using a preset window size and a preset window sliding method to obtain each image data; based on a preset sample ratio, images containing images of standing trees with discolored pine wood and images not containing images of standing trees with discolored pine wood are respectively obtained from each of the image data to obtain a pine wood image set and a background image set; the pine wood image set and the background image set are rotated and flipped using a preset rotation angle and a preset flipping method to obtain a flipped pine wood image set and a flipped background image set; the flipped pine wood image set and the flipped background image set are divided using a preset division ratio to obtain corresponding training sets, validation sets and test sets, and the model training data set is obtained based on the training set, the validation set and the test set.

[0069] In a specific implementation, a 128*128 cropping window is used to perform window sliding cropping on the target feature subset, and then a set of 10,000 pine images containing images of discolored standing trees and a set of 30,000 background images not containing discolored standing trees are selected to ensure that the ratio of positive and negative samples is 1:3, which is used to improve the generalization ability of the training model. The pine image set and the background image set are rotated 90°, 180°, and 270° and flipped horizontally and vertically for data augmentation, and the augmented pine image set is randomly divided according to the ratio of training set: validation set: test set = 8:1:1, and the augmented background data set not containing discolored standing trees is added to the training set, validation set, and test set according to the same ratio.

[0070] Step S13, optimizing the initial deep learning model based on preset model optimization rules to obtain a target deep learning model, fusing the target deep learning model with the multi-layer target separation model to obtain an initial multispectral and deep learning fusion model, and using the model training data set to train the initial multispectral and deep learning fusion model to obtain a target multispectral and deep learning fusion model.

[0071] In this embodiment, a deep convolutional network is used as a backbone network to establish an initial deep learning model, the initial convolutional layer and pooling layer of the initial deep learning model are modified, and corresponding preset attention mechanism configurations, feature pyramid network configurations, region extraction network configurations, overall loss function configurations, and optimizer configurations are performed to obtain a target deep learning model, and the target deep learning model is combined with a multi-layer target separation model to obtain an initial multi-spectral and deep learning fusion model. Specifically, the initial deep learning model is optimized based on a preset model optimization rule to obtain a target deep learning model, and the initial multi-spectral and deep learning fusion model is obtained by fusing the target deep learning model and the multi-layer target separation model, including: establishing an initial deep learning model based on a deep convolutional network, modifying the initial convolutional layer and pooling layer of the initial deep learning model, and performing corresponding preset attention mechanism configurations, feature pyramid network configurations, region extraction network configurations, overall loss function configurations, and optimizer configurations to obtain a target deep learning model; the overall loss function includes a classification loss function and a border regression loss function; the initial multi-spectral and deep learning fusion model is obtained by fusing the target deep learning model and the multi-layer target separation model.

[0072] In a specific implementation, a ResNeXt (Residual Networks with NextGeneration, a convolutional neural network architecture) deep convolutional network is used as the backbone network, and a CA (Coordinate Attention, coordinate attention mechanism) attention mechanism is embedded. The initial convolutional layer 3 is modified to 4 to apply to the model training data set input with 4 channels and a data format of TIFF. Shallow features with high spatial resolution and rich edge and texture detail information are retained, and deep features with low spatial resolution but more semantic information are extracted. Shallow features are helpful for detecting small objects and local details in discolored standing trees of pine wood or background features, and deep features can better identify features with stronger integrity.

[0073] In this embodiment, a preset attention mechanism is used to extract features of the model training data set to obtain a corresponding feature map, the feature map is fused through a configured feature pyramid network to obtain a pyramid feature map, a candidate box corresponding to each pyramid level is generated using a configured region extraction network and the pyramid feature map, and the candidate box is optimized using an overall loss function and a stochastic gradient descent algorithm, redundant candidate boxes in the candidate box are removed by a non-maximum suppression technique to obtain a target candidate box, and then the target candidate box is aligned using a region of interest alignment technique and a bilinear difference technique to obtain an aligned candidate box, objects in the aligned candidate box are classified based on the configured overall loss function, and the boundary of the aligned candidate box is adjusted to obtain a classification result and a frame coordinate corresponding to the aligned candidate box, so as to complete the training of the initial multispectral and deep learning fusion model to obtain a trained multispectral and deep learning fusion model, and the trained multispectral and deep learning fusion model is optimized using a learning rate scheduler and a preset optimization algorithm to obtain a target multispectral and deep learning fusion model.

[0074] Specifically, the method uses the model training data set to train the initial multispectral and deep learning fusion model to obtain a target multispectral and deep learning fusion model, including: inputting the model training data set into the initial multispectral and deep learning fusion model in sequence, so as to use the deep convolutional network in the initial multispectral and deep learning fusion model to extract features in the model training data set to obtain a feature map; using the feature pyramid network and the preset attention mechanism configured by the initial multispectral and deep learning fusion model to perform feature fusion on the feature map to obtain a pyramid feature map; generating candidate boxes corresponding to each pyramid level based on the configured region extraction network and the pyramid feature map, and then optimizing the candidate boxes using a binary classification loss function, a border regression loss function, and a stochastic gradient descent algorithm, and performing non-maximum suppression on the candidate boxes. The technology removes redundant candidate frames in each of the candidate frames to obtain a target candidate frame; the target candidate frame is aligned using a region of interest alignment technology and a bilinear interpolation technology to obtain an aligned candidate frame; the alignment processing includes region cropping and feature pooling operations on the target candidate frame; based on the configured binary classification loss function and the bounding box regression loss function, the objects in the aligned candidate frame are classified and the boundaries of the aligned candidate frame are adjusted to obtain classification results and bounding box coordinates corresponding to the aligned candidate frame, so as to complete the training of the initial multispectral and deep learning fusion model and obtain a trained multispectral and deep learning fusion model; the trained multispectral and deep learning fusion model is optimized using a learning rate scheduler and a preset optimization algorithm to obtain a target multispectral and deep learning fusion model.

[0075] In a specific implementation, after obtaining the feature map, FPN (Feature Pyramid Networks) is used and CA attention mechanism is embedded to fuse the feature map from top to bottom to obtain a pyramid feature map, and RPN (Region Proposal Network) is used to generate corresponding candidate boxes on each pyramid, and each candidate box is classified and position adjusted by a binary classification loss function and a bounding box regression loss function, and non-maximum suppression technology (NMS) is used to remove redundant candidate boxes in each candidate box, and the confidence is adjusted by using a stochastic gradient descent (SGD) algorithm to obtain a target candidate box, and the RoI Align (Region of Interest Alignment) layer is used to replace the RoI Pooling (Region of interest) in the original network structure. A pooling (i.e., region of interest pooling) layer is used to align the target candidate frame using the region of interest alignment technology and the bilinear interpolation technology to obtain an aligned candidate frame, and then the overall loss function is used to classify the objects in the aligned candidate frame and adjust the boundary of the aligned candidate frame to obtain the classification result and frame coordinates corresponding to the aligned candidate frame, thereby completing the training of the initial multispectral and deep learning fusion model to obtain the trained multispectral and deep learning fusion model; the learning rate is dynamically adjusted using a learning rate scheduler (cosine annealing, step attenuation) to optimize the trained multispectral and deep learning fusion model to obtain the target multispectral and deep learning fusion model.

[0076] In this embodiment, based on the data of different sizes input by the user end, the input data adapted to the target multi-spectral and deep learning fusion model is matched, the input data is processed by a regular grid cropping method, and the position of the upper left corner point of each small image block after cropping in the input data, that is, the cropping row and column number, is recorded, and the input pine wood spectral data to be detected is preprocessed according to the size of the input data and the cropping row and column number to obtain processed spectral data, and then the processed spectral data is detected, and the detected frame coordinates are affine transformed to obtain a vector detection frame, and the vector detection frame is binarized to obtain a binary image, and a preset index formula is used in combination with a multi-layer target separation algorithm to identify and eliminate areas in the binary image whose pixel values ​​meet preset conditions to obtain a eliminated image, and then a Moore voting algorithm is used to count the pixel values ​​in the eliminated image, and the eliminated image is subjected to circumscribed rectangle processing and morphological processing to obtain a processed image, and the processed image is extracted to obtain an image centroid, and the target position of the pine wood discolored standing timber is obtained based on the image centroid.

[0077] Specifically, the method uses the target multi-spectral and deep learning fusion model to identify the discolored standing trees of pine wood on the input pine wood spectral data to be detected to obtain the corresponding identification result, including: matching the input data corresponding to the target multi-spectral and deep learning fusion model based on the data input by the user end, and pre-processing the input pine wood spectral data to be detected by using the size and cropping row and column number of the input data to obtain processed spectral data; detecting the processed spectral data by using the target multi-spectral and deep learning fusion model to obtain the output frame coordinates, and performing affine transformation processing on the frame coordinates to obtain a vector Detection frame; binarizing the vector detection frame to obtain a binary image, using a preset index formula and combining a multi-layer target separation algorithm to identify and eliminate areas in the binary image whose pixel values ​​meet preset conditions to obtain a eliminated image; using a Moore voting algorithm to count the pixel values ​​in the eliminated image, and performing circumscribed rectangle processing and morphological processing on the eliminated image to obtain a processed image; the morphological processing includes corroding and dilating the eliminated image; extracting the processed image to obtain an image centroid, and obtaining the target position of the pine wood discolored standing timber based on the image centroid.

[0078] It can be understood that after obtaining the vector detection frame, the vector detection frame is binarized to obtain a binary image, where 0 is the background and 1 is the pine wood discoloration standing tree target, and then the preset index formula is used in combination with the multi-layer target separation algorithm to identify the area with a pixel value of 1 in the binary image and remove the interference target. The preset index formula is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] Among them, SDI stands for bare soil discrimination index, which is used to separate bare soil; SSI stands for shadow separation index, which is used to remove shadows in images; RSI stands for rock separation index, which is used to remove large areas of rocks in images; RGRI stands for red-green ratio index, which is used to remove other background objects in the image and then extract pine wood discoloration standing wood targets; It is a custom piecewise function used for abnormal data processing in images. a, e, h represent different difference factors, and b, c, d, f, g represent different multiplication factors.

[0085] Step S14: using the target multi-spectral and deep learning fusion model to perform pine wood discoloration standing timber identification on the input pine wood spectral data to be detected, so as to obtain corresponding identification results.

[0086] In this embodiment, the target multi-spectral and deep learning fusion model is converted into a target format model using a preset model reasoning tool, and the prediction code and virtual environment corresponding to the target format model are encapsulated into a dynamic link library based on a preset integrated development environment, so that the dynamic link library can be called through a software platform. Specifically, before using the target multi-spectral and deep learning fusion model to identify pine wood discoloration standing timber for the input pine wood spectral data to be detected to obtain the corresponding identification results, it also includes: using a preset model reasoning tool to convert the model format of the target multi-spectral and deep learning fusion model to obtain a converted model; based on a preset integrated development environment, the prediction code and virtual environment corresponding to the converted model are encapsulated into a dynamic link library, so that the dynamic link library can be called through a software platform.

[0087] In a specific implementation, ONNX Runtime (Open Neural Network Exchange Runtime, i.e., a preset model inference tool) is used to convert the model format of the target multispectral and deep learning fusion model to obtain a converted model, and the prediction code and virtual environment corresponding to the converted model are encapsulated into a DLL (Dynamic Link Library) based on Visual Studio 2019 (i.e., a preset integrated development environment), so that the dynamic link library can be called through the software platform to call the target multispectral and deep learning fusion model.

[0088] As can be seen from the above, the present application obtains a characteristic index through historical pine wood spectral data and the target enhancement index formula, so as to construct a target feature subset based on the characteristic index, obtain a model training data set by performing data preprocessing on the image data corresponding to the target feature subset, and then optimizes the initial deep learning model based on the preset model optimization rules, and combines the multi-layer target separation model to build an accurate, efficient and adaptive target multi-spectral and deep learning fusion model. In this way, the target multi-spectral and deep learning fusion model is used to identify the pine wood discoloration standing timber input to the pine wood spectral data to be detected, so as to realize the automatic, efficient and accurate identification and positioning of the pine wood discoloration standing timber, breaking the data, time and space limitations of the pine wood discoloration standing timber identification, and bringing unprecedented convenience and benefits to forestry management.

[0089] It can be seen from the above embodiments that the present application identifies pine wood discoloration standing timber based on the target multi-spectral and deep learning fusion model for the input pine wood spectral data to be detected. Therefore, the process of identifying pine wood discoloration standing timber based on the target multi-spectral and deep learning fusion model for the input pine wood spectral data to be detected is described.

[0090] See also Figure 2 As shown, the embodiment of the present invention discloses a specific method for identifying pine wood discoloration standing trees, comprising:

[0091] This embodiment first draws multiple spectral curves of various types of ground objects based on the collected historical pine wood spectral data, and uses attributes such as gradient, trend, peak, etc. to amplify the differences between different spectral curves, and constructs multiple feature indexes based on this, and uses the feature indexes to construct a target feature subset. Figure 3 A schematic diagram of constructing and packaging an initial multispectral and deep learning fusion model is provided for this embodiment. After non-overlapping grid cropping, data augmentation and proportional division of the target feature subset, a model training data set is obtained, and then an initial multispectral and deep learning fusion model is constructed.

[0092] Figure 4A flowchart for constructing an initial multi-spectral and deep learning fusion model provided in this embodiment uses a deep convolutional network as the backbone network and embeds a CA attention mechanism. The initial convolutional layer is modified to be suitable for the model training data set input, and the model training data set is feature extracted through a preset attention mechanism to obtain a corresponding feature map, and the attention mechanism is strengthened by channels and coordinates. The sensitivity to spectral fluctuations of discolored parts of pine wood is improved, and the expression of global information and spatial information of feature extraction is enhanced, so that the model can more carefully capture subtle changes in color, texture, etc. in different bands, and improve detection accuracy.

[0093] Then, the feature pyramid network is configured to fuse the feature maps from top to bottom, and the high-resolution feature maps are added layer by layer with the low-resolution feature maps to fuse information of different scales; the CA attention mechanism is embedded in the input layer of FPN, and the model can detect targets at different scales and dimensions by learning the importance weight of each feature channel and adjusting the feature maps accordingly. The generated multi-level feature pyramid strengthens the features related to the target and suppresses irrelevant background noise, so that the model maintains high detection performance when dealing with target scale changes and complex scene changes during the multi-scale feature fusion process.

[0094] By configuring the region extraction network to generate candidate boxes corresponding to each pyramid level in the pyramid feature map obtained by He Rong, the overall loss function and the stochastic gradient descent algorithm are used to optimize the candidate boxes, and the redundant candidate boxes in the candidate boxes are removed by the non-maximum suppression technology to obtain the target candidate box, so that the generated candidate box can be concentrated on the discolored area of ​​the pine wood, and only the most likely candidate area is retained; the RoI Align layer is used to replace the RoI Pooling layer in the original network structure, and the region of interest alignment technology and bilinear difference technology are used to perform regional cropping and feature pooling on the target candidate box, and the feature maps of different sizes corresponding to each candidate box are unified into a fixed size to ensure the consistency of subsequent classification and regression calculations. The accuracy of detection and segmentation is improved, while avoiding the loss of subtle discoloration features, thereby obtaining the aligned candidate box.

[0095] After obtaining the aligned candidate box, the regression layer uses a regressor to further adjust the position and size of the aligned candidate box. The frame regression branch uses a smooth L1 loss to predict and fine-tune the four coordinates of the aligned candidate box, and outputs the final bounding box coordinates to complete the training of the initial multispectral and deep learning fusion model to obtain the trained multispectral and deep learning fusion model, and dynamically adjust the learning rate using a learning rate scheduler (cosine annealing, step decay). After calculating the gradient of the overall loss function, the stochastic gradient descent algorithm is used in combination with momentum to update the convolution kernel weights and bias convolution blocks in the ResNeXt residual block in the network, the parameters of the lateral connection convolution layer and the up- and down-sampling convolution layers of the feature pyramid network part, and the parameters of the fully connected layer. Weight regularization is used to alleviate overfitting, and batch normalization is introduced at multiple locations in the network to alleviate the gradient vanishing or gradient explosion problems, so as to obtain the target multispectral and deep learning fusion model.

[0096] Then, based on the data of different sizes input by the user end, the input data adapted to the target multispectral and deep learning fusion model is matched, and the row and column numbers are cropped in combination with the input data size. The frame coordinates in the regression layer are processed by affine transformation and mapped to the corresponding position of the original image to obtain a vector detection frame, and the vector detection frame is binarized to obtain a binary image. The area whose pixel value meets the preset conditions in the binary image is identified and eliminated by using a preset index formula and a multi-layer target separation algorithm to obtain a eliminated image, and then the Moore voting algorithm is used to count the pixel values ​​of each window area in the eliminated image to remove similar interference objects in the pixels of pine wood discolored standing trees. Since the circularity processing can easily filter out large areas of discolored standing trees regardless of their shapes, the improved circumscribed rectangle processing replaces the circularity processing, which can eliminate the fine and fragmented objects in the result while retaining the pine wood discolored standing trees with irregular crown shapes; the processed image after the circumscribed rectangle processing is corroded and expanded, and the centroid is taken to obtain the target position of the pine wood discolored standing trees.

[0097] This embodiment also utilizes a preset model inference tool to perform model format conversion on the target multi-spectral and deep learning fusion model to obtain a converted model, and encapsulates the prediction code and virtual environment corresponding to the converted model into a dynamic link library based on a preset integrated development environment, so as to call the target multi-spectral and deep learning fusion model through the dynamic link library through a software platform, thereby greatly reducing the difficulty of using the model.

[0098] As can be seen from the above, this application uses a deep convolutional network as the model backbone, configures an attention mechanism, a feature pyramid network, a region extraction network, an overall loss function, and an optimizer, and combines a multi-layer target separation model to construct an initial multispectral and deep learning fusion model. The target feature subset constructed using the feature index of the historical pine wood spectrum data is used to train the initial multispectral and deep learning fusion model to obtain a target multispectral and deep learning fusion model, so as to use the target multispectral and deep learning fusion model to identify the pine wood discoloration standing wood input to be detected pine wood spectrum data, so as to obtain the target position of the pine wood discoloration standing wood. The impact of environmental factors is greatly reduced, breaking the limitations of pine wood discoloration standing wood identification in data, time and space, thereby improving the accuracy of pine wood discoloration standing wood identification.

[0099] Accordingly, see Figure 5 As shown, the present application also provides a pine wood color-changing standing tree identification device, comprising:

[0100] The feature subset construction module 11 is used to determine the feature index used to reflect the target attribute or condition of the discolored standing tree of pine wood based on the collected historical pine wood spectrum data and the basic theory of remote sensing, and to merge the feature indexes to obtain the target feature subset;

[0101] A data set acquisition module 12 is used to obtain a corresponding model training data set by performing non-overlapping grid clipping, data augmentation and proportional division on the target feature subset;

[0102] A model training module 13 is used to optimize the initial deep learning model based on a preset model optimization rule to obtain a target deep learning model, obtain an initial multispectral and deep learning fusion model by fusing the target deep learning model with a multi-layer target separation model, and train the initial multispectral and deep learning fusion model using the model training data set to obtain a target multispectral and deep learning fusion model;

[0103] The discolored standing tree identification module 14 is used to identify the discolored standing trees of pine wood by using the target multi-spectral and deep learning fusion model to input the pine wood spectrum data to be detected, so as to obtain the corresponding identification result.

[0104] As can be seen from the above, the present application obtains a characteristic index through historical pine wood spectral data and the target enhancement index formula, so as to construct a target feature subset based on the characteristic index, obtain a model training data set by performing data preprocessing on the image data corresponding to the target feature subset, and then optimizes the initial deep learning model based on the preset model optimization rules, and combines the multi-layer target separation model to build an accurate, efficient and adaptive target multi-spectral and deep learning fusion model. In this way, the target multi-spectral and deep learning fusion model is used to identify the pine wood discoloration standing timber input to the pine wood spectral data to be detected, so as to realize the automatic, efficient and accurate identification and positioning of the pine wood discoloration standing timber, breaking the data, time and space limitations of the pine wood discoloration standing timber identification, and bringing unprecedented convenience and benefits to forestry management.

[0105] In some specific implementations, the feature subset construction module 11 may specifically include:

[0106] An index enhancement formula construction unit, used for collecting historical pine wood spectrum data based on a preset imaging device, and analyzing the historical pine wood spectrum data to construct a target enhancement index formula;

[0107] The characteristic index merging unit is used to determine the characteristic index reflecting the target attribute or condition of the pine wood discoloration standing tree through the historical pine wood spectral data, the target enhancement index formula and the basic theory of remote sensing, and merge the characteristic index corresponding to the historical pine wood spectral data to obtain a target characteristic subset.

[0108] In some specific implementations, the data set acquisition module 12 may specifically include:

[0109] A grid cropping unit, used for performing non-overlapping grid cropping on the target feature subset by using a preset window size and a preset window sliding mode to obtain each image data;

[0110] An image set construction unit, for respectively acquiring images containing images of standing trees with discolored pine wood and images not containing images of standing trees with discolored pine wood from each of the image data based on a preset sample ratio, so as to obtain a pine wood image set and a background image set;

[0111] An image set flipping unit, used to rotate and flip the pine wood image set and the background image set using a preset rotation angle and a preset flipping method to obtain a flipped pine wood image set and a flipped background image set;

[0112] The image set division unit is used to divide the flipped pine wood image set and the flipped background image set according to a preset division ratio to obtain corresponding training sets, verification sets and test sets, and obtain the model training data set based on the training sets, the verification sets and the test sets.

[0113] In some specific implementations, the model training module 13 may specifically include:

[0114] A learning model building unit, used to establish an initial deep learning model based on a deep convolutional network, modify the initial convolutional layer and pooling layer of the initial deep learning model, and perform corresponding preset attention mechanism configuration, feature pyramid network configuration, region extraction network configuration, overall loss function configuration and optimizer configuration to obtain a target deep learning model; the overall loss function includes a classification loss function and a bounding box regression loss function;

[0115] The fusion model construction unit is used to obtain an initial multi-spectral and deep learning fusion model by fusing the target deep learning model and the multi-layer target separation model.

[0116] In some specific implementations, the model training module 13 may specifically include:

[0117] A feature extraction unit, used to sequentially input the model training data set into the initial multi-spectral and deep learning fusion model, so as to extract features in the model training data set using the deep convolutional network in the initial multi-spectral and deep learning fusion model, so as to obtain a feature map;

[0118] A feature fusion unit, used to perform feature fusion on the feature map using a feature pyramid network configured by the initial multi-spectral and deep learning fusion model and a preset attention mechanism to obtain a pyramid feature map;

[0119] A candidate box removal unit is used to generate candidate boxes corresponding to each pyramid level based on the configured region extraction network and the pyramid feature map, and then optimize the candidate boxes using a binary classification loss function, a bounding box regression loss function, and a stochastic gradient descent algorithm, and remove redundant candidate boxes in each of the candidate boxes by a non-maximum suppression technique to obtain a target candidate box;

[0120] A candidate frame alignment unit, used for aligning the target candidate frame using a region of interest alignment technique and a bilinear interpolation technique to obtain an aligned candidate frame; the alignment process includes performing region cropping and feature pooling operations on the target candidate frame;

[0121] A boundary adjustment unit, used to classify the objects in the aligned candidate frame and adjust the boundary of the aligned candidate frame based on the configured binary classification loss function and the frame regression loss function, respectively, to obtain the classification result and frame coordinates corresponding to the aligned candidate frame, so as to complete the training of the initial multispectral and deep learning fusion model and obtain the trained multispectral and deep learning fusion model;

[0122] The model optimization unit is used to optimize the trained multispectral and deep learning fusion model by using a learning rate scheduler and a preset optimization algorithm to obtain a target multispectral and deep learning fusion model.

[0123] In some specific implementations, the color-changing standing tree identification module 14 may specifically include:

[0124] A spectral data processing unit, used for matching the input data corresponding to the target multi-spectral and deep learning fusion model based on the data input by the user end, and pre-processing the input pine wood spectral data to be detected by using the size and cropping row and column number of the input data to obtain processed spectral data;

[0125] A spectral data detection unit, used to detect the processed spectral data through the target multi-spectral and deep learning fusion model to obtain output frame coordinates, and perform affine transformation processing on the frame coordinates to obtain a vector detection frame;

[0126] A detection frame binarization unit, used to perform binarization processing on the vector detection frame to obtain a binary image, and use a preset exponential formula in combination with a multi-layer target separation algorithm to identify and eliminate areas in the binary image whose pixel values ​​meet preset conditions, so as to obtain an image after elimination;

[0127] A pixel value elimination unit, used for counting pixel values ​​in the eliminated image by using Moore voting algorithm, and performing circumscribed rectangle processing and morphological processing on the eliminated image to obtain a processed image; the morphological processing includes erosion and expansion processing on the eliminated image;

[0128] The image extraction unit is used to extract the processed image to obtain the image centroid, and obtain the target position of the pine wood discolored standing timber based on the image centroid.

[0129] In some specific embodiments, the color-changing standing tree identification device may further include:

[0130] A model format conversion unit, used to perform model format conversion on the target multi-spectral and deep learning fusion model using a preset model reasoning tool to obtain a converted model;

[0131] The code encapsulation unit is used to encapsulate the prediction code and virtual environment corresponding to the converted model into a dynamic link library based on a preset integrated development environment, so as to call the dynamic link library through a software platform.

[0132] Furthermore, the present application also discloses an electronic device. Figure 6This is a structural diagram of an electronic device 20 shown according to an exemplary embodiment, and the content in the figure cannot be regarded as any limitation on the scope of use of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, and the computer program is loaded and executed by the processor 21 to implement the relevant steps in the method for identifying pine wood discoloration standing trees disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0133] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0134] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0135] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program that can be used to complete the pine wood discoloration standing timber identification method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs that can be used to complete other specific tasks.

[0136] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned disclosed method for identifying discolored standing pine trees is implemented. The specific steps of the method can be referred to the corresponding contents disclosed in the aforementioned embodiments, and will not be repeated here.

[0137] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0138] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0139] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0140] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0141] The technical solution provided by the present application is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technicians in this field, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for identifying pine wood discoloration standing trees, characterized in that: include: Based on the collected historical pine wood spectral data and basic remote sensing theory, characteristic indexes for reflecting the target attributes or conditions of pine wood discoloration standing trees are determined, and the characteristic indexes are combined to obtain a target characteristic subset; By performing non-overlapping grid cropping, data augmentation and proportional division on the target feature subset, a corresponding model training data set is obtained; The target deep learning model is obtained by optimizing the initial deep learning model based on the preset model optimization rules, the initial multispectral and deep learning fusion model is obtained by fusing the target deep learning model with the multi-layer target separation model, and the target multispectral and deep learning fusion model is obtained by training the initial multispectral and deep learning fusion model using the model training data set; The target multi-spectral and deep learning fusion model is used to identify the discolored standing trees of pine wood on the input pine wood spectrum data to be detected, so as to obtain the corresponding identification results.

2. The method for identifying pine wood discoloration standing trees according to claim 1, characterized in that: The characteristic index used to reflect the target attribute or condition of pine wood discoloration standing trees is determined based on the collected historical pine wood spectral data and the basic theory of remote sensing, and the target characteristic subset is obtained by merging the characteristic indexes, including: Collecting historical pine wood spectral data based on a preset imaging device, and analyzing the historical pine wood spectral data to construct a target enhancement index formula; The characteristic index used to reflect the target attribute or condition of pine wood discoloration standing trees is determined by the historical pine wood spectral data, the target enhancement index formula and the basic theory of remote sensing, and the characteristic indexes corresponding to the historical pine wood spectral data are merged to obtain a target feature subset.

3. The method for identifying pine wood discoloration standing trees according to claim 1, characterized in that: The method of performing non-overlapping grid clipping, data augmentation and proportional division on the target feature subset to obtain a corresponding model training data set includes: Performing non-overlapping grid cropping on the target feature subset by using a preset window size and a preset window sliding mode to obtain each image data; Based on a preset sample ratio, respectively acquiring images including images of standing trees with discolored pine wood and images not including images of standing trees with discolored pine wood from each of the image data to obtain a pine wood image set and a background image set; The pine wood image set and the background image set are rotated and flipped using a preset rotation angle and a preset flipping method to obtain a flipped pine wood image set and a flipped background image set; The flipped pine wood image set and the flipped background image set are divided by a preset division ratio to obtain corresponding training sets, verification sets and test sets, and the model training data set is obtained based on the training sets, the verification sets and the test sets.

4. The method for identifying pine wood discoloration standing trees according to claim 1, characterized in that: The method optimizes the initial deep learning model based on the preset model optimization rule to obtain a target deep learning model, and obtains an initial multi-spectral and deep learning fusion model by fusing the target deep learning model with the multi-layer target separation model, including: An initial deep learning model is established based on a deep convolutional network, an initial convolutional layer and a pooling layer of the initial deep learning model are modified, and corresponding preset attention mechanism configurations, feature pyramid network configurations, region extraction network configurations, overall loss function configurations, and optimizer configurations are performed to obtain a target deep learning model; the overall loss function includes a classification loss function and a bounding box regression loss function; An initial multi-spectral and deep learning fusion model is obtained by fusing the target deep learning model and the multi-layer target separation model.

5. The method for identifying pine wood discoloration standing trees according to claim 4, characterized in that: The method of using the model training data set to train the initial multispectral and deep learning fusion model to obtain a target multispectral and deep learning fusion model includes: The model training data set is sequentially input into the initial multispectral and deep learning fusion model, so as to extract features in the model training data set by using the deep convolutional network in the initial multispectral and deep learning fusion model to obtain a feature map; Using the feature pyramid network configured by the initial multi-spectral and deep learning fusion model and the preset attention mechanism to perform feature fusion on the feature map to obtain a pyramid feature map; Generate candidate boxes corresponding to each pyramid level based on the configured region extraction network and the pyramid feature map, then optimize the candidate boxes using a binary classification loss function, a bounding box regression loss function, and a stochastic gradient descent algorithm, and remove redundant candidate boxes in each of the candidate boxes using a non-maximum suppression technique to obtain a target candidate box; The target candidate frame is aligned using a region of interest alignment technique and a bilinear interpolation technique to obtain an aligned candidate frame; the alignment process includes performing region cropping and feature pooling operations on the target candidate frame; Based on the configured binary classification loss function and the bounding box regression loss function, the objects in the aligned candidate frame are respectively classified and the boundaries of the aligned candidate frame are adjusted to obtain the classification results and bounding box coordinates corresponding to the aligned candidate frame, so as to complete the training of the initial multispectral and deep learning fusion model and obtain the trained multispectral and deep learning fusion model; The trained multispectral and deep learning fusion model is optimized using a learning rate scheduler and a preset optimization algorithm to obtain a target multispectral and deep learning fusion model.

6. The method for identifying pine wood discoloration standing trees according to claim 5, characterized in that: The method of using the target multi-spectral and deep learning fusion model to identify the input pine wood spectrum data to be detected to obtain the corresponding identification results includes: Matching the input data corresponding to the target multi-spectral and deep learning fusion model based on the data input by the user end, and preprocessing the input pine wood spectral data to be detected by using the size and cropped row and column numbers of the input data to obtain processed spectral data; The processed spectral data is detected by the target multi-spectral and deep learning fusion model to obtain output frame coordinates, and the frame coordinates are affine transformed to obtain a vector detection frame; Binarizing the vector detection frame to obtain a binary image, using a preset exponential formula combined with a multi-layer target separation algorithm to identify and eliminate areas in the binary image whose pixel values ​​meet preset conditions, so as to obtain an image after elimination; Counting pixel values ​​in the eliminated image using Moore voting algorithm, and performing circumscribed rectangle processing and morphological processing on the eliminated image to obtain a processed image; the morphological processing includes erosion and expansion processing on the eliminated image; The processed image is extracted to obtain the image centroid, and the target position of the pine wood discolored standing timber is obtained based on the image centroid.

7. The method for identifying pine wood discoloration standing trees according to any one of claims 1 to 6, characterized in that: Before the target multi-spectral and deep learning fusion model is used to identify the input pine wood spectrum data to be detected to obtain the corresponding identification result, the method further includes: Using a preset model inference tool to convert the target multispectral and deep learning fusion model into a model format to obtain a converted model; Based on a preset integrated development environment, the prediction code and the virtual environment corresponding to the converted model are encapsulated into a dynamic link library, so that the dynamic link library can be called through a software platform.

8. A pine wood discoloration standing tree identification device, characterized in that: include: A feature subset construction module is used to determine a feature index reflecting the target property or condition of pine wood discoloration standing trees based on the collected historical pine wood spectrum data and basic remote sensing theory, and to merge the feature indexes to obtain a target feature subset; A data set acquisition module, used for performing non-overlapping grid cropping, data augmentation and proportional division on the target feature subset to obtain a corresponding model training data set; A model training module, used to optimize the initial deep learning model based on a preset model optimization rule to obtain a target deep learning model, obtain an initial multispectral and deep learning fusion model by fusing the target deep learning model with a multi-layer target separation model, and train the initial multispectral and deep learning fusion model using the model training data set to obtain a target multispectral and deep learning fusion model; The color-changed standing tree recognition module is used to use the target multi-spectral and deep learning fusion model to identify the color-changed standing trees of pine wood on the input pine wood spectral data to be detected, so as to obtain the corresponding recognition results.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor is used to execute the computer program to implement the method for identifying pine wood discoloration standing timber as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the method for identifying pine wood discoloration standing timber according to any one of claims 1 to 7 is implemented.

Images