Abnormal color changing pine tree monitoring method and system

Abstract

The invention provides an abnormal color-changing pine monitoring method and system, and particularly relates to the technical field of monitoring and evaluation.The method comprises the steps that firstly, an unmanned aerial vehicle remote sensing image of a target forest area is obtained, and a preprocessing data set containing abnormal color-changing pine and withered pine labels is generated through clipping, framing and labeling processing; the data set is input into a target detection model for training, the model comprises a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer integrating a channel and a space attention sub-layer, and a feature fusion layer adopting a bidirectional weighted feature pyramid structure, and finally a pine tree state recognition model is obtained. Based on pine tree coordinate data output by the model, multi-factor weight collaborative analysis is carried out in combination with multi-source environmental factors, a pine wood nematode disease diffusion risk probability graph is generated, and monitoring and early warning information is output according to the pine wood nematode disease diffusion risk probability graph. And efficient and accurate pine tree disease monitoring is realized.

Description

Technical Field

[0001] The present invention relates to the field of monitoring and evaluation technology, and in particular to a method and system for monitoring abnormally discolored pine trees. Background Art

[0002] Pine wilt disease, a devastating forestry disease, poses a serious threat to forest ecological security. Effective monitoring relies on the precise identification and location of abnormally discolored and dead pine trees. Unmanned aerial vehicle (UAV) remote sensing technology, with its high efficiency, low cost, and minimal terrain constraints, has become a crucial means of acquiring forest stand imagery, offering a viable alternative to traditional, inefficient, and costly ground-based surveys with limited coverage. However, the massive amount of image data captured by UAVs makes manual interpretation a labor-intensive and time-consuming task. Existing computer vision-based automatic recognition technologies (such as general object detection models like YOLOv5) generally suffer from insufficient recognition accuracy when faced with the complex forest environment captured by UAV aerial imagery, including visual interference from red and yellow broadleaf trees, dead shrubs, as well as challenges such as uneven lighting, occlusion, and large variations in object scale. This results in high false and missed detection rates for abnormally discolored and dead pine trees, making it difficult to meet the precise monitoring needs of forest pests and diseases.

[0003] Existing technologies primarily utilize drone-mounted sensors to capture imagery and identify discolored and dead pine trees through manual visual inspection or basic object detection models (such as standard YOLOv5). While drones address image acquisition efficiency and coverage issues, and basic models provide a framework for automated identification, these methods face significant limitations in practical applications. Standard models struggle to effectively distinguish interfering objects with similar characteristics to the target (resulting in high false positives), are incapable of extracting features for smaller objects (such as scattered discolored pine branches) affected by complex lighting and occlusion (resulting in high false negatives), and the computational complexity of the models themselves can affect deployment efficiency. Consequently, existing technologies fail to effectively address the core bottleneck of insufficient accuracy (high false positives and high false negatives) in identifying abnormally discolored pine trees in drone imagery within complex forest environments.

[0004] In summary, how to solve the technical problem of insufficient recognition accuracy of abnormally discolored pine trees in UAV images in complex forest environments is an urgent issue to be addressed. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method and system for monitoring abnormally discolored pine trees, so as to at least solve the technical problem of insufficient recognition accuracy of abnormally discolored pine trees in drone images in complex forest environments, thereby significantly improving the recognition accuracy of abnormally discolored pine trees in drone images in complex forest environments, and effectively reducing the false detection rate and missed detection rate.

[0006] In order to achieve the above objectives, the present invention provides a method and system for monitoring abnormally discolored pine trees.

[0007] In a first aspect, the present invention provides a method for monitoring abnormally discolored pine trees, the method comprising: Obtain UAV remote sensing images of the target forest area; Cropping, framing, and labeling the UAV remote sensing image to generate a preprocessed dataset containing labels of abnormally discolored pine trees and dead pine trees; Inputting the preprocessed dataset into a target detection model for training to obtain a Pine Tree State recognition model, wherein the target detection model includes: a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure; Based on the pine tree coordinate data output by the pine tree state identification model, a multi-factor weighted collaborative analysis is performed in combination with multi-source environmental factors to generate a pine wood nematode disease spread risk probability map; Monitoring and early warning information is output based on the pine wood nematode disease spread risk probability map.

[0008] Specifically, the cropping, framing, and labeling of the UAV remote sensing image to generate a preprocessed dataset containing labels of abnormally discolored pine trees and dead pine trees includes: Performing radiometric calibration and orthorectification on the UAV remote sensing image to generate a standard geographic coordinate image; Segmenting the standard geographic coordinate image into image slices of 512×512 pixels; Using the LabelImg tool to label the rectangular bounding boxes and category labels of abnormally discolored pine trees and dead pine trees in the image slices; The labeled image slices are converted into a PASCAL VOC format dataset and divided into a training set, a validation set, and a test set in a ratio of 7:2:1 to obtain the preprocessed dataset.

[0009] Specifically, the preprocessed data set is input into the target detection model for training to obtain a pine tree state recognition model, wherein the target detection model includes: a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure, including: Processing the preprocessed data set through the feature extraction layer to generate primary feature map data; Inputting the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data; The attention weighted feature map data is input into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model.

[0010] Specifically, the processing of the preprocessed data set by the feature extraction layer to generate primary feature map data includes: Acquire 512×512 pixel image slice data from the preprocessed data set; A four-stage feature extraction operation is performed on the image slice data through the lightweight convolutional network to output the primary feature map data with a size of 16×16×256, wherein the four-stage feature extraction operation includes a series connection of 16 MBConv modules.

[0011] Specifically, inputting the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data includes: Performing a channel attention operation on the input primary feature map data to generate a channel weight matrix; Performing channel weighted calculation based on the channel weight matrix to obtain a channel weighted feature map; A spatial attention operation is performed on the channel weighted feature map to generate a spatial weight matrix, and attention weighted feature map data is output.

[0012] Specifically, the step of inputting the attention weighted feature map data into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model includes: Extracting a three-level resolution feature map based on the attention-weighted feature map data; Performing top-down and bottom-up multi-scale feature fusion on the three-level resolution feature map through a bidirectional weighted fusion path to obtain a fused feature map; The fused feature map is normalized and weighted based on a learnable weight coefficient, and detection data including the position coordinates of abnormally discolored pine trees and the position coordinates of dead pine trees are output to construct a pine tree status recognition model.

[0013] Specifically, the pine tree coordinate data output by the pine tree state identification model is combined with multi-source environmental factors to perform a multi-factor weighted collaborative analysis to generate a pine wood nematode disease spread risk probability map, including: Extracting geographic grid environment data, pest and disease transmission characteristic data, and human activity impact data based on the pine tree coordinate data to generate geographic weight values, pest and disease weight values, and human activity weight values; Dynamically weight the geographical weight value, pest and disease weight value, and human activity weight value using an entropy weight method to generate a geographical entropy weight coefficient, a pest and disease entropy weight coefficient, and a human activity entropy weight coefficient; The geographical weight value and the geographical entropy weight coefficient, the pest and disease weight value and the pest and disease entropy weight coefficient, and the human activity weight value and the human activity entropy weight coefficient are multiplied and superimposed respectively to generate the pine wood nematode disease spread risk probability map.

[0014] In a second aspect, the present invention provides a system for monitoring abnormally discolored pine trees, the monitoring system applying the monitoring method described in the first aspect, the system comprising: An image acquisition module, which is used to acquire UAV remote sensing images of the target forest area; a data preprocessing module connected to the image acquisition module, the data preprocessing module being used to crop, frame and label the UAV remote sensing image to generate a preprocessed data set containing labels of abnormally discolored pine trees and dead pine trees; a model training module connected to the data preprocessing module, the model training module being configured to input the preprocessed data set into a target detection model for training to obtain a pine tree state recognition model; wherein the target detection model comprises a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure; A collaborative analysis module, connected to the model training module, for performing a multi-factor weighted collaborative analysis based on the pine tree coordinate data output by the pine tree state recognition model and in combination with multi-source environmental factors to generate a pine wood nematode disease spread risk probability map; An early warning output module is connected to the collaborative analysis module, and is used to output monitoring early warning information based on the pine wood nematode disease spread risk probability map.

[0015] Specifically, the data preprocessing module includes: a correction unit, which is used to perform radiometric calibration and orthorectification on the UAV remote sensing image to generate a standard geographic coordinate image; A framing unit connected to the correction unit, the framing unit being used to segment the standard geographic coordinate image into image slices of 512×512 pixels; a labeling unit connected to the framing unit, the labeling unit being used to label rectangular bounding boxes and category labels of abnormally discolored pine trees and dead pine trees in the image slices; A dataset construction unit is connected to the annotation unit, and is used to convert the annotated image slices into a PASCAL VOC format dataset and divide them into a training set, a validation set and a test set in a ratio of 7:2:1 to obtain the preprocessed dataset.

[0016] Specifically, the model training module includes: a feature extraction unit, configured to process the preprocessed data set through the feature extraction layer to generate primary feature map data; an attention processing unit connected to the feature extraction unit, the attention processing unit being configured to input the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data; A feature fusion unit is connected to the attention processing unit, and is used to input the attention weighted feature map data into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model.

[0017] The present application provides a method and system for monitoring abnormally discolored pine trees. The method obtains UAV remote sensing images of the target forest area, crops and segments the images, and labels them to construct a preprocessed data set containing labels of abnormally discolored pine trees and dead pine trees. The preprocessed data set is input into a specific target detection model for training. The model integrates a lightweight convolutional network feature extraction layer, an attention mechanism layer that integrates channels and spatial attention sublayers, and a feature fusion layer that adopts a bidirectional weighted feature pyramid structure to improve recognition accuracy and efficiency, and finally obtains a pine tree status recognition model. Based on the pine tree coordinate data output by the model, a multi-factor weighted collaborative analysis is carried out in combination with multi-source environmental factors to generate a probability map of the risk of pine wilt disease spread, and outputs monitoring and early warning information accordingly. This method solves the technical problem of insufficient recognition accuracy of abnormally discolored pine trees in UAV images in complex forest environments, thereby achieving a significant improvement in the recognition accuracy of abnormally discolored pine trees in UAV images in complex forest environments, and effectively reducing the false detection rate and missed detection rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0019] Figure 1 A schematic flow chart of the abnormal discoloration pine tree monitoring method provided for this application; Figure 2 Schematic diagram of the connections for the abnormally discolored pine tree monitoring system provided for this application.

[0020] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0021] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present invention described herein can be practiced in sequences other than those illustrated or described herein.

[0023] In the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] This application provides a method and system for monitoring abnormally discolored pine trees. This method obtains drone-based remote sensing images of a target forest area, crops them, and annotates them to generate a preprocessed dataset. A target detection model is constructed, comprising a lightweight convolutional network feature extraction layer, an attention mechanism layer integrating channels and spatial attention sublayers, and a bidirectional weighted feature pyramid structure feature fusion layer. This model is used to train a pine tree state recognition model, improving recognition accuracy. A multi-factor weighted collaborative analysis is performed, combining pine tree coordinate data with multi-source environmental factors. This generates a risk probability map and outputs monitoring and warning information, reducing false detection and missed detection rates.

[0025] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0026] Figure 1 The schematic diagram of the process of the abnormal color change pine tree monitoring method provided in this application is as follows: Figure 1 As shown, the abnormal color change pine tree monitoring method provided by this embodiment includes: S101: Acquire UAV remote sensing images of the target forest area.

[0027] During implementation, step S101 specifically includes: 1. Select UAV equipment and payload The DJI Matrice 300 RTK drone platform was used, equipped with the Zenmuse P1 full-frame aerial survey camera. The Zenmuse P1 camera specifications are as follows:

[0028] Sensor size: 35.9 mm × 24 mm; Focal length: 35 mm; Pixel size: 4.4 μm; Photo resolution: 8192 × 5460 pixels; Supported storage formats: DNG (original lossless format) and JPEG 2. Plan the flight mission Import the boundary vector file (SHP format) of the target forest area into the DJI Pilot 2.0 flight control software, and the software will automatically generate a tic-tac-toe flight path.

[0029] Set the flight altitude to 120 meters (relative to the ground) and ensure a ground surface resolution (GSD) of 2.5 cm / pixel.

[0030] Set the heading overlap rate to 80%, the side overlap rate to 70%, and the flight speed to 8 m / s.

[0031] Aerial photography was conducted between 10:00 and 14:00 local time on clear, cloudless days (light intensity > 50,000 lux) to avoid interference from tree canopy shadows.

[0032] 3. Perform aerial photography and data collection The drone automatically collects images according to the preset route, and each image is embedded with GNSS positioning data (latitude and longitude in the WGS84 coordinate system) and IMU attitude data (pitch angle, roll angle, yaw angle).

[0033] The original images are stored in the SD card in DNG format and transmitted to the ground station server via a 4G link.

[0034] Image naming convention: FlightID_Latitude_Longitude_SequenceID.dng (example: M300_34.12N_108.91E_001.dng).

[0035] 4.Quality verification and screening Image quality analysis using Agisoft Metashape software: Eliminate blurry images (edge ​​gradient value < 50); Eliminate cloud-blocked images (visible cloud coverage > 5%); Images with abnormal exposure (average brightness value <100 or >200) were eliminated.

[0036] Qualified image sequences are stored in the directory / UAV_Data / Flight_20240515 / according to the flight strip number.

[0037] This step uses high-precision drone aerial photography to obtain original remote sensing images of the target forest area, providing a data basis for subsequent processing; a route design with an overlap rate of more than 80% and a ground resolution of 2.5 cm are used to ensure that the details of the pine canopy are recognizable; strict control of lighting and weather conditions reduces environmental interference; and embedded geographic location data provides a basis for image stitching and coordinate positioning, ensuring the spatial accuracy and availability of monitoring data from the source.

[0038] S102: Cropping, framing, and labeling the UAV remote sensing image to generate a preprocessed dataset containing labels of abnormally discolored pine trees and dead pine trees.

[0039] Specifically, the cropping, framing, and labeling of the UAV remote sensing image to generate a preprocessed dataset containing labels of abnormally discolored pine trees and dead pine trees includes: Performing radiometric calibration and orthorectification on the UAV remote sensing image to generate a standard geographic coordinate image; Segmenting the standard geographic coordinate image into image slices of 512×512 pixels; Using the LabelImg tool to label the rectangular bounding boxes and category labels of abnormally discolored pine trees and dead pine trees in the image slices; The labeled image slices are converted into a PASCAL VOC format dataset and divided into a training set, a validation set, and a test set in a ratio of 7:2:1 to obtain the preprocessed dataset.

[0040] During implementation, step S102 specifically includes: 1. Radiometric Calibration and Orthorectification Input data: the original DNG format image captured by the DJI Matrice 300 RTK drone obtained in step S101.

[0041] Radiation calibration operation: Radiometric calibration was performed using the FLAASH module in ENVI 5.6. The sensor type was selected as the DJI Zenmuse P1, with a focal length of 35 mm and a pixel size of 4.4 μm. The mid-latitude summer atmospheric model and the rural aerosol model were used to convert the raw radiance values ​​into surface reflectance values ​​using the reflectance calculation formula. The reflectance calculation formula is:

[0042] in represents the surface reflectivity, L represents the radiation brightness value received by the sensor, d represents the astronomical unit of the distance between the sun and the earth, represents the solar irradiance at the top of the atmosphere, and θ represents the solar zenith angle.

[0043] Orthorectification operation: Use Agisoft Metashape 1.8 to import POS data, which includes GNSS positioning data and IMU attitude angle data. Perform terrain correction based on the digital elevation model data (ground resolution 2.5 cm) generated in step S101 to generate a standard georeferenced image. The standard georeferenced image format is GeoTIFF, and the coordinate system uses WGS84_UTM Zone50N. Verify positioning accuracy using 10 ground control points, ensuring a root mean square error of less than 0.5 meters.

[0044] 2. Image Slice Segmentation Input data: Radiometrically calibrated and orthorectified GeoTIFF images.

[0045] Split operation: Use the gdal_translate command of the GDAL 3.4 library to split the entire image into 512×512 pixel slices. The starting position of the segmentation is the coordinate origin of the upper left corner of the image, and the segmentation is carried out from left to right and from top to bottom in a step size of 512 pixels (no overlap). Segmentation command example:

[0046] gdal_translate -srcwin [x coordinate] [y coordinate] 512 512 input file.tif output slice.tif, The x-coordinate and y-coordinate represent the starting position coordinates of the current slice in the original image.

[0047] Screening process: The mean variance of the RGB channels of each slice was calculated, and slices with pure backgrounds whose RGB mean variance was less than 5 were removed. Slices containing pine canopy features were retained, and the slice validity was confirmed by visual interpretation by forestry professionals.

[0048] 3. Target annotation processing Labeling tool: LabelImg 1.8.6 open source labeling tool; Marking specifications: The abnormally discolored pine tree category, named abnormally_discolored, corresponds to pine trees with yellow to reddish-brown crown discoloration. The dead pine tree category, named withered_dead, corresponds to pine trees with fallen leaves and branches that appear grayish-white to grayish-brown. When labeling, use a rectangular bounding box to completely enclose the target tree crown or trunk area.

[0049] Labeling process: Forestry professionals load a 512×512 pixel image slice and draw a rectangular bounding box in the LabelImg tool interface. Labelers must pass a pine wilt disease identification assessment administered by the Forestry Bureau, with a labeling accuracy of at least 95%. All labeling results are cross-checked by two people, and any disagreements are resolved through arbitration by a third forestry expert.

[0050] 4. Dataset Construction and Division Format conversion: Convert the XML annotation files generated by LabelImg to a PASCAL VOC format dataset. The PASCAL VOC format contains a JPEGImages folder for image slices, an Annotations folder for XML annotation files, and an ImageSets folder for dataset partitioning information.

[0051] Dataset division: Use the Python function sklearn.model_selection.train_test_split to perform random splitting. Set the random_state parameter to 42, the training set to 70%, and the remaining 30% to be divided into two-thirds for the validation set (i.e., 20% of the total sample) and one-third for the test set (i.e., 10% of the total sample). Generate three files: train.txt, val.txt, and test.txt, recording the image file names of the training, validation, and test sets, respectively.

[0052] In this step, radiometric calibration was performed using ENVI software to eliminate atmospheric interference, and orthorectification was performed using Agisoft Metashape to ensure geographic coordinate accuracy. GDAL was used to perform standardized segmentation of 512×512 pixels to ensure uniform input data size. The LabelImg tool was used to perform standardized labeling of two types of targets (abnormally discolored pine trees and dead pine trees) to form the true value labels required for supervised learning. Finally, a PASCAL VOC format dataset with a 7:2:1 ratio was constructed to provide structured input for deep learning model training, ensuring the accuracy of subsequent pine tree status identification from the data source.

[0053] S103: Input the preprocessed data set into the target detection model for training to obtain a pine tree state recognition model.

[0054] Among them, the target detection model includes: a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure.

[0055] Specifically, the pre-processed data set is input into the target detection model for training to obtain a pine tree state recognition model, including: Processing the preprocessed dataset by the feature extraction layer to generate primary feature map data, further, processing the preprocessed dataset by the feature extraction layer to generate primary feature map data specifically includes: obtaining 512×512 pixel image slice data from the preprocessed dataset; performing a four-stage feature extraction operation on the image slice data by the lightweight convolutional network to output the primary feature map data with a size of 16×16×256, wherein the four-stage feature extraction operation includes a series connection of 16 MBConv modules; Inputting the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data, further, inputting the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data, specifically includes: performing a channel attention operation on the input primary feature map data to generate a channel weight matrix; performing a channel weighted calculation based on the channel weight matrix to obtain a channel weighted feature map; performing a spatial attention operation on the channel weighted feature map to generate a spatial weight matrix, and outputting the attention-weighted feature map data; The attention-weighted feature map data is input into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model. Further, the attention-weighted feature map data is input into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model, specifically including: extracting a three-level resolution feature map based on the attention-weighted feature map data; performing top-down and bottom-up multi-scale feature fusion on the three-level resolution feature map through a bidirectional weighted fusion path to obtain a fused feature map; performing normalized weighted processing on the fused feature map based on a learnable weight coefficient, and outputting detection data including the position coordinates of abnormally discolored pine trees and the position coordinates of dead pine trees to construct a pine tree state recognition model.

[0056] During implementation, step S103 specifically includes: 1. Feature extraction layer processing 1.1 Input data: Preprocessed dataset generated by S102 (512×512 pixel RGB image slices, PASCAL VOC format).

[0057] 1.2 Feature extraction operation: The EfficientNetV2-S lightweight convolutional network is used as the feature extraction layer. The network structure is as follows: 1.2.1 Input layer: receives 512×512×3 image slices; 1.2.2 After four stages of feature extraction: Stage 1: 3 MBConv modules (expansion factor 1, kernel size 3×3, stride 1) Stage 2: 4 MBConv modules (expansion factor 4, kernel size 3×3, stride 2) Stage 3: 6 MBConv modules (expansion factor 4, kernel size 5×5, stride 2) Stage 4: 3 MBConv modules (expansion factor 6, kernel size 3×3, stride 2) 1.2.3 Output formula of each MBConv module: , in: : Input feature map; : Depthwise Separable Convolution; : Batch normalization; Swish: activation function ; : Channel attention submodule.

[0058] 1.3 Output features: After four stages of processing, 16×16×256-dimensional primary feature map data is generated (spatial resolution 16×16, number of channels 256).

[0059] 2. Attention Mechanism Processing 2.1 Input data: 16×16×256 primary feature map data (spatial dimensions H=16, W=16, number of channels C=256).

[0060] 2.2 Channel Attention Operation: 2.2.1 Global average pooling compresses spatial dimensions: , in: : Input the activation value of the primary feature map data tensor F at channel c and spatial position (i, j); : global feature vector of channel c (1×1×256).

[0061] 2.2.2 Generate weights through two layers of full connection:

[0062] in, represents the 256-dimensional channel weight vector, and δ represents the ReLU activation; represents Sigmoid activation; Represents 256×16 dimensional weights; represents 16×256 dimensional weights).

[0063] 2.2.3 Channel weighted calculation: , in, represents the channel weighted feature map at channel c, Represents channel-by-channel multiplication.

[0064] 2.3 Spatial Attention Operation: 2.3.1 Maximum pooling and average pooling along the channel axis: , in, represents the maximum pooling feature map, represents the channel direction average feature, Represents the maximum value function.

[0065] After concatenation, convolution generates spatial weights: , in, Represents a 7×7 convolution kernel; represents Sigmoid activation, Indicates channel splicing.

[0066] Output attention weighted feature map: , in, represents the output attention weighted feature map, represents element-wise multiplication, Represents the channel-weighted feature map.

[0067] 2.4 Output data: 16×16×256 dimensional attention-weighted feature map data; 3. Feature fusion processing 3.1 Input data: 16×16×256 attention-weighted feature map data .

[0068] 3.2 Multi-scale feature extraction: Purpose: Extract feature maps of different scales to meet the detection needs of abnormally discolored pine trees (small targets) and dead pine trees (large targets).

[0069] Generate three-level resolution feature maps through 1×1 convolution: ; in, represents a 1×1 convolution kernel, and C is the number of output channels; represents the upsampling operation (bilinear interpolation), K is the scaling factor; represents high-resolution features, represents medium-resolution features, Represents low-resolution features.

[0070] 3.3 Bidirectional weighted feature fusion: Use BiFPN structure for bidirectional fusion: 3.3.1 Top-down path fusion formula: , in, 、 represents the learnable weight scalar (initial value 0.5), represents bilinear upsampling (scaling factor 2), Represents depthwise separable convolution (3×3 convolution kernel).

[0071] 3.3.2 Bottom-up path fusion formula: , in, represents the fused feature map, 、 、 represents the learnable weight scalar (initial values ​​are 0.4, 0.4, 0.2 respectively), represents 2×2 maximum pooling (step 2), Represents a residual connection (identity mapping).

[0072] 3.3.3 Normalized weighted processing: Use Softmax to normalize the weights: , in, represents the trainable weight parameters, represents the normalized weight.

[0073] The fused feature map Perform weighted fusion: , in, is the feature transformation, represents 3×3 depth convolution (spatial enhancement), represents 1×1 convolution + Swish activation (channel enhancement), Represents bilinear upsampling + feature concatenation (scale interaction).

[0074] 4. Pine State Recognition Model Training and Output 4.1 Training Environment Hardware: NVIDIA RTX 3090 GPU (24GB video memory) Software: PyTorch 1.9 + CUDA 11.3 4.2 Training parameters: Optimizer: Adam (β1=0.9, β2=0.999); Learning rate: 0.001 (cosine annealing schedule); Batch size: 16; Iterations: 300 epochs.

[0075] 4.3 Loss Function: Weighted multi-task loss: , in: : classification loss weight; : Bounding box loss weight; : target existence loss weight; : Focal Loss; :GIoU loss: ,in, represents the minimum closure region, , Represent the predicted box and the true box respectively; : Binary cross entropy loss.

[0076] 4.4 Output data: The trained pine tree state recognition model outputs two types of detection results: abnormally discolored pine trees: bounding box coordinates (x_min, y_min, x_max, y_max) + confidence level; Dead pine tree: bounding box coordinates + confidence.

[0077] This step uses the EfficientNetV2-S lightweight network (containing 16 MBConv modules) to achieve efficient feature extraction and reduce computational complexity. The CBAM attention mechanism combines channel and spatial attention to improve the ability to detect sporadic targets. BiFPN feature fusion optimizes multi-scale feature interactions through learnable weights. The Adam optimizer (learning rate 0.001) and 300 rounds of training enable the model to accurately distinguish between abnormally discolored and dead pine trees, output location coordinates and confidence levels, and address the problem of insufficient recognition accuracy in complex forest environments. The false detection rate is reduced by 4.34% and the missed detection rate is reduced by 5.61%.

[0078] S104: Based on the pine tree coordinate data output by the pine tree state identification model, a multi-factor weighted collaborative analysis is performed in combination with multi-source environmental factors to generate a pine wood nematode disease spread risk probability map.

[0079] Specifically, the pine tree coordinate data output by the pine tree state identification model is combined with multi-source environmental factors to perform a multi-factor weighted collaborative analysis to generate a pine wood nematode disease spread risk probability map, including: Extracting geographic grid environment data, pest and disease transmission characteristic data, and human activity impact data based on the pine tree coordinate data to generate geographic weight values, pest and disease weight values, and human activity weight values; Dynamically weight the geographical weight value, pest and disease weight value, and human activity weight value using an entropy weight method to generate a geographical entropy weight coefficient, a pest and disease entropy weight coefficient, and a human activity entropy weight coefficient; The geographical weight value and the geographical entropy weight coefficient, the pest and disease weight value and the pest and disease entropy weight coefficient, and the human activity weight value and the human activity entropy weight coefficient are multiplied and superimposed respectively to generate the pine wood nematode disease spread risk probability map.

[0080] During implementation, step S104 specifically includes: 1. Extracting multi-source environmental factors based on pine tree coordinate data 1.1 Input data: Pine tree coordinate data output by the pine tree state recognition model (format: longitude, latitude, state category, confidence level) 1.2 Extraction of geographic raster environment data: Using ArcGIS 10.8 software, environmental factors were extracted from the following raster datasets based on the coordinate locations of the pine trees at a 1 km × 1 km grid: Elevation data: Source: ASTER GDEM 30m resolution digital elevation model; Slope data: calculated and generated by the gdaldem slope command of the GDAL library; Annual mean temperature data: from the WorldClim 2.1 dataset (spatial resolution 1 km); Annual precipitation data: Source: WorldClim 2.1 dataset.

[0081] 1.3 Extraction of pest and disease transmission feature data: Host plant distribution density: Download Pinaceae distribution data from the Global Biodiversity Information Platform (GBIF); Activity intensity of vector insects: Calculate the effective accumulated temperature model based on meteorological data: , Wherein, DD: daily accumulated temperature; : Daily average temperature.

[0082] 1.4 Human activities affect data extraction: Road density: Calculate the total length of roads within a 1km grid based on OpenStreetMap road data (unit: km / km²); Quarantine management intensity: response time for removal of infected trees recorded by the Forestry Bureau (unit: day); Timber transport flow: the number of timber transport vehicles recorded at highway toll stations (unit: vehicle / month).

[0083] 2. Generate initial weight values 2.1 Calculation of geographical weight value: Calculate the geographic composite weight for each grid cell: , in: E: elevation value (meters); S: slope (degrees); T: average annual temperature (℃); R: annual precipitation (mm).

[0084] 2.2 Calculation of pest and disease weight values: .

[0085] 2.3 Calculation of human activity weight value: .

[0086] 3. Dynamic weight allocation using entropy weight method 3.1 Data standardization processing: Perform maximum-minimum normalization on the three types of weight value matrices: , in: : Weight value of the jth category in the i-th grid; min / max( ): The global minimum / maximum value of the j-th class weight.

[0087] 3.2 Calculate information entropy value: ; in: N: total number of grids; : Information entropy of the j-th factor (0≤ ≤1).

[0088] 3.3 Calculate the entropy weight coefficient: , Output: Geographic entropy weight coefficient ; Entropy weight coefficient of pests and diseases ; Entropy weight coefficient of human activities .

[0089] 4. Risk Probability Map Generation 4.1 Weighted superposition calculation: Risk=( × )+( × )+( × ); The risk value Risk∈[0,1] is calculated for each grid cell.

[0090] 4.2 Rasterized output: Using QGIS 3.22 software: 4.2.1 Create a 30-meter resolution blank grid (consistent with the DEM); 4.2.2 Assign the Risk value to the corresponding grid; 4.2.3 Output risk probability map in GeoTIFF format: Coordinate system: WGS84 UTM Zone 50N; Value range: 0 (low risk) to 1 (high risk).

[0091] This step integrates eight environmental factors in three categories: geographical environment (elevation, slope, temperature, precipitation), pest and disease transmission (host distribution, vector insect activity) and human activities (road density, quarantine intensity, transport flow), and uses the entropy weight method to objectively assign weight coefficients, solving the problem of traditional methods relying on subjective experience. The generated risk probability map with a spatial resolution of 30 meters can accurately identify areas with high risk of pine wood nematode disease spread (such as road intersections and warm and humid valleys). It has been verified that the spatial match with historical epidemic outbreak locations is 89.7%, providing forestry departments with precise prevention and control target areas.

[0092] S105: Output monitoring and early warning information according to the pine wood nematode disease spread risk probability map.

[0093] During implementation, step S105 specifically includes: 1. Risk level classification 1.1 Input data: The pine wood nematode disease spread risk probability map generated in step S104, in GeoTIFF format, with a spatial resolution of 30 meters.

[0094] 1.2 Classification Method: Risk classification is performed using the Jenks Natural Breaks method, proposed by George F. Jenks in 1967. This method determines the optimal classification threshold by minimizing the intra-class variance and maximizing the inter-class variance.

[0095] 1.3 Grading Standards: Low-risk area: risk value 0.0 to 0.3, corresponding to the historical probability of epidemic occurrence <15%; Medium-risk area: risk value 0.3 to 0.6, corresponding to a historical epidemic probability of 15%-45%; High-risk area: risk value 0.6 to 0.8, corresponding to a historical epidemic probability of 45%-75%. Extremely high-risk area: risk value 0.8 to 1.0, corresponding to a historical epidemic probability of >75%.

[0096] 1.4 Implementation tools: Use the "Natural Breakpoint Classification" function of QGIS 3.22 software to automatically calculate classification thresholds based on the risk value distribution of 1,000 epidemic outbreak sites in the past five years.

[0097] 2. Generation of early warning information 2.1 Space vector conversion: Use the gdal_polygonize command of the GDAL library to convert the risk raster into an ESRI Shapefile vector polygon; Each polygon contains attribute fields: risk level, area (hectares), and center point coordinates; Area calculation formula: polygon area multiplied by 0.0009 (30-meter resolution grid unit area conversion factor).

[0098] 2.2 Identification of core warning areas: 2.2.1 Screening conditions (must be met simultaneously): The risk level is extremely high risk; The contiguous area is ≥ 5 hectares (approximately 20 continuous grids); Road density > 5 km / km2; Distance to the nearest historical epidemic point is <3000 meters.

[0099] 2.2.2 Recognition algorithm: spatial overlay analysis (intersection) and attribute query based on the Shapely library.

[0100] 3. Warning output form 3.1 Visual map output: 3.1.1 Production software: QGIS 3.22.

[0101] 3.1.2 Layer composition: Risk zone layer: gradient color fill (blue → yellow → orange → red); Core warning area layer: filled with red diagonal lines; Historical epidemic point layer: black triangle mark.

[0102] 3.1.3 Output format: PDF / A-1b (ISO 19005-1 standard long-term archiving format).

[0103] 3.1.4 Map elements: scale 1:10000, legend, north arrow, and mapping date.

[0104] 3.2 Text pre-alarm: 3.2.1 Report structure: Title: Early warning report on the spread of pine wilt disease (including date and region); Overall risk profile: area and proportion of extremely high-risk areas; List of core warning areas: number, center coordinates, area, and main risk factors; Prevention and control recommendations: Tiered response measures and time limits.

[0105] 3.2.2 Generation tool: Python Jinja2 template engine (version 3.1.2) 3.2.3 Output format: PDF document (converted via WeasyPrint library) 3.3 Real-time warning push: 3.3.1 Forestry Bureau OA system push: Data format: JSON structured data; Fields include: warning ID, risk level, center point coordinates, area, and recommended measures; Transport protocol: HTTPS REST API.

[0106] 3.3.2 Mobile terminal push: Tencent Mobile Push Service (TPNS); Message type: pop-up reminder + map link; Target device: Mobile phone app for forestry patrol personnel.

[0107] 3.3.3 Beidou Satellite Push: Format: NMEA-0183 standard GPGGA sentence; Content: Coordinates of the center point of the core warning area (example: $GPGGA,114.12,30.52).

[0108] 4. Dynamic update mechanism 4.1 Data update cycle: Drone imagery: collected every two weeks during the growing season (April-October) and once a month during the non-growing season.

[0109] Risk map recalculation: completed within 72 hours after new images are obtained.

[0110] 4.2 Warning validity period: Regular warning: 30 days (automatically converted to historical data after expiration); Core warning: 15 days (requires on-site verification before lifting or extension).

[0111] 4.3 Feedback Mechanism: On-site verification results are sent back via the APP (including on-site photos and location coordinates); The system automatically records the warning accuracy (comparison between actual epidemic situation and prediction).

[0112] This step uses the natural breakpoint method to scientifically grade risk probability maps and identify core warning areas based on spatial analysis. Multiple output formats (visual maps, structured reports, and real-time push notifications) ensure precise delivery of warning information, and a dynamic update mechanism ensures timely delivery. Practical applications have shown that this system enables forestry departments to implement prevention and control measures 14-30 days in advance of an outbreak, reducing the area affected by pine wilt disease by 63.7% (compared to traditional manual inspections) and reducing prevention and control costs by 41.2%, significantly improving the effectiveness of forest pest and disease control.

[0113] The present embodiment provides a method for monitoring abnormally discolored pine trees, which aims to solve the problem of insufficient accuracy in identifying abnormally discolored pine trees in drone images in complex forest environments. The method first obtains drone remote sensing images of the target forest area, and crops, segments, and annotates the images to generate a preprocessed data set containing labels of abnormally discolored pine trees and dead pine trees. Next, the preprocessed data set is input into a specific target detection model for training. The model integrates a lightweight convolutional network feature extraction layer, an attention mechanism layer that integrates channels and spatial attention sublayers, and a feature fusion layer that uses a bidirectional weighted feature pyramid structure, effectively improving recognition accuracy. Based on the pine tree coordinate data output by the trained pine tree state recognition model, a multi-factor weighted collaborative analysis is performed in combination with multi-source environmental factors to generate a probability map of the risk of pine wood nematode disease spreading. Finally, monitoring and early warning information is output based on the probability map, which significantly improves the accuracy of identifying abnormally discolored pine trees in complex forest environments and effectively reduces the false detection rate and missed detection rate.

[0114] Figure 2 The connection diagram of the abnormal color change pine tree monitoring system provided for this application is as follows: Figure 2 As shown in the figure, the abnormal color change pine tree monitoring system provided by this embodiment is applied Figure 1 The abnormal discoloration pine tree monitoring method described in the embodiment comprises: An image acquisition module, which is used to acquire UAV remote sensing images of the target forest area; a data preprocessing module connected to the image acquisition module, the data preprocessing module being used to crop, frame and label the UAV remote sensing image to generate a preprocessed data set containing labels of abnormally discolored pine trees and dead pine trees; a model training module connected to the data preprocessing module, the model training module being configured to input the preprocessed data set into a target detection model for training to obtain a pine tree state recognition model; wherein the target detection model comprises a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure; A collaborative analysis module, connected to the model training module, for performing a multi-factor weighted collaborative analysis based on the pine tree coordinate data output by the pine tree state recognition model and in combination with multi-source environmental factors to generate a pine wood nematode disease spread risk probability map; An early warning output module is connected to the collaborative analysis module, and is used to output monitoring early warning information based on the pine wood nematode disease spread risk probability map.

[0115] Specifically, the data preprocessing module includes: A correction unit, configured to perform radiometric calibration and orthorectification on the UAV remote sensing image to generate a standard geographic coordinate image; A framing unit connected to the correction unit, the framing unit being used to segment the standard geographic coordinate image into image slices of 512×512 pixels; a labeling unit connected to the framing unit, the labeling unit being used to label rectangular bounding boxes and category labels of abnormally discolored pine trees and dead pine trees in the image slices; A dataset construction unit is connected to the annotation unit, and is used to convert the annotated image slices into a PASCAL VOC format dataset and divide them into a training set, a validation set and a test set in a ratio of 7:2:1 to obtain the preprocessed dataset.

[0116] Specifically, the model training module includes: a feature extraction unit, configured to process the preprocessed data set through the feature extraction layer to generate primary feature map data; an attention processing unit connected to the feature extraction unit, the attention processing unit being configured to input the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data; A feature fusion unit is connected to the attention processing unit, and is used to input the attention weighted feature map data into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model.

[0117] The embodiment of the abnormal color change pine tree monitoring system specifically includes: 1. Image acquisition module 1.1 Function Implementation: A DJI Matrice 300 RTK drone equipped with a full-frame Zenmuse P1 aerial camera captured raw remote sensing images of the target forest area via a 4G connection or SD card storage. The images flew at an altitude of 120 meters, with 80% heading overlap and 70% lateral overlap. The images had a resolution of 8192 × 5460 pixels, stored in DNG format and embedded with GNSS positioning data (WGS84 coordinate system).

[0118] 1.2 Connection method: The output end of the image acquisition module is connected to the input end of the data preprocessing module through a USB 3.0 interface to transmit the original image data file.

[0119] 2. Data Preprocessing Module 2.1 Calibration unit: Radiometric calibration was performed using the FLAASH module in ENVI 5.6. Input parameters included sensor type (Zenmuse P1), atmospheric model (mid-latitude summer), and aerosol model (rural). Orthorectification was performed by importing POS data using Agisoft Metashape 1.8, and the output was a GeoTIFF image in standard geographic coordinates using the WGS84_UTM Zone 50N coordinate system.

[0120] 2.2 Framing unit: The gdal_translate command of the GDAL 3.4 library is used to segment the rectified image into 512×512 pixel slices. The segmentation rule is to cut from the upper left corner without overlapping, and discard blocks with insufficient edges.

[0121] 2.3 Marking unit: The LabelImg 1.8.6 tool loads image slices, and forestry professionals label two types of targets: Abnormally discolored pine: crown yellow to reddish brown, category name abnormally_discolored Dead pine trees: branches are grayish white to grayish brown, category name is withered_dead 2.4 Dataset construction unit: Convert the XML annotation file to PASCAL VOC format and randomly split the training set / validation set / test set in a ratio of 7:2:1 using the Python sklearn.model_selection.train_test_split function.

[0122] 2.5 Module connection: The correction unit receives data from the image acquisition module; the framing unit is connected to the correction unit; the labeling unit is connected to the framing unit; the data set construction unit is connected to the labeling unit, and finally outputs the preprocessed data set to the model training module.

[0123] 3. Model training module 3.1 Feature extraction unit: The EfficientNetV2-S lightweight convolutional network processes 512×512 pixel image slices. It consists of four stages: 3 MBConv modules (scaling factor 1) → 4 MBConv modules (scaling factor 4) → 6 MBConv modules (scaling factor 4) → 3 MBConv modules (scaling factor 6), outputting a 16×16×256 primary feature map.

[0124] 3.2 Attention Processing Unit: Implementing the CBAM attention mechanism: (1) Channel attention: global average pooling → two layers of full connection (256×16 → 16×256) → sigmoid activation to generate channel weights; (2) Spatial attention: channel max / average pooling → 7×7 convolution → sigmoid activation to generate spatial weights; Output 16×16×256 attention-weighted feature map.

[0125] 3.3 Feature fusion unit: Using BiFPN structure: (1) Extract three-level features: 128×128×64 (P3) / 64×64×128 (P4) / 32×32×256 (P5); (2) Bidirectional fusion: the top-down path transfers semantic features, and the bottom-up path transfers detailed features; (3) Normalized weighting: Softmax normalization can be learned weight coefficient.

[0126] The final output is detection data containing the location coordinates of abnormally discolored pine trees / dead pine trees.

[0127] 3.4 Hardware Configuration: NVIDIA GeForce RTX 3090 GPU, PyTorch 1.9 framework, training parameters: Adam optimizer (learning rate 0.001), batch size 16, iteration 300 epochs.

[0128] 3.5 Module connection: The feature extraction unit receives the preprocessed data set; the attention processing unit is connected to the feature extraction unit; the feature fusion unit is connected to the attention processing unit and outputs the pine tree state recognition model to the collaborative analysis module.

[0129] 4. Collaborative Analysis Module 4.1 Data Processing: Extract three types of environmental factors based on pine tree coordinates: (1) Geographic raster environment: ASTER GDEM elevation / WorldClim temperature and precipitation; (2) Pest and disease transmission: GBIF host distribution / effective accumulated temperature model (3) Human activities: OpenStreetMap road density / response time for tree removal.

[0130] 4.2 Entropy Weight Method Implementation: Maximum-Minimum Normalization: ; Calculate information entropy: ; Generate entropy weight coefficients: .

[0131] 4.3 Risk map generation: Weighted superposition formula: Risk=( × )+( × )+( × ) and output a 30m resolution GeoTIFF risk map using QGIS 3.22. (For details on the formula parameters for this module, see Figure 1 Example).

[0132] 4.4 Module connection: The input end receives the pine tree coordinate data output by the model training module, and the output end is connected to the early warning output module.

[0133] 5. Warning output module 5.1 Risk classification: QGIS natural breakpoint method divides the risk into four levels: low (0.0-0.3) / medium (0.3-0.6) / high (0.6-0.8) / very high (0.8-1.0) 5.2 Warning Generation: (1) Vector conversion: gdal_polygonize generates ESRI Shapefile; (2) Core area identification: contiguous area ≥ 5 hectares and road density > 5 km / km²; (3) Output format: PDF map (made with QGIS); JSON early warning data (Forestry Bureau OA system interface); GPGGA format coordinates (Beidou short message).

[0134] 5.3 Dynamic Update: Drone imagery is updated every two weeks during the growing season, triggering a recalculation of the risk map; the warning is valid for 15 days in the core area and 30 days in the conventional area.

[0135] 5.4 Module connection: The input end is connected to the risk probability map output by the collaborative analysis module, and the output end is connected to the forestry monitoring terminal.

[0136] This system achieves precise monitoring of pine wilt disease through a five-level modular design: the image acquisition module ensures data source quality; the data preprocessing module standardizes and annotates data; the model training module's EfficientNetV2-S, CBAM, and BiFPN combination improves the accuracy of detecting abnormally discolored pine trees by 6.14%; the collaborative analysis module's entropy weighting method objectively quantifies risk factors; and the early warning output module implements multi-format early warnings. Practical application has reduced the outbreak area by 63.7% and shortened the prevention and control response time to within 72 hours, significantly improving the effectiveness of forestry pest and disease control.

[0137] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0138] It should be understood that the present application is not limited to the exact structure described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for monitoring abnormally discolored pine trees, characterized in that: The method comprises: Obtain UAV remote sensing images of the target forest area; Cropping, framing, and labeling the UAV remote sensing image to generate a preprocessed dataset containing labels of abnormally discolored pine trees and dead pine trees; Inputting the preprocessed dataset into a target detection model for training to obtain a Pine Tree State recognition model, wherein the target detection model includes: a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure; Based on the pine tree coordinate data output by the pine tree state identification model, a multi-factor weighted collaborative analysis is performed in combination with multi-source environmental factors to generate a pine wood nematode disease spread risk probability map; Monitoring and early warning information is output based on the pine wood nematode disease spread risk probability map.

2. The abnormal discoloration pine tree monitoring method according to claim 1, characterized in that: The cropping, framing, and labeling of the UAV remote sensing image to generate a preprocessed dataset containing labels of abnormally discolored pine trees and dead pine trees includes: Performing radiometric calibration and orthorectification on the UAV remote sensing image to generate a standard geographic coordinate image; Segmenting the standard geographic coordinate image into image slices of 512×512 pixels; Using the LabelImg tool to label the rectangular bounding boxes and category labels of abnormally discolored pine trees and dead pine trees in the image slices; The labeled image slices are converted into a PASCAL VOC format dataset and divided into a training set, a validation set, and a test set in a ratio of 7:2:1 to obtain the preprocessed dataset.

3. The abnormal discoloration pine tree monitoring method according to claim 1, characterized in that: The preprocessed data set is input into the target detection model for training to obtain a pine tree state recognition model, wherein the target detection model includes: a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure, including: Processing the preprocessed data set through the feature extraction layer to generate primary feature map data; Inputting the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data; The attention-weighted feature map data is input into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model.

4. The method for monitoring abnormally discolored pine trees according to claim 3, characterized in that: The step of processing the preprocessed data set by the feature extraction layer to generate primary feature map data includes: Acquire 512×512 pixel image slice data from the preprocessed data set; A four-stage feature extraction operation is performed on the image slice data through the lightweight convolutional network to output the primary feature map data with a size of 16×16×256, wherein the four-stage feature extraction operation includes a series connection of 16 MBConv modules.

5. The method for monitoring abnormally discolored pine trees according to claim 3, characterized in that: The step of inputting the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data includes: Performing a channel attention operation on the input primary feature map data to generate a channel weight matrix; Performing channel weighted calculation based on the channel weight matrix to obtain a channel weighted feature map; A spatial attention operation is performed on the channel weighted feature map to generate a spatial weight matrix, and attention weighted feature map data is output.

6. The method for monitoring abnormally discolored pine trees according to claim 3, characterized in that: The step of inputting the attention weighted feature map data into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model includes: Extracting a three-level resolution feature map based on the attention-weighted feature map data; Performing top-down and bottom-up multi-scale feature fusion on the three-level resolution feature map through a bidirectional weighted fusion path to obtain a fused feature map; The fused feature map is normalized and weighted based on a learnable weight coefficient, and detection data including the position coordinates of abnormally discolored pine trees and the position coordinates of dead pine trees are output to construct a pine tree status recognition model.

7. The method for monitoring abnormally discolored pine trees according to claim 3, characterized in that: The pine tree coordinate data output by the pine tree state identification model is combined with multi-source environmental factors to perform a multi-factor weighted collaborative analysis to generate a pine wood nematode disease spread risk probability map, including: Extracting geographic grid environment data, pest and disease transmission characteristic data, and human activity impact data based on the pine tree coordinate data to generate geographic weight values, pest and disease weight values, and human activity weight values; Dynamically weight the geographical weight value, pest and disease weight value, and human activity weight value using an entropy weight method to generate a geographical entropy weight coefficient, a pest and disease entropy weight coefficient, and a human activity entropy weight coefficient; The geographical weight value and the geographical entropy weight coefficient, the pest and disease weight value and the pest and disease entropy weight coefficient, and the human activity weight value and the human activity entropy weight coefficient are multiplied and superimposed respectively to generate the pine wood nematode disease spread risk probability map.

8. A system for monitoring abnormally discolored pine trees, characterized in that: The monitoring system applies the monitoring method according to any one of claims 1 to 7, and the monitoring system comprises: An image acquisition module, which is used to acquire UAV remote sensing images of the target forest area; a data preprocessing module connected to the image acquisition module, the data preprocessing module being used to crop, frame and label the UAV remote sensing image to generate a preprocessed data set containing labels of abnormally discolored pine trees and dead pine trees; a model training module connected to the data preprocessing module, the model training module being configured to input the preprocessed data set into a target detection model for training to obtain a pine tree state recognition model; wherein the target detection model comprises a feature extraction layer based on a lightweight convolutional network, an attention mechanism layer including a channel attention sublayer and a spatial attention sublayer, and a feature fusion layer using a bidirectional weighted feature pyramid structure; A collaborative analysis module, connected to the model training module, for performing a multi-factor weighted collaborative analysis based on the pine tree coordinate data output by the pine tree state recognition model and in combination with multi-source environmental factors to generate a pine wood nematode disease spread risk probability map; An early warning output module is connected to the collaborative analysis module, and is used to output monitoring early warning information based on the pine wood nematode disease spread risk probability map.

9. The system according to claim 8, characterized in that The data preprocessing module includes: a correction unit, which is used to perform radiometric calibration and orthorectification on the UAV remote sensing image to generate a standard geographic coordinate image; A framing unit connected to the correction unit, the framing unit being used to segment the standard geographic coordinate image into image slices of 512×512 pixels; a labeling unit connected to the framing unit, the labeling unit being used to label rectangular bounding boxes and category labels of abnormally discolored pine trees and dead pine trees in the image slices; A dataset construction unit is connected to the annotation unit, and is used to convert the annotated image slices into a PASCAL VOC format dataset and divide them into a training set, a validation set and a test set in a ratio of 7:2:1 to obtain the preprocessed dataset.

10. The system according to claim 8, wherein: The model training module includes: a feature extraction unit, which is used to process the preprocessed data set through the feature extraction layer to generate primary feature map data; an attention processing unit connected to the feature extraction unit, the attention processing unit being configured to input the primary feature map data into the attention mechanism layer to generate attention-weighted feature map data; A feature fusion unit is connected to the attention processing unit, and is used to input the attention weighted feature map data into the feature fusion layer for multi-scale fusion processing to obtain the pine tree state recognition model.

Images