High and cold grassland rat hole and naked spot recognition model construction method and system based on UAV image and deep learning

Abstract

The invention belongs to the technical field of unmanned aerial vehicle low-altitude remote sensing, and particularly relates to an alpine grassland rat hole and naked spot recognition model construction method and system based on UAV images and deep learning. By optimizing a flight acquisition strategy of the unmanned aerial vehicle and combining with a lightweight deep learning model, a full-process low-power-consumption and quick-response technical scheme from data acquisition to result analysis is constructed, the hardware cost and the field investigation period are reduced, and the defects that traditional manual investigation monitoring is high in cost, long in period and poor in precision are overcome. A private data set is constructed for a specific monitoring scene of the alpine grassland, a highly lightweight scene special identification model is trained, the limitation that a general remote sensing model is poor in effect and large in resource consumption in the specific scene can be overcome, and high-precision rapid identification and detection can still be kept on a low-cost computing platform. By simplifying the data processing flow, the technical operation threshold and implementation complexity are reduced, and a feasible solution is provided for high-frequency and small-range grassland rodent damage monitoring.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude remote sensing technology of unmanned aerial vehicles (UAVs), specifically relating to a method and system for constructing a model for identifying rodent burrows and bare spots in high-altitude grasslands based on UAV imagery and deep learning. Background Technology

[0002] The classification and dynamic monitoring of land cover in grassland ecosystems are crucial foundations for assessing grassland degradation. Based on land cover characteristics, land cover categories can be divided into four elements: vegetation, bare patches, rodent burrows, and others. Among these, rodent burrow distribution density and bare patch area (bald patch rate) are key quantitative indicators, established as core bases for rodent control thresholds and rodent-infested wasteland classification standards in both the Ministry of Agriculture and Rural Affairs industry standard "Technical Specifications for Grassland Rodent Wasteland Management" and the Sichuan Provincial local standard "Technical Regulations for Ground Rodent Control in Grasslands of the Qinghai-Tibet Plateau." With the iterative upgrade of remote sensing technology, UAV low-altitude remote sensing, with its high spatiotemporal resolution advantage, has formed a technical system covering "animal and plant collaborative monitoring"—in the animal monitoring dimension, in addition to conventional wildlife conservation research, it focuses on the dynamic tracking of rodent populations, constructing rodent early warning systems through techniques such as burrow identification and suitable habitat modeling; at the grassland ground monitoring level, it achieves accurate calculation of bald patch rate through multispectral image fusion analysis. Currently, research on grassland rodent monitoring based on UAV remote sensing mainly focuses on rodent burrow identification using algorithms such as target detection, extraction of surface rodent information from multi-source remote sensing data, and rodent habitat zoning studies combining geographic information and ecology. These technological and methodological explorations provide new methodological support for grassland degradation mechanism analysis and ecological restoration. However, there is currently no effective model for identifying rodent burrows and bare patches in alpine grasslands, while manual identification and monitoring are inefficient and have significant limitations in scale. Summary of the Invention

[0003] To address the aforementioned issues, this application provides a method and system for constructing a model for identifying rodent burrows and bare patches in high-altitude grasslands based on UAV imagery and deep learning. By constructing a target detection model for rodent burrows and a semantic segmentation model for bare patches, and using RGB imagery to infer the distribution status and relative area proportions of rodent burrows and bare patches in grasslands, this approach aims to overcome the bottlenecks of traditional manual surveys, promote the development of a precise monitoring and early warning technology system for grassland rodent infestations, and provide technical support for dynamic monitoring, precise early warning, and control decision-making.

[0004] The first technical solution of this application discloses a method for constructing a model for recognizing rodent burrows and bare spots in high-altitude grasslands based on UAV images and deep learning, including the following steps:

[0005] S1. Drone Image Acquisition: Low-altitude RGB image acquisition of the target alpine grassland using drones during the warm season in August and the cold season in November;

[0006] S2. Image segmentation dataset construction: Extract sample data of mouse holes and bare spots from the acquired RGB images, annotate them with Labelme and generate category polygon boundary coordinates, category pixel labels and masks for mouse holes and bare spots, and construct Pascal VOC format image segmentation dataset;

[0007] S3. Model Construction: Using MobileNet as the backbone network, inference was performed on the image segmentation dataset to obtain a segmentation model for rat burrows and bare spots in high-altitude grasslands based on UAV images and deep learning.

[0008] Furthermore, during the image acquisition described in S1, the drone's aerial photography altitude is ≤20m, and the shooting angle is 180° overhead.

[0009] Furthermore, the image segmentation dataset construction described in S2 is specifically as follows:

[0010] Image cropping: Used to crop original-size RGB images to 512×512 pixels;

[0011] Image annotation: labelMe is used to annotate the polygon boundaries of two target categories, mouse holes and bare spots, generating a JSON file containing category labels and boundary coordinates, i.e., a sequence of polygon coordinate points in vector format;

[0012] Mask generation: The labelme2voc.py script in labelme is used to convert the JSON coordinate file into a mask containing pixel labels of category boundaries, that is, a pixel-level representation of polygons in raster mode.

[0013] Tag category: The background pixel category is specified as 0, the bare spot pixel category as 1, and the mouse hole pixel category as 2.

[0014] Furthermore, a model for identifying rodent burrows and bare spots in alpine grasslands based on UAV imagery and deep learning, obtained through the aforementioned model construction method.

[0015] The second technical solution of this application discloses a system for recognizing rodent burrows and bare spots in high-altitude grasslands based on UAV imagery and deep learning, including the following modules:

[0016] Image acquisition module: used for acquiring images of the target alpine grassland;

[0017] Data processing module: used to preprocess the acquired images;

[0018] Segmentation and labeling module: used to perform image segmentation and category labeling on preprocessed images;

[0019] Output module: Used to process the extracted features and segment the mouse hole and bare spot regions of the target image.

[0020] The third technical solution of this application discloses a method for identifying mouse holes and bare spots in alpine grasslands based on UAV images and deep learning. The method uses the above-mentioned model or system to identify mouse holes and bare spots by inputting acquired images.

[0021] And a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor being used to execute the model or system.

[0022] Beneficial effects:

[0023] 1. Highly efficient process and rapid output: By optimizing the drone flight data collection strategy (flight altitude ≤ 20 meters) and adapting it to a lightweight model, a low-power and fast-response end-to-end solution from data collection to analysis results has been built, which effectively reduces hardware costs and field cycle, and solves the pain points of high cost and long cycle of traditional manual survey and monitoring.

[0024] 2. Outstanding performance in model efficiency and specialization: A highly lightweight, dedicated recognition model was trained using a privately constructed dataset specifically designed for monitoring alpine grassland scenarios. This model overcomes the shortcomings of general-purpose remote sensing models in specific scenarios, achieving rapid recognition and detection with high accuracy even on low-cost computing platforms.

[0025] 3. Significant advantages in ease of engineering implementation and application: This solution simplifies the data labeling and processing process and, based on the aforementioned lightweight model, significantly reduces the technical operation threshold and implementation complexity, providing a feasible and applicable solution for high-frequency, small-scale grassland monitoring. Attached Figure Description

[0026] Figure 1 Illustration for aerial image acquisition;

[0027] Figure 2 This is a schematic diagram of sliding window clipping and its corresponding annotation results.

[0028] Figure 3 This is a schematic diagram of the masking process;

[0029] Figure 4 This is a diagram showing the labeled data after cropping (from left to right: cropped, labeled, labeled (training data)).

[0030] Figure 5 Comparison results of images taken from different altitudes by drone;

[0031] Figure 6 Image segmentation results of mouse hole boundary segmentation at different flight altitudes of UAV;

[0032] Figure 7 Example of an object detection dataset;

[0033] Figure 8 The results are the validation set for mouse hole detection;

[0034] Figure 9 A comparison of YOLOv8 mouse hole detection results from original images at a height of 10m in both cold and warm seasons;

[0035] Figure 10 This is a multi-class prediction result based on deep learning image segmentation;

[0036] Figure 11 Read the training labels for different batch sizes;

[0037] Figure 12 The multi-class prediction results of the instance segmentation model;

[0038] Figure 13 This is an example of the interface of a mouse hole target detection system. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0041] Unless otherwise specified, the technical terms in this specification have the same meaning as those generally understood by those skilled in the art; however, in case of any conflict, the definitions in this specification shall prevail.

[0042] Example 1: Constructing a model for identifying rodent burrows and bare spots in high-altitude grasslands based on UAV imagery and deep learning.

[0043] This application focuses on alpine grasslands in Ruoergai County, Sichuan Province (102°08'-103°39' E, 32°56'-34°19' N), located on the eastern edge of the Qinghai-Tibet Plateau, with an altitude ranging from 2380 m to 4532 m. A model for identifying rodent burrows and bare spots in alpine grasslands is constructed.

[0044] S1. Drone Image Acquisition: Low-altitude RGB images of the target alpine grassland will be acquired using drones during the warm season in August and the cold season in November.

[0045] This application utilizes a DJI Mavic 2 Pro drone for high-resolution RGB image acquisition. During the low-altitude remote sensing image acquisition, to avoid image band unevenness caused by rapid cloud movement, relatively clear and windless weather was selected for drone aerial photography of the study area to ensure uniform image quality. After initial screening, a total of 2972 ​​low-altitude images of uniform quality were collected. A schematic diagram of the aerial image acquisition is shown below. Figure 1 As shown.

[0046] S2. Image Segmentation Dataset Construction: Extract sample data of mouse holes and bare spots from the acquired RGB images, and annotate them using Labels (e.g., ...). Figure 2 (As shown) and generate the category polygon boundary coordinates, category pixel labels, and masks (such as) for mouse holes and bare spots. Figure 3 As shown), the Pascal VOC format image segmentation dataset is finally constructed (e.g., Figure 4 As shown), specifically:

[0047] Image cropping: Used to crop the original size (5472×3648) RGB image to 512×512 pixels. Figure 2-5 ;

[0048] Image annotation: labelMe is used to annotate the polygon boundaries of two target categories, mouse holes and bare spots, generating a JSON file containing category labels and boundary coordinates, i.e., a sequence of polygon coordinate points in vector format;

[0049] Mask generation: The labelme2voc.py script in labelme is used to convert the JSON coordinate file into a mask containing pixel labels of category boundaries, that is, a pixel-level representation of polygons in raster mode.

[0050] Tag category: The background pixel category is specified as 0, the bare spot pixel category as 1, and the mouse hole pixel category as 2.

[0051] It should be noted that, in this process, the imagery constructed from the image dataset was also studied based on the drone's shooting altitude, such as... Figure 5 The image shows a comparison of the shooting results taken by the drone at 5m, 10m, 15m, and 20m respectively. Figure 6 To compare the processed images, it can be seen from the figures that there is a significant nonlinear constraint between aerial altitude and the accuracy of ground feature recognition. When the flight altitude increases from 5m to 20m, the sharpness of the mouse hole boundary (edge ​​gradient based on the Sobel operator) decreases by 33.27%, and the Pearson correlation coefficient between aerial altitude and aerial image quality reaches -0.9722 (p < 0.001, R0). 2 =0.9451). Especially when the flight altitude exceeds the 20m threshold, similar to the above situation, the motion blur effect is enhanced due to the vertical increase law of wind speed in the atmospheric boundary layer (the wind speed increases by 0.28±0.05m / s for every 1m increase in altitude). Therefore, the shooting altitude is determined to be ≤20m, and 10m is the optimal altitude.

[0052] In a further implementation, a mouse hole dataset is constructed using RoboFlow and LabelImg. An object detection dataset is constructed, and high-resolution images are processed using a sliding window cropping method. The sample data includes both positive and negative samples. Figure 7 ).

[0053] S3. Model Construction: First, a target detection model for rat holes in high-altitude grasslands is constructed by training the YOLO series target detection models on a private dataset of rat holes in high-altitude grasslands.

[0054] 1. Object Detection

[0055] Based on the PyTorch framework, YOLOv5, YOLOv8, and YOLO11 models were trained on a custom high-altitude grassland mouse burrow dataset. Ultimately, YOLOv8 was determined to be the most suitable model for mouse burrow detection, with a learning rate (lr) of 0.0019406. YOLOv8n achieved optimal results with single-class detection enabled, achieving an accuracy of 84.16% and an mAP50 of 84.89% (Table 1).

[0056] Table 1 Performance evaluation results of different YOLO models

[0057] .

[0058] Note: The evaluation metrics of this object detection model are calculated based on the confusion matrix of the object bounding box prediction. The suffix "single" indicates that single-class detection is enabled, while "multi" indicates that single-class detection is not enabled. "n" and "s" refer to the model size, namely "nano" and "small" respectively.

[0059] YOLOv8 validation set detection results in single-class and multi-class settings are as follows: Figure 8As shown, with an image size of 640, single-class detection did not show any missed detections, while multi-class detection showed either missed detections or duplicate detections. The average confidence score of the former was better than that of the latter.

[0060] In a further implementation, the impact of different sliding window verification overlap rates on the average confidence score of the model was investigated for images of the original size at a height of 10m. Based on a dataset with a cropped size of 640, predictions were made for images of the original size. The overlap rate range was set to [0.01, 0.1], and the overlap rate interval was set to 0.25. The highest average confidence score of 0.475 was obtained when the overlap rate was 0.01, and the highest confidence score of 0.89 was obtained when the overlap rate was 0.25. The number of detected mouse holes was compared with the number of holes in the images as interpreted visually. The optimal overlap rate in this experiment was 0.01. The detection results are as follows: Figure 9 As shown.

[0061] 2. Image segmentation

[0062] Using MobileNet as the backbone network, inference was performed on the image segmentation dataset to obtain a segmentation model for mouse burrows and bare spots in alpine grasslands based on UAV images and deep learning.

[0063] In this step, a field survey of grassland rodent burrows in the study area of ​​Ruoergai County was first conducted to determine the degree of damage in the sample plots. Based on the number of effective rodent burrow entrances within the sample plots, three sample plots with different levels of damage from plateau pikas were selected: plots I, II, and III, representing mild, moderate, and severe damage, respectively. All sample plots contained plateau pikas as the damaging rodent species. The total number of burrow entrances within the sample plots was counted using the burrow-blocking and burrow-robbing method. Specifically, in each sample plot with different levels of damage, three sample circles were randomly selected, with a point chosen as the center and a radius of 28.2 m using a measuring tape, forming a circle with an area of ​​2500 m². 2 The circle is defined, and the total number of burrows within the circle is counted. After the count is completed, all burrows in the area are filled with soil. After 48 hours, the number of burrows reopened by the plateau pika is counted again and recorded as the number of valid burrows (as shown in Table 2).

[0064] Table 2 Information on gopher burrows in the test samples

[0065] .

[0066] Subsequently, based on the semantic segmentation model of DeepLabV3 Plus using PyTorch, MobileNet and Xception were selected as the backbone networks, and the pre-trained weights of the backbone networks were loaded to construct segmentation models for rat burrows and bare spots in high-altitude grasslands, as shown in Table 3.

[0067] Table 3 Hyperparameters and training settings during the initial training phase

[0068] .

[0069] Note: The "freeze" and "unfreeze" training parameters in this table refer to the transfer learning strategy based on pre-trained weights (ImageNet). Freeze phase: Fix the backbone network weights of the pre-trained model and train only the newly added classification layer; Unfreeze phase: Unlock all network layer weights for fine-tuning.

[0070] By constructing the dataset and performing pre-inference on models with Xception and MobileNet as the backbones, it was found that Xception exhibited severe overfitting on the dataset in this study, with network complexity exceeding practical requirements. In contrast, MobileNet showed good fit on the dataset, without underfitting or overfitting. Therefore, MobileNet was chosen as the backbone for inference on the image segmentation dataset, resulting in a segmentation model for mouse burrows and bare spots in high-altitude grasslands based on UAV images and deep learning. Using MobileNet as the feature extraction backbone yielded good results without pre-trained model inference; improvements in mIoU and mPA indicate improved performance in segmentation tasks using MobileNet as the backbone. During model inference, the network image input size was 512×512, with a downsampling factor of 16, and the number of classes was 3 (background-bare spot-mouse burrow) and 2 (background-mouse burrow), respectively.

[0071] like Figure 10 The image shows the binary classification prediction results of the input image using this model. As can be seen, the model can accurately predict mouse holes, achieving a high level of accuracy on datasets with irregular edges. Under the initial training hyperparameter settings, the model's mean precision (MP) was 91.72%. After adjusting the batch size and optimizing the dataset samples, the performance improved by approximately 1.45%. The model effectively avoids the risk of training interruption due to out-of-memory errors (OOM) and has a lightweight advantage. The final accuracy was 93.17%, as shown in Table 4. Verification through manual visual interpretation showed that the optimal interpretation height for the adjusted model was 10m.

[0072] During the unfrozen training phase, the model overfitted at epoch=78. Considering the imbalance of samples, the sample weight coefficients were redistributed, which improved the accuracy of naked spot segmentation. However, the overall performance was lower than that of single-class segmentation, with an accuracy of 74.41%. The results are shown in Table 4.

[0073] Table 4 Evaluation Results of Multi-Classification Model

[0074] .

[0075] 3. Instance segmentation

[0076] In a further implementation of this study, a YOLO-based model was used to achieve good segmentation results in mouse hole detection and instance segmentation on the dataset used in this study. The input image size was 640×640 pixels, identical to all the data construction steps described in S2 except for the sliding window cropping preset size of 640×640. Then, different training batch sizes (8 and 16) were set to compare the dataset's fit to the model and determine the model's suitability. Figure 11 As shown.

[0077] Finally, in the further implementation of this study, a training batch size of 8 was determined to yield better results, as shown in Table 5. The best performance was achieved for hole detection alone (mAP50=0.995), with fast inference speed (total speed approximately 21.7ms / image), as shown in Table 5. Figure 12 As shown.

[0078] Table 5. Model Performance Evaluation Results

[0079] .

[0080] Meanwhile, based on the research process and methods of step S2, a preliminary YOLOv8-based rodent burrow monitoring system for high-altitude grasslands was constructed. This system can detect and count the number of rodent burrows in low-altitude visible light UAV images by uploading UAV imagery and setting the overlap rate (sliding window step size), window size, and confidence threshold. It can achieve rapid counting of rodent burrows in a single image. Figure 13 As shown.

[0081] Therefore, those skilled in the art will recognize that although embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Therefore, the scope of the present invention should be understood and recognized as covering all such other variations or modifications.

Claims

1. A method for constructing a model for identifying rodent burrows and bare spots in high-altitude grasslands based on UAV imagery and deep learning, characterized in that, Includes the following steps: S1. Drone Image Acquisition: Low-altitude RGB image acquisition of the target alpine grassland using drones during the warm season in August and the cold season in November; S2. Image segmentation dataset construction: Extract sample data of mouse holes and bare spots from the acquired RGB images, label them with labels and generate category polygon boundary coordinates, category pixel labels and masks for mouse holes and bare spots, and construct PascalVOC format image segmentation dataset; S3. Model Construction: Using MobileNet as the backbone network, inference was performed on the image segmentation dataset to obtain a segmentation model for rat burrows and bare spots in high-altitude grasslands based on UAV images and deep learning.

2. The model construction method according to claim 1, characterized in that, When acquiring images as described in S1, the drone's aerial photography altitude is ≤20m, and the shooting angle is 180° overhead.

3. The model construction method according to claim 1, characterized in that, The image segmentation dataset construction described in S2 is specifically as follows: Image cropping: Used to crop original-size RGB images to 512×512 pixels; Image annotation: LabelMe is used to annotate the polygon boundaries of two target categories, mouse holes and bare spots, generating a JSON file containing category labels and boundary coordinates, i.e., a sequence of polygon coordinate points in vector format; Mask generation: The Labelme2voc.py script in Labelme is used to convert the JSON coordinate file into a mask containing pixel labels of category boundaries, that is, a pixel-level representation of polygons in raster mode; Tag category: The background pixel category is specified as 0, the bare spot pixel category as 1, and the mouse hole pixel category as 2.

4. A model for identifying rodent burrows and bare spots in alpine grasslands based on UAV imagery and deep learning, obtained by the model construction method according to any one of claims 1-3.

5. A system for identifying rodent burrows and bare spots in high-altitude grasslands based on UAV imagery and deep learning, characterized in that, Includes the following modules: Image acquisition module: used for acquiring images of the target alpine grassland; Data processing module: used to preprocess the acquired images; Segmentation and labeling module: used to perform image segmentation and category labeling on preprocessed images; Output module: Used to process the extracted features and segment the mouse hole and bare spot regions of the target image.

6. A method for identifying rodent burrows and bare spots in high-altitude grasslands based on UAV imagery and deep learning, characterized in that, Using the model described in claim 4 or the system described in claim 5, input UAV images to identify mouse holes and bare spots.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor being configured to execute the model of claim 4 or the system of claim 5.

Images