A method and system for monitoring debris flow based on artificial intelligence vision, and a readable storage medium

Abstract

This invention relates to the field of debris flow disaster monitoring, specifically to a debris flow monitoring method, system, and readable storage medium based on artificial intelligence vision. The method involves collecting and preprocessing multi-source debris flow image data; dividing the multi-source debris flow image data into several datasets; improving the attention mechanism of the neural network model; training the neural network model using the several datasets until the model converges; and using the trained neural network model to process the debris flow image sequence to obtain debris flow event characteristics. By introducing an improved attention mechanism into the neural network model, pixel-level accurate recognition is achieved, effectively filtering out environmental interference, adapting to different natural scenarios, and enabling real-time and accurate monitoring and early warning of debris flows, thus improving the reliability of debris flow identification.

Description

Technical Field

[0001] This invention relates to the field of debris flow disaster monitoring, and in particular to a debris flow monitoring method, system, and readable storage medium based on artificial intelligence vision. Background Technology

[0002] Debris flows, as a geological hazard, are characterized by their suddenness, high speed, and destructive power, posing a significant threat to the safety of railways, highways, and surrounding facilities. Current technologies employ methods such as rainfall threshold comparison, vibration sensing, and infrasound detection for debris flow monitoring. However, debris flow sites are complex environments with weak surface textures and variable backgrounds. Furthermore, debris flows often occur during rainy weather, where raindrops and airflow disturbances can interfere with monitoring results, causing non-target areas to be misidentified as debris flow zones. While existing debris flow monitoring methods have improved monitoring capabilities to some extent, they still cannot overcome these interferences and cannot provide accurate and real-time early warning information in complex environments; the monitoring effect is particularly limited under varying terrain and complex weather conditions. Therefore, identifying the main debris flow movement and filtering background interference in natural scenarios with varying terrain and complex weather conditions is of great significance for improving the real-time performance and reliability of debris flow monitoring and early warning. Summary of the Invention

[0003] Considering that existing debris flow monitoring methods are easily affected by factors such as variable terrain and complex weather conditions in natural scenarios, they cannot provide real-time and accurate monitoring and early warning of debris flows, thus reducing the reliability of debris flow identification.

[0004] To address the above problems, this invention provides a debris flow monitoring method based on artificial intelligence vision, the method comprising:

[0005] Step S1: Collect multi-source debris flow image data; preprocess the multi-source debris flow image data, wherein the preprocessing includes screening preprocessing, normalization preprocessing, multi-dimensional annotation preprocessing, and enhancement preprocessing;

[0006] Step S2: Divide the multi-source debris flow image data into several datasets;

[0007] Step S3: Improve the attention mechanism of the neural network model, and train the neural network model using the aforementioned datasets until the neural network model converges;

[0008] Step S4: Use the trained neural network model to process the debris flow image sequence to obtain debris flow event characteristics.

[0009] Preferably, in step S1, collecting multi-source debris flow image data includes:

[0010] Based on the work logs of several debris flow image sources, the image data generation attributes of each debris flow image source are determined; wherein, the image data generation attributes include the image data generation time;

[0011] Based on the image data generation attributes, the image data collection operations for each debris flow image source are determined, thereby collecting multi-source debris flow image data; wherein, the multi-source debris flow image data includes images taken by UAVs, images taken by ground monitoring points, satellite remote sensing images, and images taken by experimental cameras.

[0012] Preferably, in step S1, the multi-source debris flow image data undergoes screening preprocessing and standardization preprocessing, including:

[0013] Frame extraction is performed on the multi-source debris flow image data to obtain several image frame data;

[0014] Each image frame data undergoes quality screening preprocessing and image normalization preprocessing; wherein, the quality screening preprocessing includes screening each image frame data for resolution and distortion; and the image normalization preprocessing includes normalizing each image frame data for brightness and contrast.

[0015] Preferably, in step S1, the multi-source debris flow image data undergoes multi-dimensional annotation preprocessing, including:

[0016] The debris flow subject and debris flow edge are labeled in each image frame of the multi-source debris flow image data;

[0017] The debris flow main body annotation includes pixel annotation of the debris flow main body in the flowing state in each image frame data; the debris flow edge annotation includes point-by-point polygon annotation of the actual edge of the debris flow in each image frame data.

[0018] Preferably, in step S1, the multi-source debris flow image data undergoes enhancement preprocessing, including:

[0019] Geometric transformation enhancement, color space enhancement, and hybrid enhancement are performed on each image frame data under the multi-source debris flow image data.

[0020] The geometric transformation enhancement includes rotating, translating, or scaling each image frame data; the color space enhancement includes HSV color space dithering, brightness and contrast adjustment for each image frame data; and the hybrid enhancement includes cropping and stitching at least two image frame data.

[0021] Preferably, step S2 includes:

[0022] The category and coordinate information of each image frame data under the multi-source debris flow image data are obtained by the multi-dimensional annotation preprocessing, and each image frame data is matched with its corresponding category and coordinate information;

[0023] The multi-source debris flow image data is divided into training dataset, test dataset, and validation dataset according to the corresponding proportions of all image frames.

[0024] Preferably, step S3 includes:

[0025] A CBAM attention mechanism module is embedded in the last A2C2f module at the backbone end of the YOLOv12n network; the channel attention mechanism and spatial attention mechanism under the CBAM attention mechanism module are dually optimized;

[0026] The YOLOv12n network is trained, tested, and validated using the training dataset, test dataset, and validation dataset contained in the aforementioned datasets until the YOLOv12n network converges.

[0027] Preferably, step S4 includes:

[0028] The debris flow image to be tested is processed by frame segmentation to obtain the debris flow image sequence;

[0029] The trained neural network model is used to process the image sequence of the debris flow to be tested to obtain the dynamic characteristics of the debris flow event; wherein, the dynamic characteristics include the changes in the debris flow flow rate and the spatial coverage area of ​​the debris flow event over time.

[0030] The present invention also provides a debris flow monitoring system based on artificial intelligence vision, the system comprising:

[0031] The data collection module is used to collect multi-source debris flow image data;

[0032] The preprocessing module is used to preprocess the multi-source debris flow image data, wherein the preprocessing includes screening preprocessing, normalization preprocessing, multi-dimensional annotation preprocessing, and enhancement preprocessing.

[0033] The partitioning module is used to divide the multi-source debris flow image data into several datasets.

[0034] An improvement module is used to improve the attention mechanism of neural network models;

[0035] The training module is used to train the neural network model using the aforementioned datasets until the neural network model converges.

[0036] The processing module is used to process the debris flow image sequence under test using the trained neural network model to obtain the debris flow event characteristics.

[0037] The present invention also provides a computer-readable storage medium storing a computer program that, when executed, implements the above-described debris flow monitoring method based on artificial intelligence vision.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This invention discloses an artificial intelligence-based vision-based debris flow monitoring method, system, and readable storage medium. The method involves collecting and preprocessing multi-source debris flow image data; dividing the multi-source debris flow image data into several datasets; improving the attention mechanism of the neural network model; training the neural network model using the several datasets until the model converges; and using the trained neural network model to process the debris flow image sequence to obtain debris flow event features. By introducing an improved attention mechanism into the neural network model, pixel-level accurate recognition is achieved, effectively filtering out environmental interference, adapting to different natural scenarios, and enabling real-time and accurate monitoring and early warning of debris flows, thus improving the reliability of debris flow identification.

[0040] The debris flow monitoring method and system based on artificial intelligence vision of the present invention also have the following advantages: By improving the YOLOv12n network and introducing the CBAM attention mechanism, pixel-level accurate identification of debris flow movement areas with uncertain shapes and varying scales can be achieved. It can effectively filter out complex interferences such as cloud shadows, water surface reflections, splashing raindrops, and trees, significantly improving its anti-interference capability in real-world scenarios. Furthermore, the trained neural network model can adapt to different lighting conditions, weather, and shooting angles, exhibiting good adaptability and generalization. The entire process, from image input to flow velocity output, is fully automated, requiring no manual intervention. Relying on the efficient architecture of YOLOv12n-CBAM, it ensures that the segmentation speed meets the high frame rate requirements of real-time monitoring and early warning. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0042] Figure 1 This is a flowchart of the debris flow monitoring method based on artificial intelligence vision provided by the present invention.

[0043] Figure 2The results are based on debris flow identification using the YOLOv12n network.

[0044] Figure 3 The results are based on the improved YOLOv12n network for debris flow identification.

[0045] Figure 4 This is a structural diagram of a debris flow monitoring method based on artificial intelligence vision. Detailed Implementation

[0046] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and not for limiting the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all structures. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present invention.

[0047] The terms "comprising" and "having," and any variations thereof, used in this invention are intended to cover non-exclusive inclusion. For example, a process, method, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0048] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0049] Please see Figure 1 As shown, this invention provides a debris flow monitoring method based on artificial intelligence vision, the method comprising:

[0050] Step S1: Collect multi-source debris flow image data; preprocess the multi-source debris flow image data, including screening preprocessing, normalization preprocessing, multi-dimensional annotation preprocessing, and enhancement preprocessing.

[0051] Step S2: Divide the multi-source debris flow image data into several datasets;

[0052] Step S3: Improve the attention mechanism of the neural network model by training the neural network model using several datasets until the neural network model converges.

[0053] Step S4: Use the trained neural network model to process the debris flow image sequence to obtain debris flow event characteristics.

[0054] This invention improves the neural network model by introducing an attention mechanism, achieving pixel-level accurate recognition, effectively filtering out environmental interference, adapting to the actual conditions of different natural scenarios, and enabling real-time and accurate monitoring and early warning of debris flows, thereby improving the reliability of debris flow identification.

[0055] Further, in step S1, multi-source debris flow image data is collected, including:

[0056] Based on the work logs of several debris flow image sources, the image data generation attributes of each debris flow image source are determined; among them, the image data generation attributes include the image data generation time.

[0057] Based on the image data generation attributes, the image data collection operations for each debris flow image source are determined, thereby collecting multi-source debris flow image data; among which, multi-source debris flow image data includes images taken by UAVs, images taken by ground monitoring points, satellite remote sensing images, and images taken by experimental cameras.

[0058] Debris flows are closely related to the natural environment, and are understood to be influenced by two main factors: geology and weather. Geological factors include topographic slope, soil composition, and soil compaction; weather factors include rainfall. To obtain historical debris flow data, it is necessary to collect data on the occurrence of historical debris flows from different perspectives, which can involve collecting raw video data from various sources. Preferably, the on-site environment can be filmed using devices such as drones, fixed ground monitoring points, remote sensing satellites, and high-speed cameras to obtain the corresponding raw video data. These devices simultaneously generate work logs during the filming process, which comprehensively record the video image generation status of each device. This video image generation status may include, but is not limited to, the time interval within which each device actually filmed and generated image data. It is understood that only image data generated by each device within the aforementioned time interval can accurately reflect the debris flow occurrence. Each device generates an image data stream during operation. Based on the aforementioned time interval, image data with matching timestamps is collected from the image data stream, and the image data collected from all devices is integrated to form multi-source debris flow image data. For the matching image data generated by each device, the aforementioned multi-source debris flow image data may include, but is not limited to, images taken by drones, images taken by ground monitoring points, satellite remote sensing images, and images taken by experimental cameras; among them, the aforementioned drone images refer to video images taken by drones or other low-altitude aircraft at the debris flow site; the aforementioned ground monitoring point images refer to video images taken by several fixed ground monitoring points distributed at the debris flow site; the aforementioned satellite remote sensing images refer to remote sensing images taken by remote sensing satellites at the debris flow site; and the aforementioned experimental camera images refer to video images taken by high-speed cameras set up in the debris flow flume test scenario.

[0059] Further, in step S1, the multi-source debris flow image data undergoes screening preprocessing and standardization preprocessing, including:

[0060] Frame extraction was performed on the multi-source debris flow image data to obtain several image frame data.

[0061] Each image frame data undergoes quality screening preprocessing and image normalization preprocessing; the quality screening preprocessing includes screening each image frame data for resolution and distortion; the image normalization preprocessing includes normalizing each image frame data for brightness and contrast.

[0062] As discussed above, multi-source debris flow image data includes images captured by drones, ground monitoring points, satellite remote sensing, and experimental cameras. These images differ in video format and image quality. To achieve standardization and consistent processing of multi-source debris flow image data, the data is first segmented and extracted into frames according to a preset video frame rate, resulting in several image frames. Each image within a multi-source debris flow image dataset can yield several image frames. Next, each image frame is filtered for resolution and distortion. For example, frames with resolution below a preset resolution threshold or distortion above a preset distortion threshold are removed, while frames with resolution above the preset resolution threshold and distortion below the preset distortion threshold are retained. Furthermore, brightness and contrast are normalized for each image frame to achieve standardized image transformation.

[0063] Furthermore, in step S1, multi-dimensional annotation preprocessing is performed on the multi-source debris flow image data, including:

[0064] Perform debris flow subject annotation and debris flow edge annotation on each image frame data under the multi-source debris flow image data;

[0065] The debris flow main body annotation includes pixel annotation of the debris flow main body in the flowing state in each image frame data; the debris flow edge annotation includes point-by-point polygon annotation of the actual edge of the debris flow in each image frame data.

[0066] In multi-source debris flow imagery, each image covers the entire area where the debris flow occurred, encompassing both sub-regions where debris flow occurred and sub-regions where no debris flow occurred. Understandably, the sub-regions where debris flow occurred contain all relevant elements and are crucial for identifying the actual debris flow situation. The sub-regions where no debris flow occurred serve as the background for debris flow activity, containing numerous non-debris flow objects that interfere with debris flow identification. To accurately distinguish between debris flow-related and interfering elements in each image frame, annotation of both types of elements is necessary. Specifically, debris flow entities are annotated in each image frame of the multi-source debris flow imagery, with annotations limited to the debris flow entities in a flowing state, excluding stationary sediment deposits, pure water flows, flash floods, landslides, and other non-debris flow objects. Furthermore, debris flow edge annotation is performed on each image frame data under the multi-source debris flow image data. For example, the actual edge of the debris flow is annotated point by point with polygons, and the vertex spacing of the polygon annotations is controlled to ensure boundary accuracy. For areas with blurred boundaries, the area with concentrated solid particles is used as the reference. For partial occlusion, the complete boundary of the visible part is annotated, and occlusion attribute markers are added to the annotation file to ensure the reliability of data annotation.

[0067] Further, in step S1, the multi-source debris flow image data undergoes enhancement preprocessing, including:

[0068] Geometric transformation enhancement, color space enhancement, and hybrid enhancement are performed on the image frames of the multi-source debris flow image data.

[0069] Among them, geometric transformation enhancement includes rotating, translating, or scaling each image frame data; color space enhancement includes HSV color space dithering, brightness and contrast adjustment of each image frame data; and blending enhancement includes cropping and stitching at least two image frame data.

[0070] Geometric transformation enhancement, color space enhancement, and blending enhancement are applied to each image frame data to achieve multimodal conversion and improve the richness of the image frame data. In practice, geometric transformation enhancement using at least one of rotation, translation, or scaling can diversify the geometric shape of the image frame data. Color space enhancement using HSV color space dithering and brightness and contrast adjustments can diversify the color form of the image frame data. Blending enhancement involves cropping and stitching at least two image frames and blending them according to appropriate weights.

[0071] Further, step S2 includes:

[0072] Obtain the category and coordinate information of each image frame data under the multi-source debris flow image data by multi-dimensional annotation preprocessing, and match each image frame data with its corresponding category and coordinate information;

[0073] The multi-source debris flow image data is divided into training dataset, test dataset, and validation dataset according to the corresponding proportions.

[0074] In practice, irregular polygonal bounding boxes can be used to represent the category and coordinate information of each image frame. Specifically, these boxes include the center coordinates, length, width, and category of the bounding box, and are stored in an .xml file, corresponding one-to-one with each image frame. These .xml files need to be converted to .txt files using a suitable program to adapt to the YOLOv12n network's reading format for subsequent YOLOv12n network training. Specifically, all image frame data can be randomly divided into training, testing, and validation datasets in an 8:1:1 ratio, used for training the neural network model, validating the training process, and conducting final performance testing, respectively. The dataset contains two main folders: images and labels. Each main folder contains three subfolders: train, test, and validation, ensuring a strict correspondence between the target images and the standard information, thus constructing the final debris flow target detection dataset.

[0075] Further, step S3 includes:

[0076] The CBAM attention mechanism module is embedded in the last A2C2f module at the backbone end of the YOLOv12n network; the channel attention mechanism and spatial attention mechanism under the CBAM attention mechanism module are optimized.

[0077] The YOLOv12n network is trained, tested, and validated using several datasets, including training, testing, and validation datasets, until the YOLOv12n network converges.

[0078] The A2C2f attention mechanism of the YOLOv12n network has adjustable parameters. This invention optimizes three aspects of debris flow: adjusting the region segmentation strategy, adjusting the receptive field of attention, and optimizing channel attention. Adjusting the region segmentation strategy involves optimizing the number of regions based on the proportion of the image occupied by the debris flow. Adjusting the receptive field of attention involves setting the kernel size of the Position Perceiver in the A2C2f module. The kernel size can be from 5×5 to 9×9, preferably 7×7, to capture the large-scale morphological features and movement trends of the debris flow. Optimizing channel attention involves adjusting the compression ratio of the multilayer perceptron (MLP) in the channel attention branch of the A2C2f module to retain more channel information and enhance the sensitivity to debris flow characteristic channels.

[0079] The CBAM (Convolutional Block Attention Module) attention mechanism is integrated at key locations in the YOLOv12n network. For example, the CBAM attention mechanism module is embedded in the last A2C2f module at the backbone end of the YOLOv12n network. This module performs dual optimization of the extracted high-level semantic features in both channel and spatial dimensions, enhancing the saliency of the main debris flow features and suppressing background interference. The role of the CBAM attention mechanism module in debris flow detection is as follows: In channel attention, it automatically learns and strengthens the feature channels most important for debris flow detection while suppressing irrelevant channels; in spatial attention, after multi-scale fusion, it further refines the debris flow boundary and filters out background noise; in lightweight design, the CBAM attention mechanism module, through efficient design, adds only a small number of parameters and computational cost, without affecting inference speed.

[0080] CBAM attention mechanisms also include channel attention and spatial attention. Channel attention, by modeling the dependencies between channels, adaptively recalibrates the weights of each channel given an input feature map. Where H and W are spatial dimensions, and C is the channel. The calculation process for channel attention is as follows:

[0081]

[0082] In the above formula, The calculated channel attention weights are multiplied channel-by-channel with the input feature map F to obtain the channel attention-optimized feature map F'; AvgPool(F) represents the global average pooling of the feature map F in spatial dimension H×W, and the output dimension is... The vector; MaxPool(F) represents the global max pooling of feature map F in spatial dimension H×W, and outputs a vector of dimension H×W. The vector; MLP() represents a shared multilayer perceptron, whose structure is usually: input C channels → dimension reduction to C / r (r is the preset dimension reduction coefficient) → ReLU (linear rectified function) activation → dimension increase to C channels; This represents the Sigmoid activation function, used to normalize weights to the (0, 1) interval.

[0083] Spatial attention mechanisms focus on the spatial locations with the most information in the feature maps, where the previously obtained feature maps... As input, the calculation process is as follows:

[0084]

[0085] In the above formula, The calculated spatial attention weights are multiplied pointwise by the input feature map F' to obtain the final output feature map of the CBAM attention mechanism; AvgPool(F') represents the average pooling performed along the channel dimension, with an output dimension of... The feature map; MaxPool(F') represents the max pooling operation along the channel dimension, with an output dimension of... Feature map; f 7×7 [] indicates a standard convolutional layer with a 7×7 kernel, used to generate a spatial attention weight map; This represents the Sigmoid activation function, used to normalize weights to the (0, 1) interval.

[0086] The hardware environment configuration, software environment configuration, and training iteration number hyperparameter settings for model training of the present invention are shown in Tables 1-3 below.

[0087] Table 1

[0088] Hardware Name model processor 13th Gen Intel(R) Core(TM) i7-13650HX (2.60 GHz) RAM 24.0 GB operating system Windows 11 GPU NVIDIA GeForce RTX 4060 Laptop GPU

[0089] Table 2

[0090] Software Name model Python Python 3.11.13 Anaconda Conda 24.9.2 OpenCV OpenCV 4.9.0 Pytorch PyTorch 2.6.0 + cu124 CUDA CUDA 12.4

[0091] Table 3

[0092] Hyperparameter name Parameter value Image size 640×640 Batch size 4 Number of iterations 200 Learning rate 0.01 Optimizer Stochastic gradient descent

[0093] The convergence determination and selection of the YOLOv12n network includes: after training is complete, evaluating the detection and generalization performance of the YOLOv12n network using a test dataset. The test dataset is input into the trained YOLOv12n network, and its precision, accuracy (P), recall (R), harmonic mean (F1-score), average precision (AP), and mean average precision (mAP) are calculated. The closer the values ​​of these metrics are to 1, the better the detection performance of the YOLOv12n network. Simultaneously, to meet real-time detection requirements, the frame rate per second (FPS) needs to reach at least 30. Multiple YOLOv12n networks are trained by setting different hyperparameters, and the best-performing YOLOv12n network is selected based on a comprehensive evaluation of various metrics, ultimately applied to real-time debris flow detection.

[0094] Further, step S4 includes:

[0095] The debris flow image to be tested is processed by frame segmentation to obtain the debris flow image sequence;

[0096] The trained neural network model is used to process the image sequence of debris flow to obtain the dynamic characteristics of the debris flow event; among which, the dynamic characteristics include the changes in debris flow flow rate and coverage area over time.

[0097] In practice, the images of the debris flow to be tested are processed frame by frame to obtain the image sequence. This image sequence is then input into a trained neural network model (i.e., the improved YOLOv12n-CBAM model). This neural network model performs real-time inference on each frame of the image sequence and outputs the discrimination result. Because the neural network model integrates the CBAM attention mechanism, it can accurately identify the irregular boundaries of debris flows in complex natural environments, effectively distinguish debris flows from easily confused objects such as flash floods and pure water flows, and intelligently filter out moving interference such as wind-blown vegetation, splashing raindrops, and water surface reflections. It can also obtain the time-varying characteristics of the debris flow flow rate and coverage area of ​​the debris flow event.

[0098] Please see Figure 2 and Figure 3 For real-world images of an upstream area of ​​a river collected at time point 1, the debris flow identification results were obtained based on the YOLOv12n network and the improved YOLOv12n network (i.e., the improved YOLOv12n-CBAM model). The proportion of images identifying wildness debris flow based on the YOLOv12n network was 0.85, while the proportion of images identifying wildness debris flow based on the improved YOLOv12n network was 0.98. This shows that the improved YOLOv12n network of this invention can more accurately identify the real-time situation of wildness debris flow.

[0099] Please see Figure 4 As shown, this invention provides a debris flow monitoring system based on artificial intelligence vision, the system comprising:

[0100] The data collection module is used to collect multi-source debris flow image data;

[0101] The preprocessing module is used to preprocess multi-source debris flow image data, including screening preprocessing, normalization preprocessing, multi-dimensional annotation preprocessing, and enhancement preprocessing.

[0102] The partitioning module is used to divide multi-source debris flow image data into several datasets.

[0103] An improvement module is used to improve the attention mechanism of neural network models;

[0104] The training module is used to train a neural network model using several datasets until the neural network model converges.

[0105] The processing module is used to process the debris flow image sequence under test using the trained neural network model to obtain the debris flow event characteristics.

[0106] The debris flow monitoring system based on artificial intelligence vision of the present invention has the same operation and effect as the debris flow monitoring method based on artificial intelligence vision described above, and the description of the debris flow monitoring system based on artificial intelligence vision will not be repeated here.

[0107] In one embodiment of the present invention, the present invention also provides a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed, implements the debris flow monitoring method based on artificial intelligence vision as described above.

[0108] In one embodiment of the present invention, the present invention also provides a computer device, the computer device including at least a memory and a processor, wherein a computer program is stored in the memory, and the computer program, when executed by the processor, implements the debris flow monitoring method based on artificial intelligence vision as described above.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of a necessary general-purpose hardware platform, or by a combination of hardware and software. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a computer product. The present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Other embodiments may also be used. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A debris flow monitoring method based on artificial intelligence vision, characterized in that, The method includes: Step S1: Collect multi-source debris flow image data; preprocess the multi-source debris flow image data, wherein the preprocessing includes screening preprocessing, normalization preprocessing, multi-dimensional annotation preprocessing, and enhancement preprocessing; Step S2: Divide the multi-source debris flow image data into several datasets; Step S3: Improve the attention mechanism of the neural network model, and train the neural network model using the aforementioned datasets until the neural network model converges; Step S4: Use the trained neural network model to process the debris flow image sequence to obtain debris flow event characteristics.

2. The method according to claim 1, characterized in that, In step S1, multi-source debris flow image data is collected, including: Based on the work logs of several debris flow image sources, the image data generation attributes of each debris flow image source are determined; wherein, the image data generation attributes include the image data generation time; Based on the image data generation attributes, the image data collection operations for each debris flow image source are determined, thereby collecting multi-source debris flow image data; wherein, the multi-source debris flow image data includes images taken by UAVs, images taken by ground monitoring points, satellite remote sensing images, and images taken by experimental cameras.

3. The method according to claim 2, characterized in that, In step S1, the multi-source debris flow image data undergoes screening preprocessing and normalization preprocessing, including: Frame extraction is performed on the multi-source debris flow image data to obtain several image frame data; Each image frame data undergoes quality screening preprocessing and image normalization preprocessing; wherein, the quality screening preprocessing includes screening each image frame data for resolution and distortion; and the image normalization preprocessing includes normalizing each image frame data for brightness and contrast.

4. The method according to claim 3, characterized in that, In step S1, the multi-source debris flow image data undergoes multi-dimensional annotation preprocessing, including: The debris flow subject and debris flow edge are labeled in each image frame of the multi-source debris flow image data; The debris flow main body annotation includes pixel annotation of the debris flow main body in the flowing state in each image frame data; the debris flow edge annotation includes point-by-point polygon annotation of the actual edge of the debris flow in each image frame data.

5. The method according to claim 4, characterized in that, In step S1, the multi-source debris flow image data undergoes enhancement preprocessing, including: Geometric transformation enhancement, color space enhancement, and hybrid enhancement are performed on each image frame data under the multi-source debris flow image data. The geometric transformation enhancement includes rotating, translating, or scaling each image frame data; the color space enhancement includes HSV color space dithering, brightness and contrast adjustment for each image frame data; and the hybrid enhancement includes cropping and stitching at least two image frame data.

6. The method according to claim 5, characterized in that, Step S2 includes: The category and coordinate information of each image frame data under the multi-source debris flow image data are obtained by the multi-dimensional annotation preprocessing, and each image frame data is matched with its corresponding category and coordinate information; The multi-source debris flow image data is divided into training dataset, test dataset, and validation dataset according to the corresponding proportions of all image frames.

7. The method according to claim 6, characterized in that, Step S3 includes: A CBAM attention mechanism module is embedded in the last A2C2f module at the backbone end of the YOLOv12n network; the channel attention mechanism and spatial attention mechanism under the CBAM attention mechanism module are dually optimized; The YOLOv12n network is trained, tested, and validated using the training dataset, test dataset, and validation dataset contained in the aforementioned datasets until the YOLOv12n network converges.

8. The method according to claim 7, characterized in that, Step S4 includes: The debris flow image to be tested is processed by frame segmentation to obtain the debris flow image sequence; The trained neural network model is used to process the image sequence of the debris flow to be tested to obtain the dynamic characteristics of the debris flow event; wherein, the dynamic characteristics include the changes in the debris flow flow rate and the spatial coverage area of ​​the debris flow event over time.

9. A debris flow monitoring system based on artificial intelligence vision, characterized in that, The system includes: The data collection module is used to collect multi-source debris flow image data; The preprocessing module is used to preprocess the multi-source debris flow image data, wherein the preprocessing includes screening preprocessing, normalization preprocessing, multi-dimensional annotation preprocessing, and enhancement preprocessing. The partitioning module is used to divide the multi-source debris flow image data into several datasets. An improvement module is used to improve the attention mechanism of neural network models; The training module is used to train the neural network model using the aforementioned datasets until the neural network model converges. The processing module is used to process the debris flow image sequence under test using the trained neural network model to obtain the debris flow event characteristics.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed, implements the debris flow monitoring method based on artificial intelligence vision as described in any one of claims 1-8.

Images