Avalanche monitoring self-perfecting system and method based on YOLO

Abstract

The application discloses a YOLO-based avalanche monitoring self-improvement system and method, and belongs to the technical field of intelligent geological disaster monitoring. The method collects multi-source image data through a UAV and satellite remote sensing, and realizes initial avalanche detection through YOLO. A time sequence consistency verification mechanism is used to screen high-confidence pseudo labels, and a dynamic time attenuation weight strategy is combined to construct an enhanced training set. When the pseudo labels accumulate or the performance decreases, model retraining is automatically triggered, performance evaluation is used to decide model updating or rollback, and a self-learning closed loop is formed. The application innovatively deeply integrates YOLO detection, automatic pseudo label distillation, physical parameter inversion and edge end spread deduction, realizes the whole-chain self-improvement of detection-deduction-warning-optimization, continuously improves the detection accuracy, reduces the artificial labeling amount, supports complex terrain and extreme environments such as strong reflection and dynamic blur, and reduces the early warning delay to the millisecond level.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of geological disaster detection, and particularly relates to a snow avalanche monitoring self-improvement system and method based on YOLO. BACKGROUND

[0002] As a common sudden geological disaster in high-altitude mountainous areas, snow avalanche has the characteristics of fast outbreak and strong destructive power, and poses a serious threat to life and property safety, transportation and communication facilities in plateau and mountainous canyon areas. Influenced by the Indian Ocean monsoon, complex terrain and extreme climate, the snow thickness, density and morphology in this area change dramatically seasonally, and there are complex environmental disturbances such as strong ultraviolet light reflection, local shadow, low temperature and low oxygen, so the snow avalanche monitoring needs to meet the core needs of accurate identification, real-time response and long-term self-adaptation.

[0003] The existing snow avalanche monitoring technology based on deep learning mostly uses a fixed model architecture trained once, which has obvious defects: firstly, the model parameters are fixed and lack self-evolution ability, making it difficult to adapt to the dynamic change characteristics of snow in plateau areas, resulting in continuous degradation of detection accuracy over time; secondly, the recognition ability of small-scale initial snow avalanche signs is insufficient, and such weak features are easily covered by environmental noise, so the existing single threshold screening mechanism cannot effectively detect them; thirdly, the real-time performance and stability are insufficient in extreme environments, on the one hand, dynamic blur and device jitter will exacerbate detection errors, and on the other hand, model updating relies on human intervention, which cannot respond to monitoring needs in a timely manner. Therefore, there is an urgent need for a snow avalanche monitoring technology that can automatically iterate and optimize, accurately identify special targets, and adapt to extreme environments to solve the bottleneck of existing technologies. SUMMARY

[0004] In order to solve the problems existing in the prior art, the present application provides a snow avalanche monitoring self-improvement system and method based on YOLO, which significantly improves the accuracy, timeliness and environmental adaptability of snow avalanche monitoring by constructing a complete self-improvement closed-loop monitoring method.

[0005] The technical scheme adopted by the present application is as follows:

[0006] The present application discloses a snow avalanche monitoring self-improvement method based on YOLO, comprising the following steps:

[0007] Step S1: Collecting multi-source image data of the target area through a UAV and a satellite remote sensing system, and constructing an initial training data set;

[0008] Step S2: Training an initial snow avalanche detection model based on the YOLO architecture, and the initial snow avalanche detection model outputs target frame, confidence and snow avalanche category information;

[0009] Step S3: Deploying the initial snow avalanche detection model to an edge device or a cloud, and performing inference on the real-time collected images to generate detection results;

[0010] Step S4: Pseudo-label annotation is performed on high-confidence samples in the detection result, and the pseudo-label needs to meet a time consistency verification condition: the same geographic coordinate is detected in continuous images of not less than 3 periods, and the position deviation is less than 5%, and is confirmed as an effective pseudo-label;

[0011] Step S5: The effective pseudo-label samples and the original annotation samples are mixed according to a dynamic time attenuation weight strategy to form an enhanced training set, and the weight strategy is that the weight of a newly collected sample is α t exponentially attenuates with time t, and α t =e -λt , wherein λ ∈ [0.01, 0.1];

[0012] Step S6: When the number of pseudo-labels exceeds a preset threshold N = 500 or the detection performance decreases by more than Δ = 5%, model retraining is automatically triggered to generate a new version of model V t+1 ;

[0013] Step S7: The performance of the new and old versions of the model is compared on the validation set, and if the average precision value mAP of the new version of the model V t+1 is improved by more than a threshold Δ', the new version of the model is replaced as the main model; otherwise, version rollback is performed, and the old version of the model V t is maintained.

[0014] Step S8: The optimized model is automatically distributed to the monitoring nodes to realize a self-improving closed loop of the model.

[0015] Further, the pseudo-label annotation and screening process specifically includes:

[0016] A double-threshold screening mechanism is set, wherein an initial trace class pseudo-label adopts a low threshold θ low ∈ [0.65, 0.75], and a flowing snow body class pseudo-label adopts a high threshold θ high ∈ [0.85, 0.95]; for the initial trace samples screened by the low threshold, a geometric shape constraint condition is added: the length-width ratio of the boundary box of a microcrack class pseudo-label needs to be greater than 3, and the circularity of the boundary box of a local subsidence class pseudo-label needs to be greater than 0.7; the samples that do not meet the geometric shape constraint or the time consistency verification are down-weighted by a coefficient β = 0.3 in the enhanced training set.

[0017] Further, before the reasoning in step S3, a weak feature enhancement processing of an avalanche is performed on the input image:

[0018] An adaptive contrast stretching algorithm is adopted to map the gray scale interval [μ-σ, μ+σ] to [0, 255] with the mean value μ of the snow background gray scale as a reference, wherein σ is a local standard deviation;

[0019] After the first convolutional layer of YOLO's Backbone layer, an avalanche microcrack morphology-guided attention activation function is inserted. This function performs morphological correlation operations with the feature map through a preset elongated convolutional kernel template, which strengthens the feature response weights of microcracks and local subsidence areas. The output feature map of the attention activation function has its response values ​​in reflective and shadow areas suppressed to less than 30% of the original values.

[0020] Furthermore, during the reasoning process in step S3, the following corrections are performed to address the dynamic ambiguity issue caused by high-speed avalanche flow:

[0021] Synchronously acquire three-axis angular velocity and acceleration information from the UAV IMU data to estimate image motion blur kernel parameters, including the blur direction angle θ. blur With fuzzy length L blur ;

[0022] Based on the blur kernel parameters, Wiener filtering deconvolution operation is performed on the real-time image to obtain a deblurred image;

[0023] Inter-frame difference operation is performed on two consecutive deblurred images. Regions with a pixel change rate exceeding 30% are marked as candidate regions for flowing snow bodies, and their weight is increased by 1.5 times in subsequent detection.

[0024] Furthermore, in the data acquisition step S1, a multimodal physical parameter inversion step is added:

[0025] Snow density ρ was retrieved using hyperspectral data. An empirical relationship model between the ratio of snow reflectance in the 980nm and 1100nm bands and density was established to obtain the density distribution map.

[0026] The snow moisture content w was retrieved using SAR data, and the moisture content distribution map was calculated using a semi-empirical model of C-band backscattering coefficient and moisture content.

[0027] By deploying micro-vibration sensors in the target area, the spectral characteristics of the frictional vibration signal of snow particles are collected, and the dominant frequency amplitude A is extracted. vib As an indicator of micro-vibration intensity;

[0028] The density ρ, moisture content w, and micro-vibration intensity A are used to determine the density ρ, moisture content w, and micro-vibration intensity A. vib The data is concatenated with the RGB data of the image along the channel dimension to form six-channel input data, which is used for model training and inference.

[0029] Furthermore, in the model training phase of step S2, a critical trigger state prediction branch is added:

[0030] A fully connected network branch is added as a bypass to the YOLO Head layer, with the input being the pooling vector of the feature map and the physical parameters ρ, w, and A.vib The concatenated vector;

[0031] The branch outputs the avalanche trigger probability P. trigger ∈[0,1], the training labels are determined based on the critical state moments marked by historical avalanche data;

[0032] During the retraining process in step S6, for density ρ∈[0.38,0.42]g / cm³, slope∈[28°,32°] and micro-vibration intensity A vib For samples approaching the historical trigger threshold, the weight of the loss function is increased to α. t =0.8.

[0033] Furthermore, after the inference in step S3 is completed, a real-time avalanche propagation simulation is performed at the edge device:

[0034] Based on the avalanche initial position coordinates (x0, y0), target box area S, and estimated velocity vector v output by YOLO detection, a simplified Sutherland spread model is invoked.

[0035] The spread model uses terrain slope, aspect, and obstacle distribution as constraints, and calculates the polygonal boundary of the danger zone and the time T for reaching the key terrain point locally on the edge device. arr ;

[0036] When T arr Less than the preset warning time T warn A local audible and visual alarm is triggered after 60 seconds, without waiting for instructions from the cloud.

[0037] Furthermore, after model distribution and optimization in step S8, a self-optimization step for inference accuracy is added:

[0038] Obtain the polygon R representing the actual avalanche spread in subsequent satellite imagery. actual , and the range of the deduction R pred Calculate IoU deviation;

[0039] IoU deviation Terrain parameters and initial detection parameters are used as new samples, labeled with the actual spread range, and included in the pseudo-label library; during the next retraining, the spread estimation error is used as an auxiliary loss term L. spread =||R pred -R actual ||², jointly optimize the parameters of the detection model and the inference model.

[0040] Furthermore, the initial model training in step S2 includes a lightweight pruning step oriented towards avalanche features:

[0041] For each convolutional layer channel of YOLO's Backbone, its correlation with snow spectral and motion features is calculated; the top 60% of channels in the correlation ranking are retained, and channels related to soil and vegetation detection are removed to generate a lightweight model specifically for avalanches; the number of model parameters after pruning is reduced by more than 40%, and the inference latency in a low temperature environment of -30℃ is less than 80ms, which meets the real-time requirements of avalanche early warning.

[0042] Furthermore, before the reasoning in step S3, an image correction step based on the attitude jitter of the UAV is added:

[0043] In the strong wind environment accompanying an avalanche, the pitch and roll rate of change in IMU data are read in real time; when the angular velocity exceeds the threshold ω thr When the angular displacement is 5° / s, the image is compensated by inverse affine transformation based on the angular displacement to eliminate image shift and rotation caused by UAV jitter; the variance of the detection confidence fluctuation of the compensated image is reduced by more than 50%, ensuring detection stability under wind speed >15m / s conditions.

[0044] Furthermore, the present invention also provides a YOLO deep learning-based avalanche monitoring self-improvement system, applied in the aforementioned YOLO deep learning-based avalanche monitoring self-improvement method, comprising:

[0045] The data acquisition module integrates drones, satellite remote sensing systems, and ground sensors to collect multi-source images and physical parameter data;

[0046] The model deployment and inference module includes edge and cloud deployment units, which perform image preprocessing and real-time inference.

[0047] The pseudo-label processing module generates valid pseudo-labels through threshold filtering, time-series verification, and weight allocation.

[0048] The model training optimization module automatically triggers retraining and uses performance evaluation to achieve version iteration or rollback.

[0049] The model distribution module uses an incremental update strategy to push optimized models to each monitoring node; and

[0050] The edge early warning module integrates propagation simulation and local alarm units to form a self-perfecting closed loop.

[0051] The beneficial effects of this invention are as follows:

[0052] (1) This invention achieves accurate identification of the entire avalanche stage, ensuring high confidence in the detection of flowing snow bodies and strengthening the effective capture of initial signs. Through a dual threshold screening mechanism and geometric morphology constraints, exclusive screening criteria are set for initial features such as microcracks and local settlement. Combined with temporal consistency verification to eliminate erroneous samples, the missed detection rate and false judgment rate are effectively reduced, solving the problem of insufficient identification of weak initial signs by traditional technology.

[0053] (2) This invention reduces the cost of manual intervention and constructs an efficient model self-evolution system. By automatically generating effective pseudo-labels and combining them with a dynamic time decay weight strategy, the model can continuously absorb monitoring data in new environments without manual re-labeling and training. At the same time, it automatically triggers retraining and version rollback mechanisms to ensure stable improvement of model performance and significantly reduce the monitoring and maintenance costs in remote plateau areas.

[0054] (3) This invention suppresses environmental interference such as reflection and shadow by enhancing weak avalanche features and using targeted attention mechanism; combined with dynamic blur correction and equipment jitter compensation technology, it ensures detection accuracy under extreme weather conditions; the combination of lightweight model architecture and edge real-time simulation function realizes local rapid processing of monitoring data and early warning response, meeting the emergency avoidance needs of avalanche disasters. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of training on a predetermined avalanche video frame in an embodiment of the present invention. Figure 1 ;

[0056] Figure 2 This is a schematic diagram of training on a predetermined avalanche video frame in an embodiment of the present invention. Figure 2 ;

[0057] Figure 3 This is a schematic diagram of the method in an embodiment of the present invention. Detailed Implementation

[0058] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. Example 1:

[0059] This embodiment discloses a self-improving method for avalanche monitoring based on YOLO deep learning, aiming to solve the technical problems of existing avalanche monitoring systems, such as fixed models, inability to adapt to dynamic changes in snow accumulation in high-altitude areas, target drift, and high manual maintenance costs. The method in this embodiment constructs a self-learning closed loop, enabling the system to automatically iteratively optimize the detection model without human intervention, thereby improving its ability to identify the entire avalanche evolution process.

[0060] Specifically, the method described in this embodiment consists of eight core steps, as referred to Figures 1-3Each step corresponds to a different functional module in the system architecture, and achieves closed-loop data flow:

[0061] Step S1: Multi-source data acquisition and initial dataset construction

[0062] By collaborating with a drone swarm and a satellite remote sensing system, multi-source image data of the target monitoring area is collected. Drones equipped with optical cameras acquire high-resolution RGB images to capture subtle surface deformations during the initial stages of an avalanche.

[0063] The satellite system provides hyperspectral and SAR data, used to analyze the physical properties of snow cover and penetrate shallow clouds and fog, respectively. The collected data is uploaded wirelessly to a cloud data center or ground monitoring station. After geometric correction and radiometric calibration, an initial training dataset containing positive avalanche samples, negative samples, and background samples is constructed. This dataset provides the model with fundamental knowledge representations, covering typical distributions of landforms, illumination, and snow cover conditions.

[0064] Step S2: Training the initial avalanche detection model

[0065] Based on the YOLOv8 or YOLOv9 single-stage object detection architecture, supervised learning is performed on the initial training dataset; the model input is multi-source image data, and the output includes the object bounding box coordinates, confidence score, and avalanche category label.

[0066] During training, the CIoU loss function is used to optimize bounding box localization accuracy, and FocalLoss is used to handle the sample class imbalance problem, ultimately generating initial model weights with basic detection capabilities. This model serves as a baseline version for subsequent self-improvement processes, and its performance is evaluated on the test set, with metrics such as precision and recall recorded as performance baselines.

[0067] Step S3: Real-time inference and detection output

[0068] The trained initial model is deployed to the inference nodes of the monitoring system, including the drone edge computing unit and the cloud server.

[0069] During the inspection process, the drone performs frame-by-frame inference on the real-time acquired image stream to generate detection results; the cloud server then processes the satellite wide-field imagery in batches. The detection results are output in a standardized format, including target category identifiers, normalized bounding box coordinates, confidence scores, and timestamp information. All detection results are stored in a time-series database, providing a data foundation for subsequent pseudo-label screening.

[0070] Step S4: Pseudo-tag generation and temporal consistency verification

[0071] The detection results generated by real-time inference are filtered to generate pseudo-labels for model retraining. A temporal consistency verification mechanism is introduced into the filtering process: the system maintains a geographic coordinate index table to record the detection history of each spatial unit in different acquisition phases.

[0072] If an avalanche target is detected at the same geographic coordinates in multiple consecutive time phases, and the offset distance of the center point of the detection bounding box in each phase is less than a preset proportional threshold, then the detection result is deemed to have spatiotemporal stability and is upgraded to a valid pseudo-label. This mechanism effectively eliminates false detections caused by clouds, shadows, or transient noise, ensuring the reliability of pseudo-labels.

[0073] Step S5: Enhance training set construction and dynamic weight strategy

[0074] The pseudo-labeled samples that passed verification in step S4 are mixed with the original manually labeled samples to form an enhanced training set. To improve the model's adaptability to recent data distributions, a dynamic time decay weight strategy is introduced: the weight of newly collected samples in the loss function decays exponentially with the collection time interval, and the decay coefficient λ is adaptively adjusted according to the data update frequency. The latest samples within the time window receive the highest weight, while the weights of historical samples gradually decrease, thus enabling the model training to focus on the current snow accumulation state and preventing outdated data distributions from dominating model parameter updates.

[0075] Step S6: Retraining Trigger and Model Version Iteration

[0076] The system employs a dual-condition trigger mechanism to automatically initiate model retraining: Condition 1 is that the accumulated number of pseudo-label samples reaches a preset threshold N; Condition 2 is that the online monitoring detection performance index decreases by more than a preset magnitude Δ relative to the historical peak. When either condition is met, the GPU training server is automatically invoked to load the enhanced training set and perform model fine-tuning. After training is complete, a new version of the model V is generated. t+1 Its version number is associated with the time-series index, forming a traceable model evolution chain.

[0077] Step S7: Performance Evaluation and Model Rollback Decision

[0078] New version model V t+1 Performance evaluation was conducted on an independent validation set, calculating key metrics such as precision, recall, and mAP.

[0079] If V t+1 mAP compared to the old version V t If the performance improvement exceeds the performance gain threshold Δ', the version is confirmed to be effective, and a model replacement operation is performed; if performance does not improve or even deteriorates, the version rollback mechanism is triggered, and V is restored. t The main model, and V t+1Versions marked as failures are stored in the repository. This decision-making process ensures that the online model always maintains optimal performance and avoids the risk of performance degradation.

[0080] Step S8: Model Distribution and Distributed Collaborative Update

[0081] The new version of the model, whose performance has been confirmed, is pushed to all monitoring nodes via OTA technology, including the UAV's onboard computing unit and the regional monitoring center. The push process uses an incremental update strategy, transmitting only the weight difference portion to reduce communication bandwidth consumption.

[0082] The cloud-based centralized training and edge-based local inference form a collaborative architecture: the cloud is responsible for the training and optimization of complex models, while the edge is responsible for the real-time inference of lightweight models. The two form a data loop through pseudo-label feedback, enabling the entire monitoring system to improve itself and evolve continuously.

[0083] This embodiment also discloses a system, including:

[0084] The data acquisition module integrates drones, satellite remote sensing systems, and ground sensors to collect multi-source images and physical parameter data;

[0085] The model deployment and inference module includes edge and cloud deployment units, which perform image preprocessing and real-time inference.

[0086] The pseudo-label processing module generates valid pseudo-labels through threshold filtering, time-series verification, and weight allocation.

[0087] The model training optimization module automatically triggers retraining and uses performance evaluation to achieve version iteration or rollback.

[0088] The model distribution module uses an incremental update strategy to push optimized models to each monitoring node; and

[0089] The edge early warning module integrates propagation simulation and local alarm units to form a self-perfecting closed loop.

[0090] The system's various modules work together to achieve a closed-loop operation across the entire chain, from data acquisition to early warning optimization.

[0091] The data acquisition module specifically integrates a cluster of UAVs equipped with optical cameras, hyperspectral sensors, IMUs, and micro-vibration sensors to provide high-resolution images and attitude signals; the satellite remote sensing system outputs SAR data and wide-field images, and the ground sensor network supplements micro-vibration signals and terrain DEM data. All data are corrected and then uploaded to the database.

[0092] In the model deployment and inference module, the edge unit deploys a lightweight pruned YOLO model and integrates preprocessing functions such as adaptive contrast stretching, jitter compensation, and Wiener filtering deblurring to maintain a small inference latency in low-temperature environments; the cloud unit processes satellite wide-field images and forms a collaborative inference architecture with the edge unit.

[0093] The pseudo-label processing module sets differentiated thresholds for initial signs and mobile snow bodies, filters valid samples by verifying the temporal consistency of three consecutive image periods and geometric morphological constraints, and performs weight reduction processing on unqualified samples.

[0094] The model training optimization module is equipped with a GPU cluster. When the number of pseudo-labels accumulates to 500 or the performance drops by more than 5%, retraining is automatically triggered. A new critical trigger state prediction branch and a propagation inference error auxiliary loss term are added. Version iteration or rollback is achieved through mAP metric evaluation.

[0095] The model distribution module uses LoRa emergency communication and incremental update technology, transmitting only the weight difference portion, which is adapted to the network environment in remote mountainous areas and pushes the optimized model to nodes such as drones and ground monitoring stations.

[0096] The edge warning module has a built-in simplified Sutherland spread model that extrapolates the danger zone and arrival time based on the detection results. When the arrival time is less than 60 seconds, a local audible and visual alarm is triggered without waiting for cloud instructions. Example 2:

[0097] This embodiment further optimizes the method of Embodiment 1 by providing a pseudo-label screening optimization scheme for the differences in characteristics of different avalanche evolution stages. It aims to solve the problems of missed detection of initial signs and false detection of flowing snow caused by a single confidence threshold, and improve the adaptability and accuracy of pseudo-labels in different avalanche stages.

[0098] This embodiment, based on the method in Embodiment 1, introduces a hierarchical screening mechanism in the pseudo-label generation stage. This mechanism categorizes detected targets into initial signs and flowing snow bodies based on the differences in their physical characteristics and visual features. For initial signs, which exhibit low contrast and weak boundary features in the image, resulting in relatively low confidence levels in the model output but possessing specific geometric shapes, a lower confidence threshold is used for screening, along with additional geometric shape constraints.

[0099] Specifically, for targets with microcracks, the aspect ratio of the bounding box must meet the condition of being slender greater than 3, and for targets with localized subsidence, the roundness of the bounding box must meet the condition of being nearly circular greater than 0.7. For flowing snow bodies, due to their obvious features and high contrast, a higher confidence threshold is used for direct screening.

[0100] By employing the dual threshold and geometric constraints described above, the system can effectively capture weak signals of initial signs without reducing the accuracy of detecting flowing snow bodies. For samples that do not meet the temporal consistency verification or geometric constraints, their weights in the augmented training set are assigned a weighting coefficient β, with a value ranging from 0.2 to 0.5. This allows them to participate in training without excessively affecting the model parameters, thereby ensuring the quality of pseudo-labels while fully utilizing all available information.

[0101] The method of this embodiment was applied to a typical high mountain canyon monitoring area in a plateau region. This area has a complex environment with mixed distribution of glaciers and snow and severe canyon shadows.

[0102] In the early stages of system operation, the use of a single high threshold of 0.9 resulted in a large number of microcrack signs being missed, and the initial sign samples in the pseudo-label library accounted for less than 10%.

[0103] After applying this embodiment, the detection rate of microcrack-type false tags increased by 3.2 times, the detection rate of local settlement-type tags increased by 2.1 times, and the detection rate of mobile snow bodies remained stable.

[0104] After three iterations, the model's recall rate for initial avalanche signs increased from 58% to 84%, and the false positive rate decreased from 12% to 6.5%, validating the effectiveness of dual thresholds and geometric constraints in complex mountainous environments. Example 3:

[0105] This embodiment further optimizes the method of Embodiment 1 by providing an image preprocessing scheme for low contrast and high noise interference of initial avalanche signs. It aims to solve the problem of microcracks and local subsidence features being masked by strong ultraviolet reflection and canyon shadows in plateau areas, and improve the model's sensitivity to detect weak targets.

[0106] This embodiment performs two-stage feature enhancement on the input image before model inference, as follows:

[0107] The first stage is adaptive contrast stretching. This algorithm statistically analyzes the mean μ and standard deviation σ of the grayscale in the snow background region of the image, and linearly maps the grayscale range [μ-σ,μ+σ] containing the target features to the complete 8-bit grayscale dynamic range [0,255]. This significantly amplifies the grayscale difference between the microcracks and the background snow, while avoiding oversaturation.

[0108] The second stage is morphology-guided attention activation, which inserts a specially designed attention module after the first convolutional layer of YOLO Backbone. This module incorporates two morphological convolutional kernel templates: elongated and circular, corresponding to the typical shapes of microcracks and local subsidence, respectively. When the feature map passes through this module, the morphological templates and the feature map undergo correlation operations to calculate a similarity response map. Regions with response values ​​higher than 1.5 times the average response are weighted and their weights are suppressed. In particular, for bright areas caused by strong reflections and shadow areas caused by canyon terrain, their feature response values ​​are compressed to less than 30% of their original values, thereby suppressing environmental noise interference and highlighting target features.

[0109] The two-stage processing described above dynamically adjusts the representation of the input image and intermediate feature maps without changing the main structure of the model, thereby enhancing the model's ability to detect low-contrast targets while maintaining the stability of target detection under normal lighting conditions.

[0110] At a high-altitude monitoring point in a plateau region, the strong afternoon ultraviolet radiation causes intense reflection on the snow surface, resulting in a generally low confidence level (below 0.5) for traditional methods. After applying this embodiment, the image undergoes adaptive stretching and attention enhancement, improving the peak signal-to-noise ratio (PSNR) of micro-crack features by 8.3 dB and reducing the false detection rate in shadowed areas by 72%. Over 30 days of continuous monitoring, the system's average confidence level for detecting initial avalanche signs increased from 0.61 to 0.78, advancing the effective detection time by an average of 2.1 hours, providing a valuable window for early warning. Validation demonstrates that this method is significantly adaptable to the unique high-altitude, high-light environment of plateau regions. Figure 1 and Figure 2 In the middle, the model is being trained on a known avalanche image, where... Figure 2 The confidence level is 0.85. Example 4:

[0111] This embodiment further optimizes the method of Embodiment 1 and provides a correction scheme for image dynamic blur caused by high-speed avalanche flow. It aims to solve the problems of image trailing and target deformation caused by snow flow velocity of up to tens of meters per second, and improve the detection accuracy and positioning accuracy of the model for fast-moving avalanche bodies.

[0112] This embodiment introduces motion-state-aware deblurring preprocessing before the image is input to the detection model. The system synchronously reads the three-axis angular velocity and linear acceleration data output from the IMU unit on the UAV, and combines this with the UAV's flight velocity vector and the exposure time parameters of the imaging sensor to dynamically estimate the motion blur kernel at the moment of image acquisition. This blur kernel is described by a two-dimensional Gaussian function, with its major axis corresponding to the motion direction angle θ. blur The standard deviation of the major axis corresponds to the fuzzy length L. blurThe standard deviation of the minor axis is set to a fixed value. After obtaining the blur kernel parameters, the Wiener filtering algorithm is used to perform deconvolution operation on the original image to restore the high-frequency details of the image and reduce the motion blur effect.

[0113] To further enhance feature extraction of flowing snow, this embodiment performs inter-frame difference operation on two consecutive frames of images after deblurring to calculate the grayscale change rate at each pixel position.

[0114] Pixel regions with a rate of change exceeding a preset threshold are marked as dynamic candidate regions, corresponding to the leading edge of snow flow. During subsequent detection, the feature map weights of these dynamic candidate regions are increased by a fixed factor, guiding the model to focus on the moving target and compensating for minor artifacts that may be introduced during deblurring. This method combines motion state perception with image restoration, achieving adaptive correction of dynamic blur and enabling online processing without the need for pre-collected blurred samples.

[0115] On a steep mountain slope in a high-altitude region prone to large-scale slab avalanches, the flow velocity after an avalanche can reach 45 m / s. Without applying this embodiment, the snow shadow in the image is 20-30 pixels long, and the YOLO model misclassifies the shadow as background or misses it.

[0116] After applying this embodiment, the IMU data is used to estimate the blur kernel in real time, and the deblurring process improves the sharpness of the snow body edges by 60%, while the inter-frame difference accurately locks the flow front. The system achieved a detection recall rate of 91% in three real avalanche events, and the bounding box localization error decreased from an average of 15 pixels to 6 pixels, verifying the effectiveness of this method for high-speed flowing targets. Example 5:

[0117] This embodiment further optimizes the method of Embodiment 1 by providing a scheme for predicting the critical state of avalanches by integrating multi-source remote sensing data and ground sensor information. It aims to solve the technical bottleneck that relying solely on image features cannot accurately determine whether an avalanche is about to be triggered, and to achieve a quantitative assessment of the physical state before a disaster occurs.

[0118] This embodiment expands the dimensions of data acquisition by introducing three types of physical parameter inversion techniques.

[0119] The first category is snow density inversion, which uses the reflectance ratio of a hyperspectral sensor at the 980nm and 1100nm bands to establish an empirical regression model with snow density, generating a density distribution map with a spatial resolution of meters.

[0120] The second category is snow moisture content inversion. Using the backscattering coefficient of the C-band of the SAR sensor and combined with a semi-empirical water cloud model, the liquid water content of snow is calculated to obtain a moisture content distribution map.

[0121] The third category is micro-vibration signal acquisition, which involves deploying microelectromechanical system vibration sensors in the monitoring area to capture the frictional vibration of snow particles during the stress accumulation stage, and extracting the main frequency amplitude in the vibration spectrum as a micro-vibration intensity index.

[0122] The three types of physical parameters mentioned above are fused with the RGB image at the data layer to form a multi-channel input tensor. During model training, a new physical parameter fusion branch is added. This branch extracts semantic information from the mid-layer feature map of YOLOBackbone, concatenates it with the physical parameter vector, and then inputs it into an independent fully connected layer. The training objective is to predict the avalanche trigger probability. The loss function of this branch is weighted and merged with the original detection loss, enabling the model to understand the relationship between physical parameters and disaster risk while learning the target location. During the inference phase, this branch outputs probability values, providing a quantitative basis for early warning decisions.

[0123] In a typical snow-covered basin in a plateau region, where the snow layer depth reaches 1.5 meters, traditional imaging methods can only detect avalanches that have already occurred. After deploying this embodiment, the density and water content retrieved from hyperspectral and SAR data show that a high-density layer of 0.42 g / cm³ and a layer with 12% water content exist at the bottom of the snow layer, reaching a critical state. Micro-vibration sensors recorded an abnormal increase in the dominant frequency amplitude 3 hours before the avalanche, and the trigger probability of the model's physical parameter branch output reached 0.78.

[0124] The system issued a yellow alert 2.5 hours in advance and subsequently detected a slab avalanche. Compared to detection methods that rely solely on images, this embodiment advanced the warning time window by at least 2 hours, verifying the core value of physical parameter inversion in criticality prediction. Example 6:

[0125] This embodiment further optimizes the method of Embodiment 1 and provides a solution that integrates critical trigger probability prediction function in the YOLO detection framework. It aims to solve the problem that the existing technology can only detect avalanches that have already occurred and cannot quantify and warn of the critical state of disasters, thus achieving a leap from post-event detection to pre-event prediction.

[0126] This embodiment adds a trigger probability prediction branch in addition to the standard YOLO detection head layer. The input to this branch is a concatenated vector of two parts:

[0127] The first part is a 512-dimensional semantic feature vector obtained by global average pooling of the deep feature map output by YOLO Backbone. The second part is a three-dimensional physical parameter vector from Example 5, including density ρ, water content w, and micro-vibration intensity A. vibThe concatenated vector is input into a three-layer fully connected network, each layer containing 256 neurons and a ReLU activation function. The final output is a single sigmoid-activated neuron with a value range of [0,1], representing the avalanche trigger probability P. trigger .

[0128] The training samples are labeled according to the timeline of historical avalanche events: data M hours before the actual avalanche event is marked as the critical state, and data in non-critical periods are marked as the normal state.

[0129] During the retraining phase, a dynamic weighting strategy is implemented for samples whose physical parameters are close to the critical threshold: when the density ρ is within the 95% confidence interval of the historical trigger value, the slope is close to the critical angle, and the micro-vibration intensity A... vib When the noise level exceeds three times the background noise, the weight of this sample in the loss function increases to a higher level (e.g., 0.8), forcing the model to focus on learning the feature patterns of the critical state. This branch shares the backbone parameters with the original detection branch, achieving end-to-end joint training and avoiding error accumulation during multi-stage training.

[0130] On a high-risk slope near a major transportation route in a plateau region, the system was deployed and monitored continuously for 45 days. During this period, physical parameters showed that the snow density gradually increased from 0.35 to 0.41 g / cm³, and the micro-vibration intensity peaked 4 hours before the eruption. The model trigger probability branch output continuously increased from 0.12 to 0.81, and the system automatically triggered a red alert when the probability exceeded 0.75.

[0131] One and a half hours after the warning was issued, a large-scale dry snow avalanche occurred, affecting the road below. Thanks to the timely warning, traffic management authorities closed the road in advance, preventing vehicle damage. Statistics show that this branch achieved an 87% accuracy rate in predicting critical conditions, with a false alarm rate of 13%, significantly better than the random guessing level of traditional methods. Example 7:

[0132] This embodiment further optimizes the method of Embodiment 1 by providing a real-time simulation scheme for integrating an avalanche dynamics model at the edge device, aiming to solve the problem of insufficient early warning timeliness caused by cloud transmission delay, achieve millisecond-level local early warning response, and meet the stringent requirements of early warning speed for the high-speed spread of avalanches.

[0133] This embodiment embeds a simplified avalanche spread simulation submodule into the UAV edge computing unit. This module takes three parameters from the YOLO detection output as input: the initial avalanche location coordinates (x0, y0), the target bounding box area S (used to estimate the initial size), and the flow velocity vector v estimated through inter-frame differencing. The module incorporates a simplified Sutherland spread model, which is based on a digital elevation model (DEM) and considers slope, aspect, surface roughness, and obstacle distribution. It uses the finite difference method to quickly solve the motion equations of snow flow locally.

[0134] The simulation calculations are performed in parallel on the GPU of an edge device, generating polygonal boundaries of the danger zone and the estimated time T to reach key terrain points (such as villages, roads, and bridges). arr The early warning logic is completely localized: when T... arr When the timeout period is less than the preset emergency warning threshold, the edge device directly drives the attached audible and visual alarm to issue a warning signal and initiates a data link broadcast to notify surrounding devices and personnel, without waiting for cloud-based decision-making instructions. This architecture compresses the warning latency from seconds of network transmission to milliseconds of local computation, while also reducing the computational load on the cloud server.

[0135] In a steep slope area traversed by a seasonal herding trail in a high-altitude region, where there is no reliable mobile communication coverage, this embodiment was deployed. The drone cruised at an altitude of 5 km. When a small-scale avalanche was detected, with an initial area of ​​200 m² and an estimated flow velocity of 25 m / s, the edge detection module calculated within 80 ms that the danger zone would cover the herding trail below, with an arrival time of approximately 90 seconds.

[0136] The device immediately triggered a high-volume alarm and LED flashing, and broadcast the warning via LoRa to the terminals worn by the herders. The herders received the warning 70 seconds before the avalanche arrived, successfully driving their livestock away to safety. Post-event analysis showed that the IoU between the projected range and the actual impact range reached 0.72, and the warning time error was less than 15 seconds, verifying the crucial role of edge simulation in communication blind spots. Example 8:

[0137] This embodiment further optimizes the scheme of Embodiment 7 by providing a scheme for closed-loop optimization of the spread projection model using actual post-disaster data. It aims to solve the problem of insufficient projection accuracy caused by the fixed parameters of the initial projection model and its inability to adapt to different terrain and snow conditions, and to achieve continuous evolution of projection capabilities.

[0138] This embodiment adds a feedback learning mechanism to the existing embodiment 7. After the avalanche event, the system uses post-event high-resolution satellite imagery or drone re-examination to manually or automatically delineate the actual avalanche spread polygon R. actual The actual range is compared with the range R predicted by the edge extrapolation module. predSpatial alignment is performed, and the Intersection over Union (IoU) is calculated as a deviation measure. The IoU of the triples is then used. deviation The terrain parameter vector and the initial detection parameter vector serve as new training samples. The terrain parameters include slope, aspect, and roughness; the detection parameters include initial location, scale, and flow velocity; and the sample label is the actual spread range R. actual .

[0139] This sample was included in the pseudo-label library. During the next model retraining, not only was the YOLO detection model optimized, but an auxiliary loss term L was also added. spread The loss term is defined as the L2 norm distance between the predicted spread range and the actual spread range. Through backpropagation, this loss simultaneously adjusts the YOLO detection parameters and the dynamic parameters of the Sutherland model, enabling joint optimization of detection and inference within a unified framework. This closed-loop mechanism allows the inference model to learn error patterns from each real-world event, gradually improving its adaptability to avalanche movement patterns under different terrain conditions.

[0140] The model was validated in three different geomorphic regions of the plateau (canyon type, plateau type, and glacier hybrid type). The initial model had an IoU of 0.58 in the canyon region, which was due to underestimating the increase in flow velocity caused by the canyon funnel effect. After feedback learning from five avalanche events, the system identified that the friction coefficient needed to be reduced by 30% and the flow velocity weight increased by 20% in the canyon region.

[0141] After retraining, the inference IoU for the canyon region improved to 0.78, the IoU for the plateau region increased from 0.65 to 0.82, and the IoU for the glacier-mixed region, considering the ice-snow interface slippage effect, significantly improved from 0.52 to 0.71. Validation shows that the feedback learning mechanism improved the average accuracy of the inference model by 22%, achieving adaptive optimization for different terrain conditions. Example 9:

[0142] This embodiment further optimizes the method of Embodiment 1 by providing a scheme for lightweight compression of the model for avalanche-specific features. It aims to solve the problem of inference latency exceeding the limit caused by computing power limitations on edge devices for complex YOLO models, and ensure that real-time detection performance can still be maintained under low temperature and low power consumption conditions.

[0143] In this embodiment, after model training and before deployment, a channel pruning operation targeting avalanche features is performed. The pruning process does not rely on a general importance metric, but instead calculates the correlation between each convolutional layer channel and the core avalanche features. For each output channel of the backbone layer, its feature map is compared with the snow spectral feature template using cosine similarity calculation, and simultaneously with the motion feature template using structural similarity calculation. The weighted sum of the two is taken as the avalanche correlation score for that channel.

[0144] All channels were sorted by score, and the top 60% of highly relevant channels were retained, while channels highly correlated with non-avalanche features such as soil texture and vegetation color were removed. The pruned model parameters and computational cost were reduced by approximately 40%. To compensate for potential accuracy loss due to pruning, the weights of the remaining channels were fine-tuned, and the model was trained a small number of times using an avalanche augmentation dataset to restore and solidify detection accuracy. The resulting lightweight model achieved an inference latency of less than 80ms on an edge GPU, meeting millisecond-level early warning requirements.

[0145] In tests conducted in a high-altitude region at -30°C during winter, the unpruned YOLOv8 model experienced an inference latency of 135ms on a drone edge device, with GPU throttling and unstable latency due to the low temperature. After pruning using this embodiment, the number of model parameters decreased from 11.1M to 6.7M, the inference latency stabilized at 72ms, and there was no performance degradation after 12 hours of continuous operation at -30°C. Regarding detection accuracy, the pruned and fine-tuned model achieved an mAP of 0.812 on the validation set, a decrease of only 0.015 from the original model's 0.827, which is within an acceptable range. The validation demonstrates that this method effectively addresses the constraints of computing power and power consumption on edge devices while ensuring detection accuracy.

[0146] Example 10:

[0147] This embodiment discloses a solution to eliminate the impact of UAV attitude jitter on image quality in an avalanche-triggered environment accompanied by strong winds. It aims to solve the problems of image blurring and target offset caused by wind speed exceeding a threshold, and ensure the input stability of the detection algorithm.

[0148] This embodiment introduces an IMU-based real-time attitude compensation mechanism in the image acquisition stage. The IMU unit mounted on the UAV outputs triaxial angular velocity and acceleration data at a frequency of 1kHz. At the moment of image exposure, the system records the instantaneous attitude angles (pitch angle, roll angle) and their rate of change. When the angular velocity exceeds a preset jitter threshold, it is determined to be a strong wind interference mode.

[0149] The compensation process employs an inverse affine transformation: based on the angular displacement measured by the IMU, the rotation matrix and translation vector of the image sensor relative to its stable attitude are calculated, constructing an inverse transformation matrix. An affine transformation is performed on the original image to compensate for pixel shifts and rotations caused by jitter frame-by-frame, outputting a geometrically stable image sequence. The compensated image is input into the detection model, significantly reducing the temporal variance of its prediction confidence. This compensation mechanism works in cascade with the deblurring module, first compensating for geometric jitter and then correcting motion blur, forming a two-stage image stabilization pipeline. This ensures that even under extreme conditions with wind speeds exceeding 15 m / s, the detection box localization error remains within an acceptable range.

[0150] Example 11:

[0151] This embodiment, based on the method of Embodiment 1, provides a pseudo-label secondary screening method based on avalanche kinematic feature constraints. It aims to address the misjudgment problem of traditional temporal consistency verification failing to distinguish avalanches from non-target dynamic phenomena such as glacier slippage and residual snow on rocks. By introducing avalanche-specific trajectory morphology and motion direction constraints, the purity of the pseudo-label library is further improved, enhancing the model's ability to learn the dynamic evolution patterns of avalanches.

[0152] Specifically, this embodiment adds a dynamic trajectory rationality judgment layer based on the timing consistency verification described in step S4 of embodiment 1.

[0153] Once the avalanche candidate area of ​​the same geographic unit passes three consecutive time series verifications, the system automatically extracts the bounding box center coordinate sequence (x1,y1), (x2,y2), and (x3,y3) of the target in each image, fits its two-dimensional motion trajectory curve using the least squares method, and calculates the first-order motion direction vector and second-order curvature features of the trajectory.

[0154] The trajectory curvature κ is calculated using the discrete point curvature formula:

[0155] κ=|(x'·y''-y'·x'')| / (x'²+y'²) 3 / 2

[0156] Where x' and y' are first-order difference vectors, and x'' and y'' are second-order difference vectors. The system presets the curvature threshold range for avalanche types as follows:

[0157] For slab avalanches, which slide down as a whole with an approximately straight trajectory, the curvature threshold is set to κ<0.1; for dry snow avalanches, which exhibit diffusion, a slightly larger curvature is allowed, and the threshold is set to κ<0.3.

[0158] If the curvature of the fitted trajectory exceeds 0.5, it is determined that the target motion path does not conform to the physical laws of avalanches and is therefore rejected.

[0159] Simultaneously, the system combines the slope direction vector S of the monitored area with the digital elevation model to calculate the angle θ between the trajectory movement direction and S. Avalanche movement is strictly driven by gravity, and its main movement direction should deviate from the slope direction by less than 30°. If the angle θ > 30°, it indicates that the target movement violates the principle of gravity dominance, and may be lateral slippage of clouds or glaciers driven by lateral winds; such false labels are also removed.

[0160] The pseudo-labels, after being filtered by trajectory constraints, have their confidence weights further increased by 1.2 times in the augmented training set to enhance the model's learning of avalanche dynamic features. This module, together with geometric morphology constraints and temporal consistency verification, forms a three-level screening architecture, progressively improving the quality of pseudo-labels.

[0161] This invention is not limited to the optional embodiments described above, and anyone can derive other various forms of products based on the inspiration of this invention. The specific embodiments described above should not be construed as limiting the scope of protection of this invention; the scope of protection of this invention should be determined by the claims, and the specification can be used to interpret the claims.

Claims

1. A self-completion method for avalanche monitoring based on YOLO, characterized in that, The method comprises the following steps: Step S1: collecting multi-source image data of a target area by a UAV and a satellite remote sensing system, and constructing an initial training data set; Step S2: training an initial avalanche detection model based on a YOLO architecture, wherein the initial avalanche detection model outputs a target frame, a confidence and avalanche category information; Step S3: deploying the initial avalanche detection model to an edge device or a cloud end to perform inference on a real-time collected image, and generating a detection result; Step S4: pseudo-labeling high-confidence samples in the detection results, setting a double-threshold screening mechanism, wherein the initial sign class pseudo-label adopts a low threshold θ low ∈ [0.65, 0.75], and the flowing snow body class pseudo-label adopts a high threshold θ high ∈ [0.85, 0.95]; performing time sequence consistency verification on the pseudo-labels, confirming that a same geographical coordinate is detected in continuous images for not less than 3 times and a position deviation is less than 5% as an effective pseudo-label; for the initial sign samples screened by the low threshold, adding geometric shape constraint conditions: the aspect ratio of the bounding box of the micro crack class pseudo-label needs to be greater than 3, and the circularity of the bounding box of the local subsidence class pseudo-label needs to be greater than 0.7; samples that do not meet the geometric shape constraint or the time sequence consistency verification are subjected to weight reduction processing in the enhanced training set by a coefficient β = 0.3; Step S5: mix the effective pseudo-label samples and the original labeled samples according to a dynamic time decay weight strategy to form an enhanced training set, and the weight strategy is that the weight of the newly collected sample is α t Exponentially decays over time t, α t =e -λt Where λ ∈ [0.01, 0.1] Step S6: When the number of pseudo labels exceeds a preset threshold N=500 or the detection performance decreases by more than Delta=5%, automatically trigger model retraining to generate a new version of the model V t+1 ; Step S7: Compare the performance of the new and old version models on the validation set. If the average precision value mAP of the new version model V t+1 increases by more than the threshold value Δ', replace the main model; otherwise, perform version rollback and maintain the old version model V t ; Step S8: automatically distributing the optimized model to a monitoring node to realize a self-improvement closed loop of the model.

2. The YOLO-based avalanche monitoring self-correcting method of claim 1, wherein, Before the inference in the step S3, a weak avalanche feature enhancement processing is performed on the input image: An adaptive contrast stretching algorithm is adopted to map the gray interval [μ-σ, μ+σ] to [0, 255] based on the mean value μ of the snow background gray value, wherein σ is a local standard deviation; After the first convolutional layer of the Backbone layer of YOLO, an avalanche microcrack morphology guided attention activation function is inserted, which performs a morphology correlation operation between a preset elongated convolution kernel template and a feature map to strengthen the feature response weight of the microcrack and the local subsidence area; the output feature map of the attention activation function is suppressed to less than 30% of the original value in the response value of the reflection area and the shadow area.

3. The YOLO-based avalanche monitoring self-correcting method of claim 1, wherein, In the inference process of the step S3, the following correction is performed for the dynamic blur problem caused by the high-speed flow of the avalanche: Synchronously acquire three-axis angular velocity and acceleration information in the unmanned aerial vehicle IMU data, estimate image motion blur kernel parameters, including blur direction angle θ blur and blur length L blur ; Based on the blur kernel parameters, a Wiener filter inverse convolution operation is performed on the real-time image to obtain a deblurred image; Frame difference operation is performed on the two consecutive deblurred images, the area with a pixel change rate of more than 30% is marked as a candidate area of flowing snow, and the weight of the area is increased by 1.5 times in the subsequent detection.

4. The YOLO-based avalanche monitoring self-correcting method of claim 1, wherein, In the data collection of the step S1, a multi-modal physical parameter inversion step is added: The hyperspectral data is used to invert the snow density ρ, and the density distribution map is obtained by establishing an empirical relationship model of the ratio of the snow reflectivity at 980 nm and 1100 nm wave bands and the density; The SAR data is used to invert the snow water content w, and the water content distribution map is calculated by a semi-empirical model of the C-band backscattering coefficient and the water content; Through the micro-vibration sensor deployed in the target area, the frequency spectrum characteristics of the snow particle friction vibration signal are collected, and the main frequency amplitude A is extracted vib as a micro-vibration intensity index The density p, the water content w, the micro-vibration intensity A vib Spliced with image RGB data in channel dimension, forming six-channel input data for model training and inference; In the model training stage of the step S2, a critical trigger state prediction branch is added: A fully connected network branch is added beside the Head layer of YOLO, and the input is the pooling vector of the feature map and the splicing vector of the physical parameters ρ, w, A vib ​ The full connection network branch outputs an avalanche triggering probability P trigger ∈ [0, 1], and the training label is determined according to a critical state moment annotated according to historical avalanche data. During the retraining process of step S6, for the density ρ ∈ [0.38, 0.42] g / cm3, the slope ∈ [28°, 32°] and the micro-vibration intensity A vib The sample reaching the historical trigger threshold, the loss function weight is raised to a t = 0.

8.

5. The YOLO-based avalanche monitoring self-correcting method of claim 1, wherein, After the inference of the step S3 is completed, the avalanche spread real-time deduction is performed at the edge device end: Based on the avalanche initial position coordinates (x0, y0), the target frame area S and the estimated flow velocity vector v output by the YOLO detection, a simplified Satherland spread model is called; The simplified Sutherland spread model takes the terrain slope, slope direction and obstacle distribution as constraint conditions, and generates the dangerous area polygon boundary and the time T to reach the key terrain point by local calculation on the edge device arr ; When T arr Less than the preset warning time T warn = 60 seconds, trigger local sound and light alarm, without waiting for cloud instructions.

6. The YOLO-based avalanche monitoring self-correcting method of claim 5, wherein, After the model distribution and optimization of the step S8, a deduction accuracy self-optimization step is added: acquiring a polygon R of the actual avalanche spread in the subsequent satellite image actual , and the deduced range R pred computing the IoU deviation; IoU deviation , terrain parameters, initial detection parameters as new samples, whose labels are actual spread range, into the pseudo-label library; in the next retraining, the spread deduction error as an auxiliary loss term L spread =||R pred -R actual ||², jointly optimize the detection model and the deduction model parameters.

7. The YOLO-based avalanche monitoring self-correcting method of claim 1, wherein, The initial model training of the step S2 comprises a lightweight pruning step for the avalanche features: The correlation degree of each convolutional layer channel of the Backbone of YOLO with the snow spectrum features and the motion features is calculated; The top 60% channels in the correlation degree ranking are retained, the channels related to rock-soil body and vegetation detection are removed, and a lightweight model dedicated to the avalanche is generated; the model parameter quantity after pruning is reduced by more than 40%, the inference delay in a low temperature environment of-30℃ is less than 80 ms, and the real-time requirement of the avalanche early warning is met.

8. The YOLO-based avalanche monitoring self-correcting method of claim 1, wherein, Before the reasoning of the step S3, an image correction step based on the posture jitter of the unmanned aerial vehicle is added: In the strong wind environment accompanying the avalanche, the pitch angle and the roll angle change rate in the IMU data are read in real time; When the angular velocity exceeds the threshold value ω thr =5° / s, the image is compensated by reverse affine transformation according to the angular displacement, eliminating the image offset and rotation caused by the unmanned aerial vehicle jitter.

9. A self-completion system for avalanche monitoring based on YOLO deep learning, characterized in that, The YOLO deep learning-based avalanche monitoring self-improvement method in any one of claims 1-8 comprises: A data acquisition module integrates unmanned aerial vehicles, satellite remote sensing systems and ground sensors to collect multi-source image and physical parameter data; A model deployment and reasoning module includes edge and cloud deployment units to perform image preprocessing and real-time reasoning; A pseudo-label processing module generates effective pseudo-labels through threshold screening, time sequence verification and weight allocation; A model training and optimization module automatically triggers retraining and realizes version iteration or rollback through performance evaluation; A model distribution module uses an incremental update strategy to push the optimized model to each monitoring node; and An edge early warning module integrates spread deduction and local alarm units to form a self-improving closed loop.

Images