A multi-view redundancy and round-trip defrosting snow depth observation system and method

Abstract

This invention discloses a snow depth observation system and method with multi-view redundancy and rotating defrosting, including a perception layer and an edge computing unit deployed in the field. The perception layer includes a multi-camera array, environmental sensors, and an intelligent heating bracket. The multi-camera array includes at least one wide-angle camera and at least two high-precision cameras with longer focal lengths, used for panoramic acquisition and sub-centimeter-level focused observation at different scale segments, respectively. The intelligent heating bracket has a built-in heating film to support zoned temperature control. The edge computing unit performs intelligent defrosting, image enhancement, AI recognition, and multi-source observation fusion calculation based on auxiliary source weighting according to the environment and image quality. This invention avoids single-point failures through multi-view hardware redundancy, utilizes intelligent rotating heating to balance low power consumption and anti-frost, and combines frequency domain defogging and a reliability weighting algorithm driven by auxiliary sources to solve the problems of observation interruption and misjudgment caused by frost, blurring, and occlusion in high-altitude and cold environments, achieving all-weather high-precision observation.

Description

Technical Field

[0001] This invention relates to the field of meteorological observation and remote sensing technology, and more specifically, to a snow depth observation system and method with multi-view redundancy and rotating defrosting. Background Technology

[0002] Snow depth, as a key meteorological parameter characterizing snow cover, plays an irreplaceable and fundamental role in weather and climate monitoring systems in ecologically sensitive areas such as plateaus and frigid regions. Traditional manual observation methods, limited by labor costs and environmental accessibility, cannot meet the demands for high-frequency, continuous, and automated monitoring. Therefore, automatic snow depth observation systems based on optical imaging have gradually become the mainstream technological approach. These systems typically rely on a fixedly installed camera aligned with a standard snow depth scale, extracting scale information through image recognition algorithms to achieve non-contact measurement. In early applications, this approach has achieved good results in mid- and low-latitude temperate regions due to its advantages of simple structure, convenient deployment, and controllable cost. Its core logic lies in transforming complex physical measurement problems into computer vision tasks, using image processing technology to achieve remote, real-time data acquisition.

[0003] However, as automatic snow depth monitoring technology expands to extreme high-altitude and frigid environments such as the Qinghai-Tibet Plateau and the Tianshan Mountains, the fundamental premise of "clear and usable images" upon which the aforementioned single-camera imaging architecture relies faces a fundamental challenge. Specifically, in the high-altitude winter conditions characterized by low temperatures, high humidity, and strong winds, frost or ice crystals easily condense on the camera lens surface, causing scattering and refraction distortion of incident light. This results in severely blurred images, a sharp drop in contrast, and even the complete loss of effective texture information. More importantly, this optical degradation is not an occasional interference but a systemic failure mode directly induced by environmental physical conditions. Existing technologies often attempt to alleviate this problem from a single perspective: some employ constant-power continuous heating to maintain lens temperature, which, while suppressing frost, consumes extremely high energy and is unsuitable for unattended sites relying on solar power; others rely on a single high-resolution camera to improve the signal-to-noise ratio, but cannot avoid blind spots caused by occlusion (such as snowfall, bird droppings, or localized snow cover); still others introduce simple image enhancement algorithms, but their robustness is significantly insufficient under combined interference such as coexisting frost and fog, and drastic changes in lighting. Fundamentally, none of the above strategies break through the inherent paradigm of "single-point perception - single-point failure," and their technical approaches are inherently fragile. Once the core imaging unit's performance deteriorates due to environmental interference, the entire observation chain is interrupted.

[0004] A deeper contradiction lies in the fact that simply improving the performance of a single sensor or adding passive protection measures not only fails to balance energy efficiency constraints and reliability requirements, but may also amplify the overall system risk due to over-reliance on a particular technological path. For example, while using a telephoto lens to pursue high precision can improve the ability to identify local scales, it simultaneously reduces the field of view and exacerbates the sensitivity to lens cleanliness; conversely, choosing a wide-angle lens to expand the field of view sacrifices the pixel resolution of distant scales, reducing the feasibility of sub-centimeter-level measurements. This trade-off between performance dimensions is amplified dramatically under the complex and variable micro-meteorological conditions of high-altitude areas, making it difficult for any single camera configuration to maintain stable output under all operating conditions. Furthermore, even with the introduction of a preliminary image quality assessment mechanism, without cross-validation of multi-source information, the system still cannot distinguish between "real snow depth changes" and "image distortion artifacts," making it highly susceptible to misjudgment. Therefore, when dealing with the chain of observation failures caused by low-temperature frost in high-altitude environments, the existing technology system has exposed structural defects such as insufficient sensing redundancy, delayed environmental response, and lack of quantitative basis for data credibility. The fundamental problem lies in the failure to build a multimodal sensing fusion framework with inherent fault tolerance and dynamic adaptability.

[0005] In summary, overcoming the reliability bottlenecks of traditional single-camera snow depth observation systems in extreme cold environments caused by lens frost, image blurring, and obstructed viewpoints, and constructing a technical system that deeply integrates hardware redundancy, environmental adaptive control, intelligent image analysis, and multi-source data collaborative decision-making to achieve continuous snow depth observation in all weather conditions with high confidence and low power consumption, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0006] The present invention provides a snow depth observation system and method with multi-view redundancy and rotating defrosting, in order to solve the problems of decreased reliability and observation interruption caused by lens frost, image blurring and viewpoint obstruction in the existing single-camera snow depth observation system in extreme environments such as plateau and high-altitude cold regions.

[0007] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention provides a snow depth observation system with multi-view redundancy and rotating defrosting, including a sensing layer deployed in the field and an edge computing unit for data processing and control.

[0009] The sensing layer includes a multi-camera array, a comprehensive environmental sensor, an intelligent heating bracket, and a standard snow depth gauge.

[0010] The multi-camera array is mounted on the intelligent heating bracket and includes at least one wide-angle camera and at least two high-precision cameras with different focal lengths. The focal length of the high-precision cameras is greater than that of the wide-angle cameras. All cameras are aligned with the standard snow depth scale. The wide-angle camera is used to acquire a panoramic image of the entire scale, and the high-precision cameras are used to focus on different scale segments of the scale to achieve sub-centimeter resolution observation.

[0011] The integrated environmental sensor includes a temperature and humidity sensor installed at the camera mounting location, used to monitor the lens microenvironment parameters in real time;

[0012] The intelligent heating bracket is equipped with an embedded heating film for independent temperature control of each camera mounting position in the multi-camera array.

[0013] The edge computing unit is electrically connected to the multi-camera array, the integrated environmental sensor, and the intelligent heating bracket, respectively. It is used to execute an intelligent defrosting temperature control strategy based on the lens micro-environment parameters and image quality, and to perform enhancement processing, AI recognition, and multi-source observation fusion calculation based on auxiliary source-driven credibility weighting on the acquired images to output the final snow depth value.

[0014] Furthermore, each camera in the multi-camera array is mounted on the same horizontally arranged mounting rod at different pitch angles; wherein, the pitch angle of the wide-angle camera relative to the horizontal plane is set to a negative angle to cover the entire area of ​​the scale; the at least two high-precision cameras include a first high-precision camera focused on the bottom to middle scale segment of the scale, and a second high-precision camera focused on the middle to top scale segment of the scale, the first high-precision camera having a negative pitch angle relative to the horizontal plane, and the second high-precision camera having a positive pitch angle relative to the horizontal plane, thereby forming a multi-view redundant layout; the integrated environmental sensor also includes at least one of an anemometer, a light sensor, and a visibility sensor; the standard snow depth scale adopts a vertical column structure, the surface is coated with a high-contrast color-coded coating, and geometric patterns including circles, triangles, or squares are set at preset intervals as graphic identifiers for image recognition algorithm positioning; the system identifies the graphic identifiers during the initialization phase, constructs the perspective transformation matrix between the image coordinate system and the scale physical coordinate system, and corrects the distortion parameters of the wide-angle camera.

[0015] Furthermore, the intelligent heating bracket is an alloy structural component with an internally embedded distributed heating film. The heating film is evenly arranged along the length direction, and each group of heating films corresponds to a camera mounting position. The edge computing unit is equipped with a polling heating mechanism, which controls the intelligent heating bracket to heat different cameras sequentially in non-emergency situations. The duration of a single heating session does not exceed 90 seconds, and the interval between two heating sessions is not less than 5 minutes. It also ensures that at any given time, at least one camera is in a non-heated state to maintain continuous observation.

[0016] Furthermore, the edge computing unit is used to execute the intelligent defrosting temperature control strategy, including:

[0017] When the temperature and humidity sensor detects that the ambient temperature is below 2 degrees Celsius and the relative humidity is above 80%, it enters the anti-frost warning mode.

[0018] In anti-frost alert mode, the image sharpness index Q is calculated in real time. The image sharpness index Q is a sharpness index based on gradient magnitude. The square root of the sum of squares of the first-order differences of each pixel in the horizontal and vertical directions is calculated by the Sobel operator, and the statistical mean of a single frame or the statistical value after smoothing multiple frames is taken as the index value.

[0019] When the image clarity index Q is lower than the set threshold of 0.15, the edge computing unit controls the smart heating bracket to start predictive low-power heating, with the power set to 30% of the rated power; if the index Q does not recover within a preset time, it switches to full-power defrosting mode.

[0020] Furthermore, the edge computing unit has a built-in AI image recognition enhancement module to perform dual preprocessing before the image is input into the recognition model:

[0021] Frequency domain dehazing is performed: the image is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients using discrete wavelet transform. The transmittance map is estimated based on the atmospheric scattering model and the distance from the pixel to the scale, and the high-frequency coefficients are corrected. The image is then reconstructed using inverse wavelet transform.

[0022] Perform spatial contrast enhancement: Employ a contrast-limited adaptive histogram equalization algorithm to divide the image into 8×8 local blocks and limit the contrast gain to improve the intensity of local edge response;

[0023] The AI ​​recognition model is built on the MobileNetV2 backbone network and SSD detection head, and outputs a recognition confidence score that includes the scale reading and the corresponding value range between 0 and 1.

[0024] Secondly, the present invention provides a snow depth observation method with multi-view redundancy and rotary defrosting, employing the system described in the first aspect, including:

[0025] Step S1: Acquire images of a standard snow depth scale using the multi-camera array, and obtain lens micro-environment parameters using the integrated environmental sensor;

[0026] Step S2: Based on the lens microenvironment parameters and image quality status, control the intelligent heating bracket to perform intelligent defrosting operation to maintain a clear lens field of view;

[0027] Step S3: Perform frequency domain dehazing and spatial domain contrast enhancement processing on the acquired images, and input the enhanced images into the AI ​​recognition model to obtain the set of ruler readings and their recognition confidence scores for each camera;

[0028] Step S4: Perform a multi-source observation fusion algorithm on the scale reading set, determine the effective observation set through density clustering analysis, and calculate the final snow depth value by combining the confidence weight driven by the auxiliary source status and historical performance.

[0029] Furthermore, the specific logic for performing the intelligent defrosting operation in step S2 includes:

[0030] It monitors ambient temperature and relative humidity in real time, and enters anti-frost alert mode when the preset frost weather conditions are met;

[0031] In alert mode, the image sharpness index Q of consecutive frames is calculated in real time. If the image sharpness index Q is continuously lower than the degradation threshold, it is determined that there is a risk of frosting or that frosting has already occurred.

[0032] Select heating power according to the degree of degradation: start heating at 30% of the rated power for 2 minutes when the index first drops. If the index does not recover, start heating at 100% of the rated power until the index recovers to the safe threshold. During the heating process, if it is in non-full power defrosting mode, use a rotating heating method to heat each camera in turn, with a certain time interval between each heating to prevent hot airflow from interfering with imaging.

[0033] Furthermore, in step S3, when performing frequency domain dehazing, the calculation formula for the transmittance map introduces a distance factor, and uses the current visibility estimated by the environmental sensor and the geometric distance from the image pixel to the scale to perform exponential decay calculation.

[0034] The AI ​​recognition model output in step S3 includes not only ruler scale readings. It also includes identifying confidence scores. The confidence score Used as a preliminary weighting basis for data fusion.

[0035] Furthermore, the multi-source observation fusion algorithm in step S4 specifically includes:

[0036] First, invalid readings with a confidence score below 0.5 are removed;

[0037] If the number of remaining readings meets the minimum clustering requirement, the DBSCAN algorithm is used to perform density clustering on the remaining readings, and a neighborhood radius is set. The minimum sample size is 2.0 cm, and the minimum sample size is 2 (MinPts). The readings are divided into several clusters.

[0038] The cluster with the largest number of samples is selected as the main cluster. If the number of samples in the main cluster is insufficient or all readings are discarded, the current observation is determined to be invalid and the system self-diagnosis process is triggered. The self-diagnosis process includes: starting full-power heating impact defrosting, comparing the current image with a snow-free background reference image to determine whether the lens is blocked by physical foreign objects, and detecting the camera drive response status.

[0039] For each valid reading within the main cluster, its fusion weight is determined by the current recognition confidence score and the historical reliability weight of the camera.

[0040] Furthermore, the historical reliability weight is dynamically updated based on the consistency between the camera's output results over the past 24 hours and the main cluster center, with the update rule being:

[0041] If the absolute difference between the current reading of a camera and the mean of the main cluster is less than the allowable deviation threshold, then the first positive adjustment coefficient is added to the historical reliability weight of the previous moment.

[0042] If the absolute difference between the current reading of a camera and the mean of the main cluster is greater than or equal to the allowable deviation threshold, then the historical reliability weight of the previous moment is reduced by a second positive adjustment coefficient; wherein, the first positive adjustment coefficient is greater than the second positive adjustment coefficient, so as to reflect the gradual trust accumulation mechanism.

[0043] The final snow depth value is calculated as follows: the readings of each camera in the main cluster are multiplied by their corresponding recognition confidence score and historical reliability weight, all multiplication results are summed, and then divided by the sum of the products of the recognition confidence scores and historical reliability weights corresponding to all readings in the main cluster. The final output includes the snow depth value and data quality indicators. The data quality indicators are divided into three levels: "high confidence", "medium confidence" and "low confidence" based on the number of samples in the main cluster, sample dispersion and the mean recognition confidence.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] (1) Hardware redundancy and multi-scale collaboration completely avoid the risk of single-point failure. This invention constructs a multi-camera array containing wide-angle and high-precision cameras, using lenses with different focal lengths to acquire panoramic views and sub-centimeter-level local details respectively. This "multi-view + multi-focal length" redundant architecture not only solves the contradiction between the field of view and resolution of a single lens, but also ensures that when a certain camera is partially blocked or fails, the system can still maintain effective observation through the other cameras, significantly improving the robustness of the system.

[0046] (2) Intelligent rotating heating mechanism to ensure uninterrupted observation link through system-level constraints. This invention innovatively adopts an intelligent heating bracket with independent temperature control in different zones and a rotating heating strategy. On the one hand, by triggering both environmental parameters and the image sharpness index Q, on-demand defrosting is achieved, significantly reducing energy consumption; on the other hand, the rotating strategy constructs system-level constraints, forcing at least one cold reference channel to be retained during any defrosting window. This not only avoids the interference of hot airflow disturbance on imaging, but also logically ensures that the observation link is never interrupted (i.e., "uninterrupted flow") under extreme defrosting conditions, completely solving the problem of observation gaps caused by synchronous heating in traditional schemes.

[0047] (3) Joint enhancement of frequency and spatial domains to overcome the recognition bottleneck under severe weather conditions. In response to the dense fog and low contrast scenes common in high-altitude and cold regions, this invention introduces dual preprocessing of frequency domain defogging and spatial contrast enhancement before AI recognition. By correcting high-frequency coefficients through wavelet transform to remove fog interference, and combining adaptive histogram equalization to improve the edge response of the scale, the AI ​​model can still maintain a high recognition rate under severe weather conditions that are difficult to distinguish with the naked eye.

[0048] (4) Data fusion based on trust accumulation enables high-confidence autonomous decision-making. This invention proposes a data fusion algorithm based on DBSCAN clustering and dynamic updating of historical reliability weights. Through an asymmetric weight adjustment mechanism of "fast reward, slow punishment" (i.e., the first positive adjustment coefficient is greater than the second positive adjustment coefficient), the system can adaptively learn the long-term performance of each camera and automatically reduce the weight of cameras with degraded performance. This fault-tolerant mechanism at the logical level can effectively eliminate false readings caused by reflections, shadows, or temporary interference, ensuring that the output snow depth data has extremely high reliability.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, embodiments of the present invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the overall structure of the snow depth observation system with multi-view redundancy and rotating defrosting according to an embodiment of the present invention.

[0052] Figure 2 This is a schematic block diagram illustrating the installation structure of the intelligent heating bracket and camera according to an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram illustrating the process of AI image recognition enhancement and data fusion performed by the edge computing unit in the control and processing layer of this invention.

[0054] Figure 4 This is a logical block diagram of the DBSCAN clustering and weight update mechanism in the multi-source observation fusion algorithm of this invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0056] like Figure 1 As shown, this invention provides a snow depth observation system and method with multi-view redundancy and rotating defrosting. Its overall architecture consists of a sensing layer, a control and processing layer, and an application layer. Data interaction and command coordination between these layers are achieved through wired or wireless communication links. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0057] The sensing layer is deployed at the field snow depth observation site and includes a multi-camera array, integrated environmental sensors, intelligent heating brackets, and power and communication modules. The multi-camera array is mounted on a horizontally arranged mounting pole, which is fixed directly in front of the standard snow depth scale, ensuring that all cameras' fields of view cover the main area of ​​the scale. The multi-camera array includes one wide-angle camera and two high-precision cameras. The wide-angle camera uses a 4mm lens with a horizontal field of view of no less than 90 degrees, used to acquire panoramic images of the entire scale and the surrounding snow cover. The two high-precision cameras are equipped with fixed-focus lenses of 12mm and 25mm respectively. The former focuses on the bottom to middle scale section, and the latter focuses on the middle to top scale section, together achieving sub-centimeter resolution observation of the entire scale range. In this configuration, the focal lengths of both high-precision cameras are significantly larger than the focal length of the wide-angle camera, and the second high-precision camera (25mm focal length) has a larger focal length than the first high-precision camera (12mm focal length), thus meeting the observation requirements of different fields of view and resolutions. Each camera is installed at a different tilt angle. The tilt angle of the wide-angle camera is set to -15 degrees, the tilt angle of the first high-precision camera is -5 degrees, and the tilt angle of the second high-precision camera is +5 degrees. This forms a multi-view redundant layout, ensuring that at least one camera has an effective field of view under conditions of snow, frost, or partial obstruction.

[0058] The standard snow depth gauge features a vertical column structure, 3 meters high, with a high-contrast color-coded coating. From bottom to top, it consists of alternating red, white, and blue stripes, each 10 centimeters wide. At key graduation positions, every 50 centimeters, graphic identifiers, including geometric patterns such as circles, triangles, and squares, are placed. The colors of these patterns contrast sharply with the background to enhance the image recognition algorithm's ability to identify the graduation positions. The gauge is made of low-temperature resistant engineering plastic, with a surface treated for UV protection and anti-aging, making it suitable for extreme climates such as high-altitude and frigid regions.

[0059] A comprehensive environmental sensor suite is integrated near the camera mounting pole, including a temperature and humidity sensor, an anemometer, and a light sensor. The temperature and humidity sensor is positioned at the camera mounting location (in this embodiment, it is mounted flush against the outer surface of any camera lens to obtain the most accurate microenvironment data), with a sampling frequency of 1 Hz, used to monitor the temperature and relative humidity parameters of the lens's microenvironment in real time. The anemometer is mounted at the top of the pole, with a measurement range of 0 to 60 m / s and an accuracy of ±0.3 m / s. The light sensor faces the sky, with a dynamic range covering 0 to 200,000 lux, used to assess the impact of current lighting conditions on image quality. All environmental sensor data is aggregated to the edge computing unit via an RS485 bus.

[0060] Combination Figure 2As shown, the intelligent heating bracket is an aluminum alloy structural component with an embedded distributed heating film. Its cross-section is a rectangular cavity structure, with three independently temperature-controlled polyimide heating films evenly arranged along its length. Each heating film corresponds to a camera mounting position. The heating film is made of a 0.1 mm thick flexible electrothermal material, covered with a high-temperature resistant insulating layer, with a maximum operating temperature of 80 degrees Celsius and a power density of 1.5 watts per square centimeter. The heating bracket is tightly fitted to the camera housing via thermally conductive silicone, ensuring efficient heat transfer to the lens area. The heating control signal is output from the edge computing unit, driven by a solid-state relay, and then applied to each heating film circuit, supporting both zoned independent temperature control and rotating heating modes.

[0061] The power supply and communication module consists of a monocrystalline silicon solar panel, a lithium iron phosphate energy storage battery, a 4G / 5G wireless communication unit, and a BeiDou short message satellite communication terminal. The solar panel has a rated power of 100 watts, with the installation tilt angle optimized according to the local latitude, ensuring daily power generation meets the system's all-weather operation requirements. The energy storage battery has a capacity of 200 Ah and supports charging and discharging over a wide temperature range of -40 to +60 degrees Celsius. The 4G / 5G communication unit is used for routine data transmission; when the public network signal is interrupted, it automatically switches to the BeiDou terminal to upload critical status information and compressed image summaries in short message format. The entire power supply and communication module is encapsulated in an IP67-rated protective enclosure, placed under a ground support, and connected to the various components of the sensing layer via a waterproof aviation connector.

[0062] The control and processing layer includes edge computing units and cloud servers. The edge computing units are integrated into the field device chassis, employing an industrial-grade ARM Cortex-A72 quad-core processor with a clock speed of 1.8 GHz, equipped with 4GB of LPDDR4 memory and 32GB of eMMC storage, and running a customized Linux operating system. Their core functions include task scheduling, image preprocessing, AI model inference, heating strategy execution, and local data caching. Figure 3 As shown, the task scheduling module triggers the image acquisition process at preset time intervals (default 10 minutes), while simultaneously monitoring the environmental sensor data stream to determine in real time whether to enter the anti-frost alert mode. The image preprocessing module receives the raw RGB image, first performing noise reduction and white balance correction, and then determines whether to start the defrosting program based on the image sharpness index Q. The image sharpness index Q is a sharpness index based on gradient magnitude. It is calculated using the Sobel operator to obtain the square root of the sum of the squares of the first-order differences of each pixel in the horizontal and vertical directions (i.e., gradient magnitude), and then takes the statistical mean of the entire image (or the mean or median across multiple frames to suppress random noise). Its mathematical formula is as follows:

[0063]

[0064] in and These represent the first-order differences of the image in the horizontal and vertical directions, respectively, obtained through convolution using the Sobel operator. and The dimensions are the image height and width. When the ambient temperature is below 2 degrees Celsius and the relative humidity is above 80%, the system enters a defrost warning mode. In this mode, if the Q value of three consecutive frames is below the threshold of 0.15, the image quality is considered degraded, and a predictive low-power heating program is initiated, with the heating power set to 30% of the rated power for 2 minutes. If the Q value does not recover to above 0.2 within a subsequent preset time (set to 5 minutes in this embodiment), the full-power defrost mode is activated, increasing the heating power to 100% and continuing until the Q value recovers to the safe threshold.

[0065] The polling heating mechanism is activated in non-emergency situations, ensuring that at least one camera is in an unheated state at any given time. This unheated path serves as a reference observation channel, maintaining clear and usable observations even when other cameras experience thermal turbulence (thermal disturbance) due to changes in air density in front of their lenses caused by heating. This effectively avoids the risk of overall optical blurring caused by the "hot wind" effect in traditional fully synchronous heating modes. For example, while the wide-angle camera is heating, at least one of the two high-precision cameras remains at room temperature. The heating sequence is executed cyclically according to a preset polling table, with each heating session lasting no more than 90 seconds and an interval of no less than 5 minutes, to reduce overall energy consumption and avoid the impact of thermal interference on imaging.

[0066] Before the image is input into the deep learning model, the AI ​​image recognition enhancement module performs dual processing: frequency domain dehazing and spatial domain contrast enhancement. Frequency domain dehazing uses a three-layer discrete wavelet transform to decompose the image into low-frequency approximation coefficients and high-frequency detail coefficients. The high-frequency portion estimates the transmittance map based on an atmospheric scattering model. Its expression is:

[0067]

[0068] in This is the atmospheric light weighting factor, with a value of 0.95. The attenuation coefficient is estimated from environmental sensor data based on current visibility. The distance from pixel x to the scale is pre-calibrated using the scale's geometric model. The transmittance map is used to correct high-frequency coefficients, and the image is then reconstructed using inverse wavelet transform. Spatial enhancement employs a contrast-limited adaptive histogram equalization algorithm, dividing the image into 8×8 local blocks and performing histogram equalization on each block while limiting the contrast gain to no more than 3.0 to prevent noise amplification. The enhanced image serves as input to a lightweight object detection model.

[0069] This target detection model is built on the MobileNetV2 backbone network and the SSD detection head. The input size is 300×300 pixels, and the output includes scale readings. and the corresponding recognition confidence score The model uses a dataset of 100,000 labeled images during training, covering scenes with different lighting, weather, snow cover, and lens contamination. It employs a joint optimization strategy using cross-entropy loss and smoothing L1 loss. After deployment on an edge computing unit, the model achieves an inference speed of 8 frames per second, meeting real-time requirements. Recognition results are temporarily stored in a local circular buffer in structured message format, awaiting use by the data fusion module.

[0070] The cloud server consists of a high-performance computing cluster deployed in a data center, receiving data streams from multiple observation stations via the MQTT protocol. The server runs multi-source observation fusion and auxiliary source weighting algorithms, maintains a historical database, and periodically (once a week by default) generates model update packages, pushing them to each edge computing unit via OTA. The historical database adopts a time-series database architecture, storing original image snapshots, environmental parameters, camera readings, confidence levels, fusion results, and system status logs, supporting multi-dimensional queries by time, geographical location, device ID, and other dimensions.

[0071] The application layer provides web and mobile data visualization interfaces, an anomaly warning mechanism, and standardized data interfaces. The data visualization interface adopts a responsive design, displaying the current snow depth value in real time. The system includes an overall confidence score (defined as the weighted average confidence score within the main cluster), the operating status of each camera (including heating status, image quality Q value, and recognition success or failure), and environmental sensor data curves. An anomaly warning mechanism has three alarm thresholds: a yellow warning is triggered when the snow depth change rate exceeds 5 cm / hour and lasts for 30 minutes; an orange warning is triggered when the snow depth exceeds 120% of the historical maximum for the same period; and a red fault alarm is triggered when the system determines "data unavailable" three times consecutively or detects a power supply voltage below 11 volts. All alarm events are notified to maintenance personnel via SMS, email, and platform pop-ups.

[0072] The standardized data interface follows the WMO meteorological data exchange standard, supports both JSON and XML formats, and provides a RESTful API for meteorological operational systems to call. Interface fields include station ID, observation timestamp, snow depth value, unit, data quality flag, and a list of cameras participating in the data fusion.

[0073] Furthermore, the fusion algorithm of this invention is executed locally on the edge computing unit to reduce dependence on communication bandwidth and improve response speed. For example... Figure 4 As shown, the algorithm first processes the set of readings output by the three cameras. Perform confidence filtering and remove Invalid readings. If the number of remaining readings is less than 2, the current observation is directly determined to be invalid, marked as "data unavailable", and a graded self-diagnosis process is triggered. This process includes three levels: The first level is 'thermal shock', which forcibly starts the overload mode (i.e., 120% rated power) of the intelligent heating bracket for 60 seconds for suspected stubborn ice, using thermal shock to peel off the ice layer; The second level is 'occlusion discrimination', if the image still does not improve after thermal shock, the system retrieves a pre-stored snow-free background reference image and compares it with the current image for structural similarity (SSIM). If the SSIM value is lower than 0.4 and a large black or white block appears in a local area, it is determined to be physical occlusion (such as bird droppings or snow burial). At this time, heating is stopped to prevent dry burning and an alarm is sent; The third level is 'hard reset', if the camera driver is detected to be unresponsive, the acquisition module is restarted by powering off via GPIO.

[0074] If at least two readings remain, perform DBSCAN density clustering analysis. The clustering parameter is set as follows: neighborhood radius. centimeters, minimum sample size The DBSCAN algorithm divides the readings into several clusters, retaining the main cluster with the largest number of samples as the effective observation set. For each reading within the main cluster... Its fusion weights are determined by the current identification confidence level. Weighted by historical reliability Jointly determined. Initial values ​​of historical reliability weights. After each successful participation in primary cluster fusion, updates will be made according to the following rules:

[0075]

[0076] in The cluster mean The allowable deviation threshold is set to 3.0 cm. The first positive adjustment coefficient (i.e., the reward coefficient) is set to 0.1; The second positive adjustment coefficient (i.e., the penalty coefficient) is set to 0.05. This satisfies the requirement. The condition is that the first positive adjustment coefficient is greater than the second positive adjustment coefficient, thus reflecting a gradual trust accumulation mechanism where rewards are fast and punishments are slow. Weight Upper and lower limits are set, ranging from [0.5, 2.0], to prevent extreme values ​​from affecting the fusion stability.

[0077] Final snow depth Calculated using the weighted average formula:

[0078]

[0079] The calculation results are rounded to one decimal place and the unit is centimeters. A graded data quality label (QF) is added: When the number of valid readings in the main cluster is ≥2, the mean confidence level of all readings is >0.8, and the sample standard deviation is <0.5cm, QF=1 (high confidence) is marked, and the data can be directly used for business messages; when the number of samples in the main cluster is 2, or the sample standard deviation is between 0.5cm and 1.5cm, QF=2 (medium confidence) is marked, and it is recommended to combine it with cross-validation with neighboring sites; in other cases, QF=3 (low confidence) is marked, which is only used as a trend reference and is not included in the historical extreme value statistics.

[0080] In a preferred embodiment of the present invention, the edge computing unit may further integrate an infrared thermal imaging module to assist in determining whether the scale is covered by snow under conditions of complete optical failure (such as dense fog or darkness). The thermal imaging image is not used in the snow depth calculation, but only serves as a binary criterion for "scale visibility" to trigger more aggressive defrosting or cleaning actions.

[0081] As a preferred implementation for high reliability, the system embeds a pressure-type snow depth sensor underground at the bottom of the scale. A degradation-based circuit breaker mechanism is added to the data fusion logic: when the optical observation source (multiple cameras) is blinded by polar night, blizzard (visibility <50 meters), or determined by self-diagnosis to be completely blocked, the fusion algorithm automatically disconnects the optical link and seamlessly switches to the pressure sensor data channel. At this time, the snow depth value output by the system is marked as a backup source inversion, and based on the inherent hysteresis characteristics of the pressure sensor, a moving average filter is performed on the output data to smooth out instantaneous pressure fluctuations caused by windblown snow.

[0082] Specifically, the system software adopts a modular design, with each functional module decoupled through a message queue. The image acquisition module, environmental monitoring module, heating control module, AI recognition module, data fusion module, and communication module all operate independently, exchanging data via shared memory and an event bus. Upon system startup, a self-test procedure is executed, including camera initialization, heating circuit continuity testing, storage media integrity verification, network connectivity detection, and an automatic scale calibration procedure. During the calibration phase, the algorithm identifies geometric identifiers on the scale, uses their preset physical intervals (e.g., 50 cm) as a reference, calculates the perspective transformation matrix between the image pixel coordinate system and the scale's physical coordinate system using the OpenCV vision library, and calculates the distortion coefficients using calibration data captured by a wide-angle camera to perform geometric correction on the image, ensuring the linearity and accuracy of subsequent pixel-to-millimeter conversion. If the self-test fails, a safety mode is entered, where only the device status is uploaded without snow depth observation.

[0083] In one specific embodiment, the system performs a model health check once daily at 00:00, calculating the average deviation between the output of each camera and the center of the main cluster over the past 24 hours. If the average deviation of a camera exceeds 5 cm for three consecutive days, it is marked as a "suspected fault," its historical weight cap is reduced to 1.0, and a device calibration suggestion is sent to the operation and maintenance platform.

[0084] In summary, this invention achieves physical redundancy through a multi-camera heterogeneous array, proactive anti-interference through intelligent heating driven by environmental perception and image quality feedback, enhances AI recognition robustness through joint frequency-spatial domain enhancement, and achieves logical fault tolerance through a DBSCAN fusion algorithm incorporating historical reliability weights. Ultimately, it constructs an anti-interference snow depth observation system capable of long-term stable operation in extreme environments such as high-altitude and frigid regions. This system not only solves the single-point failure problem but also provides highly available and high-precision automated observation methods for meteorological operations through data reliability quantification and self-learning mechanisms.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A snow depth observation system with multi-view redundancy and rotary defrosting, characterized in that, This includes a sensing layer deployed in the field, and edge computing units for data processing and control; The sensing layer includes a multi-camera array, a comprehensive environmental sensor, an intelligent heating bracket, and a standard snow depth gauge. The multi-camera array is mounted on the intelligent heating bracket and includes at least one wide-angle camera and at least two high-precision cameras with different focal lengths. The focal length of the high-precision cameras is greater than that of the wide-angle cameras. All cameras are aligned with the standard snow depth scale. The wide-angle camera is used to acquire a panoramic image of the entire scale, and the high-precision cameras are used to focus on different scale segments of the scale to achieve sub-centimeter resolution observation. The integrated environmental sensor includes a temperature and humidity sensor installed at the camera mounting location, used to monitor the lens microenvironment parameters in real time; The intelligent heating bracket is equipped with an embedded heating film for independent temperature control of each camera mounting position in the multi-camera array. The edge computing unit is electrically connected to the multi-camera array, the integrated environmental sensor and the intelligent heating bracket, respectively. It is used to execute an intelligent defrosting temperature control strategy based on the lens micro-environment parameters and image quality, and to perform enhancement processing, AI recognition and multi-source observation fusion calculation based on auxiliary source driving credibility weighting on the acquired images to output the final snow depth value. The edge computing unit is equipped with a polling heating mechanism, which controls the intelligent heating bracket to heat different cameras in sequence in non-emergency situations, and ensures that at least one camera is in an unheated state at any given time to maintain continuous observation.

2. The snow depth observation system with multi-view redundancy and rotary defrosting as described in claim 1, characterized in that, Each camera in the multi-camera array is mounted on the same horizontally arranged mounting rod at different pitch angles; The wide-angle camera is set to a negative pitch angle relative to the horizontal plane to cover the entire scale area; at least two high-precision cameras include a first high-precision camera focused on the bottom to middle scale segment of the scale, and a second high-precision camera focused on the middle to top scale segment of the scale. The first high-precision camera has a negative pitch angle relative to the horizontal plane, and the second high-precision camera has a positive pitch angle relative to the horizontal plane, thus forming a multi-view redundant layout; the integrated environmental sensor also includes at least one of an anemometer, a light sensor, and a visibility sensor. The standard snow depth scale adopts a vertical column structure, with a high-contrast color-coded coating on the surface. Geometric patterns, including circles, triangles, or squares, are set at preset intervals as graphic identifiers for image recognition algorithms to locate. During the initialization phase, the system identifies the graphic identifiers, constructs the perspective transformation matrix between the image coordinate system and the scale's physical coordinate system, and corrects the distortion parameters of the wide-angle camera.

3. The snow depth observation system with multi-view redundancy and rotary defrosting as described in claim 1, characterized in that, The intelligent heating bracket is an alloy structural component with an internally embedded distributed heating film. The heating film is evenly arranged along the length direction and each group of heating films corresponds to a camera mounting position. The heating time for a single camera does not exceed 90 seconds, and the interval between two heating sessions is not less than 5 minutes.

4. The snow depth observation system with multi-view redundancy and rotary defrosting as described in claim 1, characterized in that, The edge computing unit is used to execute the intelligent defrosting temperature control strategy, including: When the temperature and humidity sensor detects that the ambient temperature is below 2 degrees Celsius and the relative humidity is above 80%, it enters the anti-frost warning mode. In anti-frost alert mode, the image sharpness index Q is calculated in real time. The image sharpness index Q is a sharpness index based on gradient magnitude. The square root of the sum of squares of the first-order differences of each pixel in the horizontal and vertical directions is calculated by the Sobel operator, and the statistical mean of a single frame or the statistical value after smoothing multiple frames is taken as the index value. When the image clarity index Q is lower than the set threshold of 0.15, the edge computing unit controls the smart heating bracket to start predictive low-power heating, with the power set to 30% of the rated power; if the index Q does not recover within a preset time, it switches to full-power defrosting mode.

5. The snow depth observation system with multi-view redundancy and rotary defrosting according to claim 1, characterized in that, The edge computing unit has a built-in AI image recognition enhancement module to perform dual preprocessing before the image is input into the recognition model: Frequency domain dehazing is performed: the image is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients using discrete wavelet transform. The transmittance map is estimated based on the atmospheric scattering model and the distance from the pixel to the scale, and the high-frequency coefficients are corrected. The image is then reconstructed using inverse wavelet transform. Perform spatial contrast enhancement: Employ a contrast-limited adaptive histogram equalization algorithm to divide the image into 8×8 local blocks and limit the contrast gain to improve the intensity of local edge response; The recognition model is built on the MobileNetV2 backbone network and the SSD detection head, and the output includes the scale reading and the corresponding recognition confidence score with a value range between 0 and 1.

6. A snow depth observation method with multi-view redundancy and rotary defrosting, characterized in that, The system as described in any one of claims 1 to 5 comprises: Step S1: Acquire images of a standard snow depth scale using the multi-camera array, and obtain lens micro-environment parameters using the integrated environmental sensor; Step S2: Based on the lens microenvironment parameters and image quality status, control the intelligent heating bracket to perform intelligent defrosting operation to maintain a clear lens field of view; Step S3: Perform frequency domain dehazing and spatial domain contrast enhancement processing on the acquired images, and input the enhanced images into the recognition model to obtain the set of scale readings and recognition confidence scores for each camera; Step S4: Perform a multi-source observation fusion algorithm on the scale reading set, determine the effective observation set through density clustering analysis, and calculate the final snow depth value by combining the confidence weight driven by the auxiliary source status and historical performance.

7. The snow depth observation method with multi-view redundancy and rotary defrosting according to claim 6, characterized in that, The specific logic for performing the intelligent defrosting operation in step S2 includes: It monitors ambient temperature and relative humidity in real time, and enters anti-frost alert mode when the preset frost weather conditions are met; In alert mode, the image sharpness index Q of consecutive frames is calculated in real time. If the image sharpness index Q is continuously lower than the degradation threshold, it is determined that there is a risk of frosting or that frosting has already occurred. Select heating power according to the degree of degradation: start heating at 30% of the rated power for 2 minutes when the index first drops. If the index does not recover, start heating at 100% of the rated power until the index recovers to the safe threshold. During the heating process, if it is in non-full power defrosting mode, use a rotating heating method to heat each camera in turn, with a certain time interval between each heating to prevent hot airflow from interfering with imaging.

8. The snow depth observation method with multi-view redundancy and rotary defrosting according to claim 6, characterized in that, In step S3, the frequency domain dehazing includes: using discrete wavelet transform to decompose the image into low-frequency approximation coefficients and high-frequency detail coefficients, estimating the transmittance map based on the atmospheric scattering model and the distance from the pixel to the scale and correcting the high-frequency coefficients, and then reconstructing the image through inverse wavelet transform. When performing frequency domain dehazing, the calculation formula of the transmittance map introduces a distance factor, and uses the current visibility estimated by the environmental sensor and the geometric distance from the image pixel to the scale to perform exponential decay calculation. The output of the recognition model in step S3 includes not only the ruler scale readings. It also includes identifying confidence scores. The confidence score Used as a preliminary weighting basis for data fusion.

9. The snow depth observation method with multi-view redundancy and rotary defrosting according to claim 6, characterized in that, The multi-source observation fusion algorithm in step S4 specifically includes: First, invalid readings with a confidence score below 0.5 are removed; If the number of remaining readings meets the minimum clustering requirement, the DBSCAN algorithm is used to perform density clustering on the remaining readings, dividing the readings into several clusters; The cluster with the largest number of samples is selected as the main cluster. If the number of samples in the main cluster is insufficient or all readings are discarded, the current observation is determined to be invalid and the system self-diagnosis process is triggered. The self-diagnosis process includes: starting full-power heating impact defrosting, comparing the current image with a snow-free background reference image to determine whether the lens is blocked by physical foreign objects, and detecting the camera drive response status. For each valid reading within the main cluster, its fusion weight is determined by the current recognition confidence score and the historical reliability weight of the camera.

10. The snow depth observation method with multi-view redundancy and rotary defrosting according to claim 9, characterized in that, The historical reliability weight is dynamically updated based on the consistency between the camera's output results and the main cluster center over the past 24 hours, and the update rule is as follows: If the absolute difference between the current reading of a camera and the mean of the main cluster is less than the allowable deviation threshold, then the first positive adjustment coefficient is added to the historical reliability weight of the previous moment. If the absolute difference between the current reading of a camera and the mean of the main cluster is greater than or equal to the allowable deviation threshold, then the historical reliability weight of the previous moment is reduced by a second positive adjustment coefficient; wherein, the first positive adjustment coefficient is greater than the second positive adjustment coefficient, so as to reflect the gradual trust accumulation mechanism. The final snow depth value is calculated as follows: the readings of each camera in the main cluster are multiplied by their corresponding recognition confidence score and historical reliability weight, all multiplication results are summed, and then divided by the sum of the products of the recognition confidence score and historical reliability weight corresponding to all readings in the main cluster. The final output includes the snow depth value and data quality indicators. The data quality indicators are divided into three levels: "high confidence", "medium confidence" and "low confidence" based on the number of samples in the main cluster, sample dispersion and the mean recognition confidence.

Images