A method for identifying snow cover range based on air-ground cooperation

Abstract

The application provides a snow cover range identification method based on air-ground cooperation, and relates to the technical field of image recognition. The method comprises the following steps: acquiring a camera geographic domain map, a historical orthographic image of a drone and a digital elevation model of a target area; obtaining a snow index feature map through color space conversion according to the camera geographic domain map; generating a snow likelihood map through a preset mapping relationship according to the snow index feature map; obtaining a high-dimensional dynamic feature map through a preset first encoder according to the camera geographic domain map and the snow index feature map; obtaining a high-dimensional static feature map through a preset second encoder from the historical orthographic image of the drone and the digital elevation model; performing feature interaction fusion and decoding on the high-dimensional dynamic feature map and the high-dimensional static feature map to obtain an initial snow probability map; and performing fusion on the initial snow probability map and the snow likelihood map to obtain a snow probability distribution map corresponding to the target area. The application can improve the accuracy of snow monitoring.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a method for identifying snow cover range based on air-ground collaboration. Background Technology

[0002] In the process of snow disaster monitoring and risk identification, snow cover extent is a core foundational information for downstream assessment tasks such as road traffic risk and infrastructure load-bearing risk. Currently, the mainstream monitoring method includes satellite remote sensing, which, due to its wide coverage capability, has become a common tool for snow cover detection and mapping, and is widely used in both snow cover identification and large-scale monitoring.

[0003] In related technologies, satellite remote sensing monitoring is limited by revisit cycles, cloud cover, and spatial resolution, resulting in a delayed response to dynamic changes in snow cover and making real-time monitoring difficult. This leads to discrepancies between monitoring results and actual snow cover, thus affecting the accuracy of snow cover monitoring. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the accuracy of snow cover monitoring.

[0005] To address the aforementioned problems, this invention provides a method for identifying snow cover range based on air-ground collaboration.

[0006] In a first aspect, the present invention provides a method for identifying snow cover range based on air-ground coordination, comprising: Acquire camera geographic domain maps, historical orthophotos of drones, and digital elevation models of the target area; A snow cover index feature map is obtained by color space conversion based on the camera geographic domain map; A prior snow feature map is generated based on the snow index feature map through mapping transformation; A high-dimensional dynamic feature map is obtained by using a preset first encoder based on the camera geographic domain map and the snow index feature map. A high-dimensional static feature map is obtained by using the historical orthophotos of the UAV and the digital elevation model through a preset second encoder; The high-dimensional dynamic feature map and the high-dimensional static feature map are fused and decoded to obtain the initial snow accumulation probability map; The snow probability distribution map corresponding to the target region is obtained by fusing the initial snow probability map and the prior snow feature map.

[0007] Optionally, obtaining the snow index feature map through color space conversion based on the camera's geographic domain map includes: The color space conversion is performed on the camera geographic domain map to obtain the corresponding color space map; Extract the normalized brightness and normalized saturation of each pixel based on the color space diagram; Based on the normalized brightness and normalized saturation of the pixel, the corresponding snow index is obtained through a preset snow index relationship; The snow index feature map is generated based on the snow index corresponding to all the pixels.

[0008] Optionally, the snow cover index relationship satisfies: ; Wherein, K is the snow index, V is the normalized brightness, S is the normalized saturation, and α and β are preset weighting coefficients.

[0009] Optionally, the snow prior feature map includes a snow likelihood map and a snow confidence map; the step of generating the snow prior feature map from the snow index feature map through mapping transformation includes: The snow likelihood probability is obtained by using the preset mapping relationship based on the snow index corresponding to the pixel. The snow likelihood map is generated based on the snow likelihood probability corresponding to all the pixels. The corresponding snow accumulation confidence level is obtained by using a preset confidence level conversion relationship based on the normalized brightness and normalized saturation of the pixel. A snow credibility map is generated based on the snow credibility corresponding to all the pixels. The preset mapping relationship satisfies: ; Where S is the snow accumulation likelihood probability, K is the snow accumulation index, T is the preset bias threshold, M is the preset scaling factor, and clip(·, 0, 1) is the truncation function. The preset confidence conversion relationship satisfies: ; ; ; ; Where C is the snow accumulation confidence level corresponding to the pixel, I is the pixel value of the pixel corresponding to the camera geographic domain map, and V is the normalized brightness corresponding to the pixel. d V is the preset underexposure threshold. s C is the preset overexposure brightness threshold. e B is the exposure confidence level corresponding to the pixel, and B is the normalized saturation level corresponding to the pixel. fV is the preset low saturation threshold. f C is the preset high brightness threshold. f Let g be the fog confidence level corresponding to the pixel, g be a preset texture gradient threshold, and C be the density of the fog. t is the texture confidence level corresponding to the pixel, otherwise is not the case, and ∧ is the condition.

[0010] Optionally, obtaining a high-dimensional dynamic feature map based on the camera geographic domain map and the snow index feature map using a preset first encoder includes: The camera geographic domain map and the snow index feature map are aligned by projection to obtain the corresponding dynamic data tensor; The dynamic data tensor is input into the first encoder, which outputs the corresponding initial high-dimensional dynamic feature map. The high-dimensional dynamic feature map is obtained by weighted aggregation of multiple initial high-dimensional dynamic feature maps of a preset continuous time series corresponding to the current time.

[0011] Optionally, obtaining a high-dimensional static feature map based on the historical orthophotos of the UAV and the digital elevation model using a preset second encoder includes: The historical orthophotos of the UAV and the digital elevation model are aligned by projection to obtain the corresponding static data tensor; The static data tensor is input into the second encoder to output the corresponding high-dimensional static feature map.

[0012] Optionally, obtaining the initial snow accumulation probability map by feature fusion and decoding of the high-dimensional dynamic feature map and the high-dimensional static feature map includes: The high-dimensional dynamic feature map and the high-dimensional static feature map are interactively fused to obtain a fused feature map; The fused feature map is input into a preset decoder to output the initial snow accumulation probability map.

[0013] Optionally, the step of interactively fusing the high-dimensional dynamic feature map and the high-dimensional static feature map to obtain a fused feature map includes: The high-dimensional dynamic feature map is input into a preset first convolutional block attention module to output an enhanced high-dimensional dynamic feature map. The high-dimensional static feature map is input into a preset second convolutional block attention module to output an enhanced high-dimensional static feature map; The enhanced high-dimensional dynamic feature map and the enhanced high-dimensional static feature map are fused together through the interactive process to obtain the fused feature map; The fused feature map satisfies: ; Among them, F fFor the fused feature map, F A For the enhanced high-dimensional dynamic feature map, F B For the enhanced high-dimensional static feature map, Down(C) is the value of the enhanced high-dimensional dynamic feature map F. A The original channel C is compressed. 1×1 For a 1×1 convolution, σ is the activation function, and ⊙ is the element-wise multiplication.

[0014] Optionally, the snow cover probability distribution map satisfies: ; Among them, P f For the snow cover probability distribution map, P A For the initial snow cover probability map, P S For the snow accumulation likelihood plot, P C Let clip(·, 0, 1) be the snow accumulation confidence graph.

[0015] Optionally, the method also includes Based on the snow cover confidence map, the snow cover likelihood map, and the initial snow cover probability map, an anomaly mask corresponding to the target area is generated through a preset conflict determination relationship. The anomaly mask satisfies: ; Where A is the anomaly mask, P C For the snow accumulation confidence map, P A For the initial snow cover probability map, P S Let η be the snow likelihood plot, τ be the preset confidence threshold, ||(·) be the indicator function, and ∧ be the logical AND operator.

[0016] Optionally, the method further includes: Obtain the prediction uncertainty map output by the second output head of the decoder, wherein the decoder includes a first output head and a second output head in parallel, the first output head is used to output the initial snow accumulation probability map, and the second output head is used to output the prediction uncertainty map; The prediction uncertainty corresponding to each pixel is obtained based on the prediction uncertainty map; When the prediction uncertainty is greater than a preset uncertainty threshold, the corresponding pixel is determined to be an uncertain pixel. The uncertainty area is obtained based on all the uncertain pixels. The corresponding abnormal area is obtained based on the abnormality mask; When the anomaly area and the uncertainty area simultaneously meet all preset UAV retesting conditions, the historical orthophoto of the UAV is updated by the collected real-time orthophoto of the UAV. The UAV retesting conditions include the anomaly area being greater than the uncertainty area, the anomaly area being greater than a preset anomaly area threshold, and the uncertainty area being greater than a preset uncertainty area.

[0017] Secondly, the present invention provides a snow cover range identification device based on air-ground cooperation, comprising: The acquisition module is used to acquire camera geographic domain maps, historical orthophotos of drones, and digital elevation models of the target area. The conversion module is used to obtain a snow index feature map by color space conversion based on the geographic domain map of the camera; The mapping module is used to generate a snow prior feature map based on the snow index feature map through mapping transformation; The first encoding module is used to obtain a high-dimensional dynamic feature map based on the camera geographic domain map and the snow index feature map through a preset first encoder; The second encoding module is used to obtain a high-dimensional static feature map based on the historical orthophotos of the UAV and the digital elevation model through a preset second encoder. The decoding module is used to obtain an initial snow accumulation probability map by feature fusion and decoding of the high-dimensional dynamic feature map and the high-dimensional static feature map; The fusion module is used to fuse the initial snow probability map and the prior snow feature map to obtain the snow probability distribution map corresponding to the target area.

[0018] Thirdly, the present invention provides an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the air-ground cooperative method for identifying snow cover range as described in the first aspect.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the snow cover range identification method based on air-ground cooperation as described in the first aspect.

[0020] The beneficial effects of the air-ground collaborative snow cover area identification method of the present invention are as follows: By acquiring camera geographic domain maps, UAV historical orthophotos, and digital elevation models of the target area, it overcomes the limitation of a single data source from satellite remote sensing. Using the camera geographic domain map as the core data for real-time monitoring, it leverages the high-frequency acquisition characteristics of fixed cameras to solve the response lag problem caused by satellite revisit cycles. Simultaneously, the high-resolution geographic background of the UAV historical orthophotos and the refined terrain data of the digital elevation model compensate for the insufficient spatial resolution of satellite remote sensing and avoid monitoring blind spots caused by cloud cover. Based on the camera geographic domain map… By converting the color space to obtain a snow index feature map, interference factors such as changes in lighting and dynamic occlusion are effectively filtered out, solving the problem of monitoring information distortion that easily occurs in satellite remote sensing due to imaging limitations. This provides a reliable physical prior reference for subsequent snow cover identification. Based on the snow index feature map, a snow prior feature map is generated through a preset mapping relationship. This quantifies the feature information of the snow index feature map into the physical probability that each pixel belongs to snow cover, giving snow cover identification clear physical interpretability. Compared with the snow index identification method of satellite remote sensing, this further improves the accuracy of snow cover judgment and reduces the deviation from the actual snow cover situation. Based on camera... The top geographic domain map and snow index feature map are used to obtain high-dimensional dynamic feature maps through a pre-set first encoder. Real-time visual data is combined with prior physical features and time-series information to extract high-dimensional dynamic features, accurately capturing real-time dynamic changes in snow cover. Compared to the lagging dynamic perception capabilities of satellite remote sensing, this achieves immediate response to snow cover changes while improving the richness of feature representation. Based on historical UAV orthophotos and digital elevation models, a high-dimensional static feature map is obtained through a pre-set second encoder. The geographic background of the UAV orthophotos and the terrain features of the digital elevation model are mined to form high-dimensional static features, providing a stable basis for snow cover identification. The geographical context support compensates for the shortcoming of insufficient spatial resolution in capturing fine geographical features, and helps to judge the probability of snow cover in different regions. The high-dimensional dynamic feature map and the high-dimensional static feature map are fused and decoded to obtain the initial snow cover probability map. The adaptive interactive fusion mechanism combines real-time dynamic features and stable static features, and then the snow cover probability map is reconstructed through multi-scale decoding. This not only preserves the real-time change information of snow cover, but also strengthens the stability of the segmentation boundary by relying on geographical static features. It solves the problem of insufficient monitoring accuracy caused by single data features in satellite remote sensing, and reduces false detections and missed detections caused by oblique occlusion and exposure fluctuations.The snow probability distribution map for the target area is obtained by fusing the initial snow probability map and the prior snow feature map. Following the principle of physical prior reliable error correction, the fusion is completed. The physical prior is used to accurately correct the results of the deep learning model, effectively suppressing identification errors caused by white interference objects and special surface features. Compared with the monitoring results prone to deviations in satellite remote sensing, this significantly improves the accuracy of snow probability distribution judgment, making the final monitoring results more consistent with the actual snow cover situation. This comprehensively solves the shortcomings of satellite remote sensing monitoring in terms of real-time performance, resolution, and anti-interference, and significantly improves the overall accuracy of snow cover range monitoring. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a snow cover area identification method based on air-ground collaboration according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the network model of the snow cover area identification method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a snow cover range identification device based on air-ground coordination according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0023] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0024] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0025] It should be noted that the terms "one" and "more" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0026] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0027] Among related technologies, satellite remote sensing monitoring, with its advantage of wide coverage, is widely used in snow cover detection and mapping. However, it has inherent limitations that are difficult to overcome. Satellite remote sensing monitoring is constrained by factors such as long revisit cycles, cloud cover interference, and insufficient spatial resolution. This results in a delayed response to dynamic changes in snow cover, making it impossible to achieve near real-time monitoring at the minute to hour level, and hindering the timely reflection of rapid changes such as snowfall and snowmelt. Furthermore, the low spatial resolution makes it difficult to accurately depict snow cover details in key areas such as mountain roads, bridges, and the vicinity of small infrastructure. Cloud cover can also cause data loss in localized areas, leading to significant discrepancies between the monitoring results and the actual snow cover status on-site. This fails to accurately reflect changes in snow distribution and thickness, thus significantly reducing the accuracy and reliability of snow cover monitoring results and making it difficult to meet the engineering requirements for data timeliness and accuracy in snow disaster monitoring, road traffic risk assessment, and infrastructure hazard identification.

[0028] To address the problems existing in the aforementioned related technologies, embodiments of the present invention provide a method for identifying snow cover range based on air-ground collaboration.

[0029] like Figure 1 and Figure 2 As shown in the figure, an embodiment of the present invention provides a method for identifying snow cover area based on air-ground cooperation, comprising: S110 acquires camera geographic domain maps, historical orthophotos of the drone, and digital elevation models of the target area.

[0030] Specifically, the camera geographic domain map of the target area is an aligned image obtained by combining real-time frames from a fixed camera with camera calibration parameters and a digital elevation model (DEM), and completing the projection mapping through an offline-generated "pixel → geographic raster" projection lookup table. If the monitored area is approximately planar, the mapping can also be achieved by replacing the lookup table with a homography matrix. This image achieves coordinate unification with the UAV orthophoto, and can transform the oblique view of the camera into an equivalent top view under the same geographic coordinate system, providing a real-time visual observation foundation for subsequent multi-source data fusion. The historical UAV orthophoto is the core static data that characterizes the geographic background of the target area, clearly presenting basic geographic information such as surface texture, road distribution, and vegetation cover in the area. It is a key material for constructing static benchmark features in the air-ground collaborative monitoring system and can provide a stable background reference for snow accumulation identification. The digital elevation model (DEM), as the core terrain data, is not only an important basis for completing geographic projection mapping of camera frames, but can also extract derived terrain information such as slope and aspect, which can help determine the possibility of snow accumulation in different locations within the target area, such as the identification of steep slopes and other areas that are not prone to snow accumulation. It is also a core component for constructing a unified geographic monitoring benchmark and forming a basic background layer library. Acquiring these three types of data can provide multi-dimensional and multi-type basic inputs for subsequent physical prior feature construction, air-ground collaborative deep neural network model training and inference. It realizes the spatial coordinate unification of real-time dynamic observation data and static geographic background data, effectively avoiding the registration error and information silo problem caused by the inconsistency of coordinates in multi-source heterogeneous data. It also allows the model to simultaneously capture the real-time snow cover status of the target area and the stable geographic environment context, providing solid data support for improving the accuracy and robustness of snow cover monitoring.

[0031] S120, obtain the snow index feature map by color space conversion based on the geographic domain map of the camera.

[0032] Specifically, a color space conversion algorithm is used to convert the original RGB color space to the HSV color space, which is more suitable for feature expression. Color components with distinctiveness, such as brightness and chromaticity, are separated. Then, based on the numerical distribution of each color channel and the preset mapping rules, normalization and weighted calculation are performed to generate a snow index feature map that can represent the probability and confidence distribution of the existence of the target area, providing reliable feature input for subsequent target detection, localization or segmentation tasks.

[0033] S130, generate a snow prior feature map by mapping transformation based on the snow index feature map.

[0034] Specifically, after obtaining the snow index feature map, it is processed pixel by pixel according to a preset mapping transformation rule. This mapping relationship combines the physical spectral characteristics of snow in the visible light band, and quantizes the feature values ​​representing snow attributes in the snow index feature map according to a preset nonlinear function. At the same time, contextual information such as illumination, shadow, and local texture is introduced for dynamic adaptation. The weight of snow features is strengthened in high index areas, and the influence of interference information is weakened in low index areas. Finally, a snow prior feature map is generated, which may include a snow likelihood map and a snow confidence map. Transforming the spectral features of the snow index feature map into physically interpretable prior features effectively filters out noise interference from factors such as light fluctuations, white interference objects, and terrain shadows. This allows the generated snow prior feature map to retain the physical basis of the snow index while possessing stronger scene adaptability. It provides clear physical prior guidance for feature extraction by subsequent deep learning models, assisting the model in accurately locating suspected snow areas and suppressing the risk of false detection of non-snow objects. It also provides standardized physical reference for subsequent fusion decisions with model output results, significantly improving the accuracy and robustness of snow identification and making the monitoring results more consistent with the actual snow distribution characteristics.

[0035] S140, a high-dimensional dynamic feature map is obtained by using a preset first encoder based on the camera geographic domain map and the snow index feature map.

[0036] Specifically, the camera geographic domain map and the snow index feature map are combined into a 4-channel data tensor (channels 1 to 3 are RGB images of the camera geographic domain map, and channel 4 is the snow index feature map), which serves as the first encoder, namely the real-time dynamic stream encoder. With EfficientNet as the core input, this encoder does not only extract features from a single frame tensor, but also introduces temporal information through a sliding window. It selects the 4-channel tensor of the most recent T frames at time t to form the input sequence. It captures multi-scale features from local texture to global context by using learnable operators such as convolution and attention. At the same time, it shares parameters in the temporal dimension and performs weighted summation on the features of T frames. Finally, it completes the in-depth mining and representation of the real-time state, short-term changes and physical prior features of snow cover under a fixed camera viewpoint, generating a high-dimensional dynamic feature map. By deeply integrating real-time visual observation data with prior physical snow index features and incorporating temporal information to enhance the stability of feature extraction, the generated high-dimensional dynamic feature map not only fully preserves the real-time texture, lighting conditions, and instantaneous changes of snow in the target area, but also incorporates prior physical snow index guidance. This effectively compensates for the shortcomings of single visual feature extraction, which is susceptible to interference from lighting and occlusion. It provides a dynamic feature foundation with both real-time performance and physical interpretability for the subsequent physical perception interaction fusion module. At the same time, the high-dimensional feature representation capability can accurately capture the subtle texture and spatial distribution changes of snow, helping to improve the accuracy of snow segmentation in subsequent multi-scale fusion and decoding processes, and reducing the risk of false detection and false negatives caused by oblique occlusion and exposure fluctuations.

[0037] S150, a high-dimensional static feature map is obtained by using a preset second encoder based on the historical orthophoto of the UAV and the digital elevation model.

[0038] Specifically, historical orthophotos of UAVs with unified geographic coordinate alignment and digital elevation models (DEMs) are integrated and constructed into a 6-channel data tensor as input to a preset second encoder (historical static stream encoder with EfficientNet as the backbone). Channels 1 to 3 are historical orthophotos of UAVs, which can clearly provide stable geographic background information such as surface texture, road distribution, and vegetation cover in the monitored area. Channels 4 to 6 are terrain data such as elevation, slope, and aspect derived from the digital elevation model, which can help the model determine the physical probability of snow accumulation in different areas. The second encoder uses learnable operators such as convolution and attention to perform deep feature extraction on the 6-channel tensor, accurately mining multi-scale static features from local geographic texture to global terrain context, and finally generating a high-dimensional static feature map. Using the geographic background information of UAV orthophotos and the terrain features of digital elevation models as the core, a second encoder is used to achieve a deep representation of the stable geographic environment features of the monitoring area. The generated high-dimensional static feature map is not affected by real-time environmental factors such as weather and lighting, and can provide a stable geographic context benchmark for the air-ground collaborative deep neural network. It effectively complements the high-dimensional dynamic feature map output by the first encoder. It can not only help the subsequent physical perception interaction fusion module to adaptively adjust the contribution weight of dynamic and static features, but also provide a reliable geographic reference for the model when the real-time imaging conditions are poor (such as heavy fog or occlusion). It effectively suppresses the feature distortion problem caused by lighting fluctuations and dynamic occlusion, strengthens the stability of snow cover segmentation boundaries, and assists the model in judging areas that are not easy to accumulate snow, such as steep slopes. It significantly reduces the probability of false detection and false negative detection of snow cover, and improves the robustness and accuracy of snow cover monitoring in complex scenarios.

[0039] S160, the high-dimensional dynamic feature map and the high-dimensional static feature map are fused and decoded to obtain the initial snow accumulation probability map.

[0040] Specifically, the extracted high-dimensional dynamic feature map and high-dimensional static feature map are input into the physical perception interactive fusion module. First, the two feature paths are enhanced by the Convolutional Block Attention (CBAM) module. Then, spatial gating weights are learned from the two feature paths by using 1×1 convolution and Sigmoid activation function. The other branch feature is adapted and adjusted by cross-modulation. The contribution of the two feature paths is adaptively adjusted according to the real-time imaging quality and physical prior. When the real-time dynamic flow is distorted by fog, occlusion, etc., the static feature weight is increased. When the real-time observation is clear, the proportion of dynamic features is increased. Then, feature fusion is completed at multiple scales and the fused feature pyramid is output. The fused features are then input into the U-Net-style shared decoder. The shallow detail features and deep semantic information of the fusion branch A (real-time dynamic flow branch) are obtained by progressive upsampling combined with skip connections. After multi-scale fusion and decoding reconstruction, the snow segmentation head outputs the initial snow probability map. Each pixel in the map is quantified with a value of 0 or 1 to represent the probability that it belongs to snow. Abandoning the traditional simple fusion method of directly stitching or adding dual-stream features, this method achieves accurate fusion of real-time dynamic snow cover features and stable static geographic features through an adaptive interactive fusion mechanism. This allows the fusion result to be flexibly adapted to the actual imaging scene, preserving real-time snow cover change information captured by a fixed camera while combining stable geographic context constraints provided by UAV orthophotos and DEM. The multi-scale decoding and skip connections of the decoder effectively restore the fragmented boundary features of the snow cover, significantly reducing false positives and false negatives caused by illumination fluctuations, oblique occlusion, and exposure changes. The output probabilistic initial snow cover probability map also provides adjustable and verifiable basic data for subsequent physical constraints and fusion decisions, rather than irreversible hard decisions, significantly improving the accuracy and flexibility of snow cover recognition and laying a good foundation for further result correction.

[0041] S170, the snow probability distribution map corresponding to the target area is obtained by fusing the initial snow probability map and the prior snow feature map.

[0042] Specifically, the initial snow probability map and the snow prior feature map generated based on physical priors are fused together. The fusion process follows the core principle of "correction only when prior is reliable". First, the reliability of each pixel region is judged based on the reliability map corresponding to the snow prior feature map. Consistency verification is performed only in high-reliability regions on the initial snow probability map and the snow prior feature map. A preset threshold is used to determine whether there is a feature conflict between the two. Then, a "weighted fusion / rollback" strategy is adopted for conflict regions to complete the fusion decision. Dynamic weights are set according to the prior reliability. The higher the prior reliability, the greater the weight of the snow prior feature map in the fusion. If a strong conflict and high reliability are found, the result of the region is directly rolled back to the physical prior result of the snow prior feature map. Non-conflict regions retain the result of the initial snow probability map. Finally, a snow probability distribution map corresponding to the target region is generated. This fusion approach deeply combines the advantages of feature extraction with the interpretability of physical priors, effectively avoiding the false detection problems that pure deep learning models are prone to in off-site scenarios and complex imaging conditions. By leveraging the physical rules of the snow cover prior feature map, the initial results are accurately corrected, significantly reducing snow cover recognition errors caused by factors such as white interference, light fluctuations, and dynamic occlusion. At the same time, through differentiated weight settings and conflict judgment logic, false corrections by physical priors in low-confidence areas are avoided. This ensures that the generated snow cover probability distribution map not only closely matches the real-time actual state of snow cover in the target area but also has solid physical feature support, significantly improving the accuracy, robustness, and cross-scene generalization ability of snow cover probability distribution judgment. This provides more reliable data support for subsequent accurate monitoring of snow cover coverage and snow disaster risk assessment.

[0043] In this embodiment, by acquiring camera geographic domain maps, historical UAV orthophotos, and digital elevation models of the target area, the limitation of a single data source from satellite remote sensing is overcome. The camera geographic domain map serves as the core data for real-time monitoring, leveraging the high-frequency acquisition characteristics of fixed cameras to address the response lag caused by satellite revisit cycles. Simultaneously, the high-resolution geographic background of historical UAV orthophotos and the refined terrain data from the digital elevation model compensate for the insufficient spatial resolution of satellite remote sensing and avoid monitoring blind spots caused by cloud cover. A snow index feature map is obtained from the camera geographic domain map through color space conversion. This method effectively filters out interference factors such as changes in lighting and dynamic occlusion, solving the problem of distorted monitoring information that easily occurs in satellite remote sensing due to limitations in imaging conditions. It provides a reliable physical prior reference for subsequent snow cover identification. Based on the snow index feature map, a snow prior feature map is generated through a preset mapping relationship. The feature information of the snow index feature map is quantified and converted into the physical probability that each pixel belongs to snow cover, giving snow cover identification clear physical interpretability. Compared with the snow index identification method of satellite remote sensing, this method further improves the accuracy of snow cover judgment and reduces the deviation from the actual snow cover situation. Based on the camera's geographic domain map and snow index features… The snow cover identification process generates a high-dimensional dynamic feature map through a pre-set first encoder. This map combines real-time visual data with prior physical features and incorporates temporal information to extract high-dimensional dynamic features, accurately capturing real-time dynamic changes in snow cover. Compared to the lagging dynamic perception capabilities of satellite remote sensing, this method achieves timely response to snow cover changes while enhancing the richness of feature representation. Furthermore, a high-dimensional static feature map is generated using a pre-set second encoder based on historical UAV orthophotos and digital elevation models. This map extracts the geographical background of the UAV orthophotos and the terrain features of the digital elevation model to form high-dimensional static features, providing a stable geographical context for snow cover identification. This technology supports and compensates for the shortcomings of insufficient spatial resolution in capturing fine geographic features, and helps to determine the probability of snow cover in different areas. It obtains an initial snow cover probability map by fusing and decoding high-dimensional dynamic feature maps and high-dimensional static feature maps. Through an adaptive interactive fusion mechanism, it combines real-time dynamic features and stable static features, and then reconstructs and outputs the snow cover probability map through multi-scale decoding. This not only preserves the real-time change information of snow cover, but also strengthens the stability of segmentation boundaries by relying on geographic static features. It solves the problem of insufficient monitoring accuracy caused by single data features in satellite remote sensing, and reduces false detections and missed detections caused by oblique occlusion and exposure fluctuations.The snow probability distribution map corresponding to the target area is obtained by fusing the initial snow probability map and the snow prior feature map. The fusion is performed following the principle that error correction is only performed if the physical prior is reliable. This physical prior is used to accurately correct the results of the deep learning model, effectively suppressing identification errors caused by white interference objects and special surface features. Compared with the monitoring results prone to deviations in satellite remote sensing, this significantly improves the accuracy of snow probability distribution judgment, making the final monitoring results more consistent with the actual snow cover situation. It comprehensively solves the shortcomings of satellite remote sensing monitoring in terms of real-time performance, resolution, and anti-interference, and significantly improves the overall accuracy of snow cover range monitoring.

[0044] Optionally, obtaining the snow index feature map through color space conversion based on the camera's geographic domain map includes: The color space conversion is performed on the camera geographic domain map to obtain the corresponding color space map; Extract the normalized brightness and normalized saturation of each pixel based on the color space diagram; Based on the normalized brightness and normalized saturation of the pixel, the corresponding snow index is obtained through a preset snow index relationship; The snow index feature map is generated based on the snow index corresponding to all the pixels.

[0045] Optionally, the snow cover index relationship satisfies: ; Wherein, K is the snow index, V is the normalized brightness, S is the normalized saturation, and α and β are preset weighting coefficients.

[0046] Specifically, the color space of the camera geographic map, after geographic coordinate alignment, is first converted from the conventional RGB color space to the HSV color space, which is closer to human visual perception and can accurately separate color and brightness information, resulting in the corresponding HSV color space map. Then, two core feature values, normalized brightness and normalized saturation, are extracted pixel-by-pixel from this color space map. These two feature values ​​precisely correspond to the inherent physical characteristics of snow in the visible light band: high brightness and low saturation. Subsequently, based on a preset snow index relationship, the normalized brightness and normalized saturation of each pixel are substituted into the calculation to obtain the snow index corresponding to each pixel. This preset snow index relationship is dynamically adapted in conjunction with environmental reference anchor points, adjusting the weight ratio of brightness and saturation according to actual lighting conditions. The environmental reference anchor points are the global average brightness and global average saturation of the current camera geographic map. First, the average brightness value of the entire image is calculated, and the illumination is divided into three intervals: underexposed, normal, and overexposed. When the average brightness is in the normal interval, a baseline weight of α=1 and β=0.3 is used. When the average brightness is below the underexposed threshold and the image is in a low-light environment, the brightness weight is reduced and the saturation weight is increased, for example, by setting α=0.7 and β=0.5, to improve the ability of saturation to distinguish snow-covered areas. When the average brightness is above the overexposed threshold and the image is in a strong-light environment, the brightness weight is increased and the saturation weight is decreased, for example, by setting α=1.2 and β=0.2, to strengthen the high-brightness characteristics of snow and suppress misjudgment of overexposed areas. This allows for dynamic adaptation of the weight ratio of brightness and saturation according to the actual illumination conditions, improving the accuracy and robustness of the snow index under complex illumination. Finally, the snow index of all pixels in the target area is integrated to generate a snow index feature map that can quantify the probability of snow accumulation in each pixel.

[0047] In this optional embodiment, the physical spectral characteristics of snow are used as the core basis. The brightness, saturation and hue features are effectively separated through color space conversion, avoiding the problem of mutual interference of color information in RGB space. The accurately extracted normalized brightness and saturation provide a reliable feature basis for snow index calculation. The pixel-by-pixel calculation and global integration based on the preset snow index relationship ensure that the generated snow index feature map retains the real-time spatial information of the camera geographic domain map and gives snow recognition clear physical interpretability. It effectively filters the interference of factors such as drastic changes in lighting, white interference objects, and dynamic occlusion on snow recognition, greatly reducing the risk of misjudgment caused by relying solely on visual features. At the same time, it provides accurate and reliable feature support for the subsequent generation of snow likelihood maps and physical prior constraints of deep learning models, improving the robustness and accuracy of the entire snow monitoring process in complex winter scenarios.

[0048] Optionally, the step of generating a snow prior feature map based on the snow index feature map through mapping transformation includes: The snow likelihood probability is obtained by using the preset mapping relationship based on the snow index corresponding to the pixel. The snow likelihood map is generated based on the snow likelihood probability corresponding to all the pixels. The corresponding snow accumulation confidence level is obtained by using a preset confidence level conversion relationship based on the normalized brightness and normalized saturation of the pixel. A snow credibility map is generated based on the snow credibility corresponding to all the pixels. The preset mapping relationship satisfies: ; Where S is the snow accumulation likelihood probability, K is the snow accumulation index, T is the preset bias threshold, M is the preset scaling factor, and clip(·, 0, 1) is the truncation function. The preset confidence conversion relationship satisfies: ; ; ; ; Where C is the snow accumulation confidence level corresponding to the pixel, I is the pixel value of the pixel corresponding to the camera geographic domain map, and V is the normalized brightness corresponding to the pixel. d V is the preset underexposure threshold. s C is the preset overexposure brightness threshold. e B is the exposure confidence level corresponding to the pixel, and B is the normalized saturation level corresponding to the pixel. f V is the preset low saturation threshold. f C is the preset high brightness threshold. f Let g be the fog confidence level corresponding to the pixel, g be a preset texture gradient threshold, and C be the density of the fog. t is the texture confidence level corresponding to the pixel, otherwise is not the case, and ∧ is the condition.

[0049] It's important to note that the truncation function is a numerical constraint function in the snow likelihood probability calculation. Its core function is to force the intermediate results calculated by the snow index, bias threshold, and scaling factor to be limited to the range of 0 to 1. Values ​​less than 0 are truncated to 0, values ​​greater than 1 are truncated to 1, and values ​​between 0 and 1 remain unchanged. This function ensures that the snow likelihood probability always conforms to the inherent range of probability values, avoiding invalid values ​​due to environmental parameters or calculation biases. This guarantees the rationality and validity of pixel-level probability values ​​in the snow likelihood map, providing standardized physical prior data for subsequent snow probability fusion.

[0050] Specifically, after obtaining the snow index corresponding to each pixel, the snow index is first calculated pixel by pixel according to a preset mapping relationship. This mapping relationship combines the physical spectral characteristics of snow in the visible light band to transform the snow index feature value into a snow likelihood probability in the 0-1 range. At the same time, contextual information such as illumination and local texture is introduced to dynamically adjust the weights. Then, the snow likelihood probabilities of all pixels are integrated to generate a snow likelihood map, accurately quantifying the physical probability that each pixel belongs to snow. Subsequently, based on the normalized brightness and normalized saturation of the pixel, the corresponding snow confidence pair is calculated through a preset confidence conversion relationship. This relationship comprehensively considers factors such as the local stability of the snow index, the deviation from the ideal snow index, illumination fluctuations, and occlusion interference to obtain the snow confidence corresponding to each pixel. Then, the snow confidence of all pixels is integrated to generate a snow confidence map, accurately quantifying the reliability of the snow prior.

[0051] In this optional embodiment, the spectral features of the snow cover index are transformed into physically interpretable snow likelihood probabilities and confidence levels. This provides a clear prior probability reference for snow cover identification through the snow likelihood map, and delineates the reliable boundaries of prior information through the snow confidence map. This effectively filters out noise interference from factors such as illumination fluctuations, white interfering objects, and terrain shadows, providing a standardized physical reference for subsequent fusion decisions with the deep learning model output. This allows for accurate correction of model results in high-confidence regions while avoiding erroneous corrections in low-confidence regions, significantly reducing the risk of false positives and false negatives in snow cover identification and significantly improving the accuracy, robustness, and interpretability of snow cover monitoring. Optionally, such as Figure 2 As shown, the step of obtaining a high-dimensional dynamic feature map based on the camera geographic domain map and the snow index feature map through a preset first encoder includes: The camera geographic domain map and the snow index feature map are aligned by projection to obtain the corresponding dynamic data tensor; The dynamic data tensor is input into the first encoder, which outputs the corresponding initial high-dimensional dynamic feature map. The high-dimensional dynamic feature map is obtained by weighted aggregation of multiple initial high-dimensional dynamic feature maps of a preset continuous time series corresponding to the current time.

[0052] Specifically, firstly, the camera geographic domain map with aligned geographic coordinates and the snow index feature map carrying physical priors are projected and aligned, allowing the two types of data to achieve spatial matching under the same geographic reference system. This integrates them to form a structured dynamic data tensor, ensuring spatial consistency and dimensional uniformity of the data input. Subsequently, this dynamic data tensor is input into the first encoder with EfficientNet as its backbone. By leveraging learnable operators such as convolution and attention in the encoder, multi-scale features in the data are deeply mined, and an initial high-dimensional dynamic feature map that can represent the real-time state of snow cover in a single frame is output. Finally, for the current moment, multiple frames of the initial high-dimensional dynamic feature maps within a preset continuous time series are selected, and the temporal features are fused through weighted aggregation to generate the final high-dimensional dynamic feature map.

[0053] In this optional embodiment, the spatial deviation between the camera geographic domain map and the snow index feature map is eliminated by projection alignment, laying a unified data foundation for feature extraction. The deep feature extraction capability of the first encoder realizes the transformation from basic visual and physical prior features to high-dimensional dynamic features, accurately capturing the real-time texture, lighting conditions and physical characteristics of snow in a single frame. Furthermore, the weighted aggregation based on continuous time series incorporates short-term temporal change information of snow cover, effectively suppressing feature distortion caused by single-frame imaging interference factors such as lighting fluctuations, snow cover, and instantaneous noise. This allows the generated high-dimensional dynamic feature map to retain the real-time dynamic change features and physical prior guidance of snow, while also possessing stronger robustness. It can more comprehensively and stably depict the real-time state of snow under a fixed camera viewpoint, providing a high-quality and highly reliable dynamic feature foundation for subsequent adaptive fusion with high-dimensional static feature maps, and significantly reducing the risk of misjudgment in single-frame feature extraction.

[0054] Optionally, such as Figure 2 As shown, the step of obtaining a high-dimensional static feature map based on the historical orthophotos of the UAV and the digital elevation model through a preset second encoder includes: The historical orthophotos of the UAV and the digital elevation model are aligned by projection to obtain the corresponding static data tensor; The static data tensor is input into the second encoder to output the corresponding high-dimensional static feature map.

[0055] Specifically, historical UAV orthophotos and digital elevation models are projected and aligned, mapping both types of data to a pre-defined geographic raster coordinate system. This eliminates spatial location biases and integrates them into a structured static data tensor. The historical UAV orthophotos provide geographic background information such as surface texture and road vegetation, while the digital elevation model provides topographic data such as elevation, slope, and aspect, achieving spatial and dimensional unification of multi-source static data. Subsequently, this static data tensor is input into a second encoder based on EfficientNet. The encoder's learnable operators, such as convolution and attention, are used to deeply mine multi-scale static features in the data, completing the transformation from basic geographic and topographic features to high-dimensional features, and finally outputting a high-dimensional static feature map.

[0056] In this optional embodiment, projection alignment lays a spatially consistent unified data foundation for feature extraction, effectively avoiding registration errors and information silos caused by inconsistent coordinates. The deep feature extraction capability of the second encoder can accurately capture the stable geographical background and terrain patterns of the monitoring area, allowing the generated high-dimensional static feature map to have rich and refined geographical context information. Moreover, this feature map is not affected by real-time environmental factors such as lighting, weather, and occlusion, and can provide stable geographical benchmark support for subsequent fusion with high-dimensional dynamic feature maps. This helps the model judge the possibility of snow accumulation in different terrains, strengthens the stability of snow accumulation segmentation boundaries, and provides a reliable reference for the model when real-time imaging conditions are poor. It significantly reduces the risk of false detection and missed detection of snow accumulation caused by distortion of dynamic flow information, and improves the robustness and accuracy of the entire snow accumulation monitoring model in complex scenarios.

[0057] Optionally, such as Figure 2 As shown, the step of obtaining the initial snow accumulation probability map by feature fusion and decoding of the high-dimensional dynamic feature map and the high-dimensional static feature map includes: The high-dimensional dynamic feature map and the high-dimensional static feature map are interactively fused to obtain a fused feature map; The fused feature map is input into a preset decoder to output the initial snow accumulation probability map.

[0058] Specifically, a high-dimensional dynamic feature map depicting the real-time state of snow cover and a high-dimensional static feature map providing stable geographic context are input into a physical perception interactive fusion module. First, the two feature maps are enhanced separately through a convolutional block attention module (CBAM). Then, spatial gating weights are learned from the two feature maps and the contribution of the two feature maps is adaptively adjusted through cross-modulation. When real-time imaging is affected by fog, occlusion, etc., the weight of static features is increased, and when real-time observation is clear, the proportion of dynamic features is increased, so as to achieve accurate interactive fusion of the two and obtain a fused feature map. Subsequently, the fused feature map is input into a preset U-Net-style shared decoder. Through multi-scale upsampling and combined with skip connections to fuse shallow detail features and deep semantic information, the feature decoding and reconstruction are completed. Finally, the snow segmentation head outputs an initial snow probability map.

[0059] In this optional embodiment, the traditional fusion method of simply splicing or adding dual-stream features is abandoned. Instead, an adaptive interactive fusion mechanism allows the two features to complement each other. This preserves the real-time dynamic changes in snow captured by the fixed camera and combines the stable geographical and terrain constraints provided by UAV orthophotos and digital elevation models, effectively avoiding the information deficiencies of a single feature dimension. The multi-scale decoding and skip connection design of the decoder can accurately restore the fine boundary features of snow, enhance the stability of snow segmentation boundaries, and significantly reduce false detections and missed detections caused by factors such as illumination fluctuations, oblique occlusion, and exposure changes. The output probabilistic initial snow probability map is not an irreversible hard decision, but provides verifiable and adjustable basic data for subsequent fusion decisions combined with physical priors. This significantly improves the accuracy and flexibility of snow recognition, making the initial snow probability map more consistent with the actual snow cover status of the target area.

[0060] Optionally, such as Figure 2 As shown, the step of interactively fusing the high-dimensional dynamic feature map and the high-dimensional static feature map to obtain a fused feature map includes: The high-dimensional dynamic feature map is input into a preset first convolutional block attention module to output an enhanced high-dimensional dynamic feature map. The high-dimensional static feature map is input into a preset second convolutional block attention module to output an enhanced high-dimensional static feature map; The enhanced high-dimensional dynamic feature map and the enhanced high-dimensional static feature map are fused together through the interactive process to obtain the fused feature map; The fused feature map satisfies: ; Among them, F f For the fused feature map, F A For the enhanced high-dimensional dynamic feature map, F BFor the enhanced high-dimensional static feature map, Down(C) is the value of the enhanced high-dimensional dynamic feature map F. A The original channel C is compressed. 1×1 For a 1×1 convolution, σ is the activation function, and ⊙ is the element-wise multiplication.

[0061] Specifically, the high-dimensional dynamic feature map depicting the real-time dynamic state of snow cover is processed through a temporal attention mechanism and then input into a pre-defined first convolutional block attention module. This module focuses on feature extraction and enhancement, outputting an enhanced high-dimensional dynamic feature map. Simultaneously, the high-dimensional static feature map carrying stable geographic topographic information is input into a pre-defined second convolutional block attention module. After feature enhancement processing, an enhanced high-dimensional static feature map is obtained, highlighting the key information from both feature paths. Subsequently, based on a pre-defined fusion formula, the enhanced high-dimensional dynamic feature map F... A and Enhanced High-Dimensional Static Feature Map F B To perform interactive fusion, the original channels C of the enhanced high-dimensional dynamic feature map are first compressed using Down(C), then the channel dimensions are adjusted using a 1×1 Conc1×1 convolution, and adaptive weight coefficients are generated using the σ activation function. Finally, feature weighting and fusion are completed by element-wise multiplication (⊙), resulting in the fused feature map F. f .

[0062] In this optional embodiment, two independent convolutional block attention modules enhance the high-dimensional dynamic and static features respectively, accurately capturing and highlighting the core effective information in the two feature paths, filtering redundant noise, and laying a high-quality feature foundation for subsequent fusion. Based on a dedicated fusion formula using channel compression, 1×1 convolution, and element-wise multiplication, a refined interactive fusion of the two enhanced features is achieved. This ensures the adaptability of feature dimensions and gives the fusion process clear mathematical logic and interpretability. It can adaptively integrate the real-time dynamic changes of snow cover with stable geographical context features, making the fused feature map both real-time and stable. This effectively makes up for the information defects of single features, solves the problem of equal information weight in traditional dual-stream fusion, and provides more comprehensive and accurate fused feature support for the feature decoding and reconstruction of the subsequent decoder, greatly improving the accuracy and robustness of subsequent snow cover probability recognition.

[0063] The snow cover probability distribution map satisfies: ; Among them, P f For the snow cover probability distribution map, P A For the initial snow cover probability map, P S For the snow accumulation likelihood plot, P C Let clip(·, 0, 1) be the snow accumulation confidence graph.

[0064] Optionally, such as Figure 2 As shown, the method also includes Based on the snow cover confidence map, the snow cover likelihood map, and the initial snow cover probability map, an anomaly mask corresponding to the target area is generated through a preset conflict determination relationship. The credibility conversion relationship satisfies The anomaly mask satisfies: ; Where A is the anomaly mask, P C For the snow accumulation confidence map, P A For the initial snow cover probability map, P S Let η be the snow likelihood plot, τ be the preset confidence threshold, ||(·) be the indicator function, and ∧ be the logical AND operator.

[0065] Specifically, after generating the snow index feature map, the snow index K of each pixel is substituted into the data, which incorporates the classification threshold Z, the integer coefficient a, and the normal brightness fluctuation threshold V. r The calculation of the preset confidence conversion relationship of various parameters, combined with the pixel normalized brightness V and the ideal standard brightness value V mid The deviation and variance of the snow cover index within the local window D K With the maximum allowable variance threshold D max The matching degree is determined, and an occlusion mask M is introduced. O By removing occluded areas, the snow confidence score C for each pixel is obtained. The snow confidence scores of all pixels are then integrated to generate a snow confidence map P. C Then, the credibility graph P of the snow accumulation is... C Snow-like image P S With the initial snow cover probability map P A Substituting the preset conflict determination relationship, the conflict determination relationship defines high-confidence regions by using a confidence threshold η, and determines the snow accumulation likelihood map P by using a conflict threshold τ. S With the initial snow cover probability map P A The numerical deviation is indicated by the indicator function II ( Pixels in the high-confidence region whose deviation exceeds the threshold are marked as anomalies, and finally an anomaly mask A corresponding to the target region is generated.

[0066] In this optional embodiment, the credibility of snow cover is calculated through a multi-parameter coupled credibility transformation relationship. This comprehensively considers key influencing factors in winter monitoring, such as illumination deviation, local snow index stability, and occlusion interference. This allows the generated snow cover credibility map to accurately quantify the reliability of physical priors, defining scientific high-credibility region boundaries for conflict determination and avoiding miscorrection of physical priors in low-credibility regions. Furthermore, conflict determination based on preset thresholds and indicator functions can accurately identify conflict areas between deep learning model predictions and physical priors within high-credibility regions. The generated anomaly mask clearly marks low-credibility regions of the monitoring results, providing a clear correction range for subsequent physical constraints and fusion decisions. This allows the model to perform weighted fusion / back-off processing only on anomaly regions, ensuring the effectiveness of monitoring results in normal regions and providing accurate anomaly region basis for subsequent uncertainty-driven closed-loop supplementary measurements. This helps the system to specifically trigger UAV supplementary measurements, significantly reducing the false detection risk of pure deep learning models, improving the accuracy and robustness of snow cover monitoring results, and making error tracing and optimization of the entire monitoring process more targeted.

[0067] Optionally, the method further includes: Obtain the prediction uncertainty map output by the second output head of the decoder, wherein the decoder includes a first output head and a second output head in parallel, the first output head is used to output the initial snow accumulation probability map, and the second output head is used to output the prediction uncertainty map; The prediction uncertainty corresponding to each pixel is obtained based on the prediction uncertainty map; When the prediction uncertainty is greater than a preset uncertainty threshold, the corresponding pixel is determined to be an uncertain pixel. The uncertainty area is obtained based on all the uncertain pixels. The corresponding abnormal area is obtained based on the abnormality mask; When the anomaly area and the uncertainty area simultaneously meet all preset UAV retesting conditions, the historical orthophoto of the UAV is updated by the collected real-time orthophoto of the UAV. The UAV retesting conditions include the anomaly area being greater than the uncertainty area, the anomaly area being greater than a preset anomaly area threshold, and the uncertainty area being greater than a preset uncertainty area.

[0068] Specifically, firstly, the prediction uncertainty map output by the second output head of the decoder is obtained. The decoder has a first output head and a second output head in parallel. The first output head is used to output the initial snow probability map corresponding to the input fused feature map, and the second output head is used to synchronously output the prediction uncertainty map, realizing the parallel execution of snow probability prediction and model uncertainty evaluation. Then, the prediction uncertainty corresponding to each pixel is extracted according to the prediction uncertainty map. Pixels with prediction uncertainty greater than a preset uncertainty threshold are identified as uncertain pixels, and the uncertainty area is obtained by counting all uncertain pixels. At the same time, the corresponding abnormal area is obtained by counting the abnormal area based on the previously generated abnormal mask. When the abnormal area and the uncertainty area simultaneously meet the preset UAV supplementary measurement conditions, including the abnormal area being greater than the uncertainty area, the abnormal area being greater than a preset abnormal area threshold, and the uncertainty area being greater than a preset uncertainty area threshold, the real-time orthophoto of the target area is collected by the UAV to update the original historical orthophoto of the UAV.

[0069] In this optional embodiment, the prediction output and uncertainty quantification are completed simultaneously without increasing the model inference time through a dual-output head parallel structure, effectively improving monitoring efficiency. The supplementary testing triggering mechanism with multi-condition joint judgment can accurately distinguish between fluctuations caused by model uncertainty and real anomalies caused by outdated geographical background, avoiding false triggers and invalid supplementary testing, and improving the system's operational economy. By dynamically updating historical orthophotos using real-time acquired UAV images, the timeliness and accuracy of static background information can be guaranteed from the data source level, providing reliable support for subsequent high-dimensional static feature extraction, feature interaction fusion, and snow cover probability calculation, significantly improving the accuracy, robustness, and long-term stability of snow cover monitoring results, forming a closed-loop optimization mechanism of "monitoring-judgment-supplementary testing-update".

[0070] Furthermore, the average uncertainty is determined based on the prediction uncertainty of all pixels. If the average uncertainty is greater than the preset uncertainty threshold, the UAV supplementary measurement process is also triggered, and the historical orthophotos of the UAV are updated based on the real-time orthophotos of the UAV collected in real time.

[0071] like Figure 3 As shown, an embodiment of the present invention provides a snow cover range identification device 300 based on air-ground cooperation, comprising: The acquisition module 310 is used to acquire camera geographic domain map, UAV historical orthophotos and digital elevation model of the target area; The conversion module 320 is used to obtain a snow index feature map by color space conversion based on the geographic domain map of the camera; Mapping module 330 is used to generate a snow prior feature map based on the snow index feature map through mapping transformation; The first encoding module 340 is used to obtain a high-dimensional dynamic feature map based on the camera geographic domain map and the snow index feature map through a preset first encoder. The second encoding module 350 is used to obtain a high-dimensional static feature map based on the historical orthophoto of the UAV and the digital elevation model through a preset second encoder. Decoding module 360 ​​is used to obtain an initial snow accumulation probability map by feature fusion and decoding of the high-dimensional dynamic feature map and the high-dimensional static feature map; The fusion module 370 is used to fuse the initial snow probability map and the prior snow feature map to obtain the snow probability distribution map corresponding to the target area.

[0072] The air-ground coordinated snow cover range identification device of this embodiment is used to implement the air-ground coordinated snow cover range identification method as described above. Its advantages over the prior art are the same as the advantages of the air-ground coordinated snow cover range identification method over the prior art, and will not be repeated here.

[0073] like Figure 4 As shown, an electronic device 400 provided in this embodiment of the invention includes a memory 410 and a processor 420; the memory 410 is used to store a computer program; the processor 420 is used to implement the snow cover range identification method based on air-ground cooperation as described above when the computer program is executed.

[0074] Alternatively, an electronic device 400 includes a memory 410 and a processor 420 coupled to the memory 410; the memory 410 is configured to store a computer program; and the processor 420 is configured to perform the following operations when the computer program is executed: Acquire camera geographic domain maps, historical orthophotos of drones, and digital elevation models of the target area; A snow cover index feature map is obtained by color space conversion based on the camera geographic domain map; A prior snow feature map is generated based on the snow index feature map through mapping transformation; A high-dimensional dynamic feature map is obtained by using a preset first encoder based on the camera geographic domain map and the snow index feature map. A high-dimensional static feature map is obtained by using the historical orthophotos of the UAV and the digital elevation model through a preset second encoder; The high-dimensional dynamic feature map and the high-dimensional static feature map are fused and decoded to obtain the initial snow accumulation probability map; The snow probability distribution map corresponding to the target region is obtained by fusing the initial snow probability map and the prior snow feature map.

[0075] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the snow cover range identification method based on air-ground coordination as described above.

[0076] Alternatively, a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following operations: Acquire camera geographic domain maps, historical orthophotos of drones, and digital elevation models of the target area; A snow cover index feature map is obtained by color space conversion based on the camera geographic domain map; A prior snow feature map is generated based on the snow index feature map through mapping transformation; A high-dimensional dynamic feature map is obtained by using a preset first encoder based on the camera geographic domain map and the snow index feature map. A high-dimensional static feature map is obtained by using the historical orthophotos of the UAV and the digital elevation model through a preset second encoder; The high-dimensional dynamic feature map and the high-dimensional static feature map are fused and decoded to obtain the initial snow accumulation probability map; The snow probability distribution map corresponding to the target region is obtained by fusing the initial snow probability map and the prior snow feature map.

[0077] The present invention will now be described an electronic device 400 that can serve as a server or client of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. Electronic device 400 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 400 can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0078] Electronic device 400 includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or a computer program loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0079] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0080] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for identifying snow cover area based on air-ground collaboration, characterized in that, include: Acquire camera geographic domain maps, historical orthophotos of drones, and digital elevation models of the target area; A snow cover index feature map is obtained by color space conversion based on the camera geographic domain map; A prior snow feature map is generated based on the snow index feature map through mapping transformation. A high-dimensional dynamic feature map is obtained by using a preset first encoder based on the camera geographic domain map and the snow index feature map. A high-dimensional static feature map is obtained by using the historical orthophotos of the UAV and the digital elevation model through a preset second encoder; The high-dimensional dynamic feature map and the high-dimensional static feature map are fused and decoded to obtain the initial snow accumulation probability map; The snow probability distribution map corresponding to the target region is obtained by fusing the initial snow probability map and the prior snow feature map. The snow cover index relationship satisfies: ； Where K is the snow index, V is the normalized brightness, B is the normalized saturation, and α and β are preset weighting coefficients. The snow prior feature map includes a snow likelihood map and a snow confidence map; the step of generating the snow prior feature map from the snow index feature map through mapping transformation includes: Based on the snow index corresponding to each pixel, the corresponding snow likelihood probability is obtained through a preset mapping relationship. The snow likelihood map is generated based on the snow likelihood probability corresponding to all the pixels. The corresponding snow accumulation confidence level is obtained by using a preset confidence level conversion relationship based on the normalized brightness and normalized saturation of the pixel. A snow credibility map is generated based on the snow credibility corresponding to all the pixels. The preset mapping relationship satisfies: ； Where S is the snow accumulation likelihood probability, K is the snow accumulation index, T is the preset bias threshold, M is the preset scaling factor, and clip(·, 0, 1) is the truncation function. The preset confidence conversion relationship satisfies: ； ； ； ； Where C is the snow accumulation confidence level corresponding to the pixel, I is the pixel value of the pixel corresponding to the camera geographic domain map, and V is the normalized brightness corresponding to the pixel. d V is the preset underexposure threshold. s C is the preset overexposure brightness threshold. e B is the exposure confidence level corresponding to the pixel, and B is the normalized saturation level corresponding to the pixel. f V is the preset low saturation threshold. f C is the preset high brightness threshold. f Let g be the fog confidence level corresponding to the pixel, g be a preset texture gradient threshold, and C be the density of the fog. t represents the texture confidence level corresponding to the pixel, otherwise represents otherwise, and ∧ represents and; The snow cover probability distribution map satisfies: ； Among them, P f For the snow cover probability distribution map, P A For the initial snow cover probability map, P S For the snow likelihood diagram, P C Let clip(·, 0, 1) be the snow accumulation confidence graph.

2. The snow cover range identification method based on air-ground cooperation according to claim 1, characterized in that, The step of obtaining the snow index feature map through color space conversion based on the camera's geographic domain map includes: The color space conversion is performed on the camera geographic domain map to obtain the corresponding color space map; Extract the normalized brightness and normalized saturation of each pixel based on the color space diagram; Based on the normalized brightness and normalized saturation of the pixel, the corresponding snow index is obtained through a preset snow index relationship; The snow index feature map is generated based on the snow index corresponding to all the pixels.

3. The snow cover range identification method based on air-ground cooperation according to claim 1, characterized in that, The step of obtaining a high-dimensional dynamic feature map based on the camera geographic domain map and the snow index feature map through a preset first encoder includes: The camera geographic domain map and the snow index feature map are aligned by projection to obtain the corresponding dynamic data tensor; The dynamic data tensor is input into the first encoder to output the corresponding initial high-dimensional dynamic feature map. The high-dimensional dynamic feature map is obtained by weighted aggregation of multiple initial high-dimensional dynamic feature maps of a preset continuous time series corresponding to the current time.

4. The snow cover range identification method based on air-ground cooperation according to claim 1, characterized in that, The step of obtaining a high-dimensional static feature map based on the historical orthophotos of the UAV and the digital elevation model through a preset second encoder includes: The historical orthophotos of the UAV and the digital elevation model are aligned by projection to obtain the corresponding static data tensor; The static data tensor is input into the second encoder to output the corresponding high-dimensional static feature map.

5. The method for identifying snow cover area based on air-ground coordination according to claim 1, characterized in that, The step of obtaining an initial snow accumulation probability map by feature fusion and decoding of the high-dimensional dynamic feature map and the high-dimensional static feature map includes: The high-dimensional dynamic feature map and the high-dimensional static feature map are interactively fused to obtain a fused feature map; The fused feature map is input into a preset decoder to output the initial snow accumulation probability map.

6. The snow cover range identification method based on air-ground cooperation according to claim 5, characterized in that, The step of interactively fusing the high-dimensional dynamic feature map and the high-dimensional static feature map to obtain a fused feature map includes: The high-dimensional dynamic feature map is input into a preset first convolutional block attention module to output an enhanced high-dimensional dynamic feature map. The high-dimensional static feature map is input into a preset second convolutional block attention module to output an enhanced high-dimensional static feature map; The enhanced high-dimensional dynamic feature map and the enhanced high-dimensional static feature map are fused together through the interactive process to obtain the fused feature map; The fused feature map satisfies: ； Among them, F f For the fused feature map, F A For the enhanced high-dimensional dynamic feature map, F B For the enhanced high-dimensional static feature map, Down(C) is the value of the enhanced high-dimensional dynamic feature map F. A The original channel C is compressed, Conv 1×1 For a 1×1 convolution, σ is the activation function, and ⊙ is the element-wise multiplication.

7. The snow cover range identification method based on air-ground cooperation according to claim 5, characterized in that, Also includes: Based on the snow cover confidence map, the snow cover likelihood map, and the initial snow cover probability map, an anomaly mask corresponding to the target area is generated through a preset conflict determination relationship. The anomaly mask satisfies: ； Where A is the anomaly mask, P C For the snow accumulation confidence map, P A For the initial snow cover probability map, P S Let η be the snow likelihood plot, τ be the preset confidence threshold, ||(·) be the indicator function, and ∧ be the logical AND operator.

8. The snow cover range identification method based on air-ground cooperation according to claim 7, characterized in that, Also includes: Obtain the prediction uncertainty map output by the second output head of the decoder, wherein the decoder includes a first output head and a second output head in parallel, the first output head is used to output the initial snow accumulation probability map, and the second output head is used to output the prediction uncertainty map; The prediction uncertainty corresponding to each pixel is obtained based on the prediction uncertainty map; When the prediction uncertainty is greater than a preset uncertainty threshold, the corresponding pixel is determined to be an uncertain pixel. The uncertainty area is obtained based on all the uncertain pixels. The corresponding abnormal area is obtained based on the abnormality mask; When the anomaly area and the uncertainty area simultaneously meet all preset UAV retesting conditions, the historical orthophoto of the UAV is updated by the collected real-time orthophoto of the UAV. The UAV retesting conditions include the anomaly area being greater than the uncertainty area, the anomaly area being greater than a preset anomaly area threshold, and the uncertainty area being greater than a preset uncertainty area.

Images