An unmanned aerial vehicle platform-oriented ice nucleus moraine dam detection method

Abstract

The application discloses a kind of ice nucleus moraine dam detection methods for unmanned aerial vehicle platform.The method comprises: collecting moraine dam area image and pre-processing;Image is input to the ECNP-YOLOv7 detection model deployed in airborne edge computing device, and the landform features such as hot melting sinkhole, creep wrinkle, mound subsidence, adjacent snowdrift are identified;Target statistical vector is generated and ice moraine dam risk score is calculated by light statistical analysis module;Real-time monitoring communication link quality, adaptively selecting backhaul video stream rendered after OSD or narrowband data containing key information.The application improves the ice nucleus moraine dam detection precision, adapts edge computing device computing power, and can ensure reliable backhaul of key data in weak network environment.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, specifically to a method for detecting ice core and moraine dams for unmanned aerial vehicle (UAV) platforms. Background Technology

[0002] Against the backdrop of global warming, glaciers in High Mountain Asia are retreating significantly, leading to frequent glacial lake outburst floods (GLOFs), which pose a major threat to downstream communities and infrastructure. Among various types of glacial lake dams, ice-cored moraine dams are particularly vulnerable to collapse. These dams consist of an internal ice core and a surface layer of rock and soil debris, and their stability is significantly controlled by thermodynamic processes, making them highly susceptible to structural instability under conditions such as rising temperatures or thermal erosion by lake water.

[0003] Existing research indicates that the melting of ice cores buried within glacial till bars is a gradual process that directly reflects specific geomorphic features on the surface, including: thermocline sinkholes (when the ice cores buried beneath the till layer partially melt, the debris layer above the ice loses support and collapses due to gravity, visually forming a funnel-shaped or pitted depression); and creep folds (the till layer, because it contains ice cores, has a rounded outline. Due to the uneven thickness of the surface debris, the underlying ice cores melt differentially, causing the surface sediments to creep and compress, ultimately shaping similar features on the rounded base). The features include: tiny "folded" ridges; hill-like subsidence (also due to the differential melting of buried ice cores, the moraine layer presents as a large number of isolated mounds, ridges and depressions, lacking a regular orientation); and adjacent snowdrifts (the temperature of the ice buried inside the moraine dam is much lower than that of the surrounding environment. This low temperature is conducted to the surface, causing the snow in its affected area to melt significantly slower in spring and summer than in the surrounding ice-free areas, resulting in snowdrifts that remain next to the moraine dam for a long time). Therefore, identifying the above four key geomorphic features based on computer vision technology is a scientific and effective means of determining whether a dam is a moraine dam.

[0004] However, in extreme environments with high altitudes, complex terrain, and limited communication, existing technical solutions have significant drawbacks: (1) In terms of satellite remote sensing monitoring, the mainstream method at present is to use optical satellites or synthetic aperture radar for large-scale monitoring. The resolution of existing free satellite data is usually between 10 meters and 30 meters. Even commercial satellites have difficulty in clearly capturing landform features at the meter or even sub-meter level. Tiny features such as thermal melting sinkholes and creep folds are precisely direct evidence of the existence of buried ice cores. Insufficient resolution will directly lead to missed detections or misjudgments.

[0005] (2) In UAV surveying, existing solutions mostly employ Structure from Motion (SfM) technology, which involves acquiring massive amounts of high-overlapping photos and subsequently generating high-precision Digital Elevation Models (DEMs) and Digital Orthophoto Maps (DOMs) to quantitatively analyze ice core degradation. However, SfM reconstruction requires enormous storage space and high-performance workstation computing power, and the data processing workflow is cumbersome. For tasks that only need to identify specific landform features, this method is a huge waste of computing resources and operation time.

[0006] (3) In terms of cloud-based intelligent detection, with the development of IoT technology, some solutions attempt to transmit front-end video back to cloud servers to run deep learning models. However, relying on cloud computing power requires a high-bandwidth network. Most glacial lakes are located in high-altitude uninhabited areas and lack public network base station coverage. Once disconnected from the network, the entire intelligent detection system immediately fails.

[0007] (4) In terms of manual field surveys, traditional methods also include sending researchers or engineers on foot to glacial moraine dams for sampling and surveying. The areas where glacial moraine dams are located are usually steep and have complex road conditions, with potential risks such as rockfalls and ice avalanches at any time, posing a high threat to personal safety for manual operations. At the same time, due to the high-altitude hypoxic environment and physical limitations, manual surveys can usually only cover a local area of ​​the dam body, making it difficult to conduct a comprehensive and detailed investigation of large-scale, long-distance glacial moraine dams.

[0008] In summary, existing technical solutions cannot meet the urgent need for efficient, accurate, and automated identification of ice cores and moraine dams in high-altitude, cold, and grid-free environments, and there are significant technical bottlenecks that need to be overcome. Summary of the Invention

[0009] The purpose of this invention is to provide a method for detecting ice core and moraine dams for unmanned aerial vehicle (UAV) platforms, aiming to solve the technical problems of low detection accuracy, reliance on external computing power and networks, low operational efficiency, and safety risks in the prior art.

[0010] This invention provides a method for detecting ice core and moraine dams for unmanned aerial vehicle (UAV) platforms, comprising the following steps: S1. Real-time images of the glacial moraine area are acquired using imaging equipment mounted on a drone, and the real-time images are filtered and enhanced. S2. Input the preprocessed image into the ECNP-YOLOv7 detection model deployed on an airborne edge computing device to detect and identify the geomorphic features of ice cores and moraines in the image; the geomorphic features include thermomelt sinkholes, creep folds, mound subsidence, and adjacent snowdrifts; The ECNP-YOLOv7 detection model uses the YOLOv7 network as its basic architecture. The backbone network in the YOLOv7 network model is replaced with the first 7 stages of EfficientNetV2-S as the feature extraction backbone network, and the CBAM attention module is embedded between the backbone network and the head network. S3. Generate the target statistical vector of the current frame by integrating the lightweight statistical parsing module at the output end of the ECNP-YOLOv7 detection model, and perform binarization mapping on the target statistical vector to generate a binarized state vector; then perform weighted fusion on the binarized state vector based on preset asymmetric weights to calculate the glacial moraine dam risk score. S4. The airborne edge computing device monitors the quality of the communication link between the UAV and the ground station in real time, and adaptively selects to transmit the video stream rendered by OSD through the digital image transmission link or to transmit narrowband data containing the detection results through the telemetry data link based on the quality of the communication link. The video stream includes real-time footage, target detection boxes, confidence scores corresponding to the target detection boxes, target statistical vectors, glacial moraine risk scores, and a determination of whether glacial moraine exists based on a comparison of the glacial moraine risk scores with a preset scoring threshold; the narrowband data includes target statistical vectors and glacial moraine risk scores.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. High detection accuracy: Based on the comprehensive improvement of EfficientNetV2-S network, CBAM attention mechanism and normalized Wasserstein distance (NWD), the ability to extract small terrain features is significantly improved.

[0012] 2. Edge computing adaptation: By pruning the structured model, the computational load and memory usage of the model are greatly reduced, enabling it to run in real time on airborne edge computing devices with limited computing power.

[0013] 3. Strong communication adaptability: The dual-mode link and dual-threshold hysteresis switching strategy ensure that critical information can still be reliably transmitted in environments without network or with strong interference, avoiding information loss due to image transmission interruption. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0015] Figure 1 This is a schematic flowchart of an ice core and moraine dam detection method for an unmanned aerial vehicle (UAV) platform according to an embodiment of the present invention. Figure 2 This is an overall architecture diagram of the ECNP-YOLOv7 detection model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the Fused-MBConv module and the MBConv module according to an embodiment of the present invention; wherein, (a) is a schematic diagram of the MBConv module; and (b) is a schematic diagram of the Fused-MBConv module. Figure 4 This is a schematic diagram of the CBAM attention module according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the channel attention module according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the spatial attention module in an embodiment of the present invention; Figure 7 This is a display diagram of the video stream transmitted back when the digital image transmission link is enabled in an embodiment of the present invention. Detailed Implementation

[0016] 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 only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, this invention provides a method for detecting ice core and moraine dams for unmanned aerial vehicle (UAV) platforms, comprising the following steps: Step S1: Collect real-time images of the glacial moraine area using the imaging equipment mounted on the UAV, and perform filtering and enhancement processing on the real-time images; It should be emphasized that during the training phase of the ECNP-YOLOv7 detection model, image preprocessing includes filtering enhancement, horizontal flipping, vertical flipping, and region cropping. In an optional implementation, the filtering enhancement process in step S1 employs a bilateral filtering algorithm, the calculation expression of which is: ; in, The coordinates of the target pixel; Indicates the target pixel point The neighborhood centered on; For the neighborhood The coordinates of any pixel within the range; For point The corresponding pixel value (original pixel value); The pixel value corresponding to the target pixel (the new pixel value after filtering); For point The corresponding weighting coefficient (which is determined by both the spatial domain kernel and the value domain kernel).

[0018] In step S1, a high-resolution visible light imaging device mounted on a UAV is used to conduct low-altitude aerial photography of the target moraine dam area, acquiring real-time images containing one or more key geomorphic features, including thermal melting cavities, creep folds, mound subsidence, and adjacent snowdrifts. Due to the drastic changes in lighting conditions in high-altitude uninhabited areas, coupled with inherent noise from UAV electronic equipment, the original images generally suffer from quality issues such as blurring, abnormal contrast, areas that are too bright or too dark, and random noise. Furthermore, the generalization ability of deep learning models is highly dependent on the size and diversity of the training dataset. Therefore, this step performs filtering and enhancement processing after acquiring the real-time images to improve image quality, thereby providing high-quality input data for subsequent feature extraction and detection.

[0019] Specifically, the filtering enhancement process aims to suppress image noise, preserve subtle topographical details, and improve the overall contrast and brightness distribution of the image. This implementation uses a bilateral filtering algorithm. Compared to traditional methods such as Gaussian filtering and mean filtering, bilateral filtering can effectively preserve edge information while smoothing noise, and is particularly suitable for high-frequency details of key topographical features of glacial till dams, such as the edges of thermal sinkholes and creep-folded micro-ridges.

[0020] Furthermore, to improve the robustness of the ECNP-YOLOv7 detection model to changes in scale, viewpoint, and flip direction, and to alleviate the problem of insufficient data due to difficulties in on-site acquisition, this invention, when training the ECNP-YOLOv7 detection model, includes the following geometric augmentation operations in addition to performing image filtering enhancement: Horizontal Flip: Mirrors the image horizontally along the vertical central axis. This operation simulates a drone taking symmetrical views of the same terrain from different flight paths, helping to learn rotation invariance.

[0021] Vertical Flip: Mirrors the image vertically along the horizontal midline. This operation enhances the model's adaptability to inverted perspectives, taking into account the undulating terrain of glacial moraine.

[0022] Region cropping: Randomly cropping sub-regions of different sizes from the original image. On the one hand, this generates diverse local views, forcing the model to focus on terrain details rather than the global background; on the other hand, it can indirectly achieve scale transformation, simulating the observation effects of drones at different flight altitudes.

[0023] By randomly combining the above operations (e.g., flipping followed by cropping, or multiple cropping operations with different parameters), the number of original images can be multiplied, forming a large-scale and diverse training dataset. Experiments show that the dataset enhanced by bilateral filtering and geometric augmentation can significantly improve the detection accuracy and generalization performance of subsequent deep learning models in real-world, cold, and unnetworked environments.

[0024] Step S2: Input the preprocessed image into the ECNP-YOLOv7 detection model deployed on the airborne edge computing device to detect and identify the geomorphic features of ice cores and moraines in the image; the geomorphic features include thermal sinkholes, creep folds, mound subsidence, and adjacent snowdrifts.

[0025] Among them, such as Figure 2 As shown, the ECNP-YOLOv7 detection model uses the YOLOv7 network as its basic architecture. The backbone network in the YOLOv7 network model is replaced with the first 7 layers of EfficientNetV2-S as the feature extraction backbone network, and a CBAM (Convolutional Block Attention Module) attention module is embedded between the backbone network and the head network.

[0026] In one alternative implementation, such as Figure 2 and Figure 3 As shown, the backbone network consists of one CBS module, three Fused-MBConv modules, and three MBConv modules cascaded in sequence. The CBS module is used to perform convolution, batch normalization (BN), and SiLU activation operations on the preprocessed image in sequence, and output the initial feature map. Each Fused-MBConv module contains a Conv3×3 layer, an SE attention layer, and a Conv1×1 layer connected sequentially from the input to the output. The feature map is first processed by the Conv3×3 layer for spatial feature extraction, then by the SE attention layer for channel recalibration, and finally by the Conv1×1 layer for channel transformation, outputting a fused convolutional feature map. Each MBConv module contains, from input to output, a Conv1×1 dimensionality-upgrading layer, a DepthwiseConv3×3 layer, an SE attention layer, and a Conv1×1 dimensionality-reducing layer. This is used to first upgrade the feature map using Conv1×1, then perform depthwise separable convolution operations using Depthwise Conv3×3, then recalibrate the channel weights using the SE attention layer, and finally reduce the dimensionality using Conv1×1 to output a depthwise convolutional feature map. The backbone network is processed sequentially through the CBS module, three Fused-MBConv modules, and three MBConv modules, ultimately outputting multi-scale input feature maps.

[0027] In one alternative implementation, the CBAM attention module includes a channel attention module and a spatial attention module cascaded in sequence.

[0028] Among them, the CBAM attention module is a lightweight hybrid attention model, and its overall structure is as follows: Figure 4 As shown, since convolutional operations extract features by mixing channel and spatial information, the channel attention module emphasizes "which features are more important," while the spatial attention module focuses on "where is more important." Therefore, the CBAM attention module integrates the channel and spatial attention modules, aiming to emphasize the meaning and features in two important dimensions: the channel axis and the spatial axis.

[0029] Specifically, such as Figure 5 As shown, the channel attention module is used to process the input feature map. F Global max pooling and global average pooling are performed separately. The results of both pooling are then input into a shared neural network (SMLP). The two feature vectors are then fused element-wise, and finally, a sigmoid activation function is used to generate channel attention weights. And channel attention weights With input feature map F Perform element-wise multiplication and output intermediate feature maps. For the spatial attention module, the calculation formula is as follows: ; in, This represents element-wise multiplication; Indicates the Sigmoid activation operation; This indicates a global average pooling operation; This represents the global max pooling operation; This indicates the operation of a multilayer perceptron.

[0030] like Figure 6 As shown, the spatial attention module is used to receive intermediate feature maps. and the intermediate feature map Global max pooling and global average pooling are performed separately. The results of the two pooling operations are concatenated along the channel dimension, and then spatial attention weights are generated by passing them through a 7×7 convolutional layer and a sigmoid activation function. And spatial attention weights With intermediate feature map Perform element-wise multiplication to output the final enhanced feature map. The calculation formula is as follows: ; ; in, This represents element-wise multiplication; Indicates the Sigmoid activation operation; This indicates a global average pooling operation; This represents the global max pooling operation; This represents a 7×7 convolution operation.

[0031] Specifically, this application provides an improved model based on the YOLOv7 network (ECNP-YOLOv7 detection model). By constructing an efficient feature extraction backbone network and a multi-scale fusion detection head, it achieves high-precision localization and recognition of targets in input images. The following is combined with... Figure 2 and Figure 3 The technical solution of the present invention will be described in detail below.

[0032] like Figure 2 As shown, the ECNP-YOLOv7 detection model proposed in this invention mainly consists of four parts: input layer, backbone network, head network, and post-processing module.

[0033] Specifically, the model's workflow is as follows: The preprocessed image first enters the Backbone for multi-scale feature extraction; the extracted feature maps are enhanced by the CBAM attention module and then sent to the Head part for feature fusion and decoding; finally, the Head outputs three prediction tensors at different scales, which are then subjected to non-maximum suppression (NMS) to obtain the final object detection result.

[0034] First, the backbone network is responsible for extracting basic features from the input image and outputting multi-scale feature maps. The backbone network adopts a cascaded structure, which consists of one CBS module, three Fused-MBConv modules, and three MBConv modules.

[0035] The initial feature extraction (CBS module) involves the input image first passing through a CBS (Conv-BatchNorm-SiLU) module. This module performs preliminary convolution operations on the image, followed by batch normalization to accelerate convergence and stabilize training. Finally, it passes through the SiLU activation function to output a feature map containing initial texture and edge information.

[0036] Shallow Feature Fusion (Fused-MBConv Module): The subsequent three Fused-MBConv modules are used to extract features from shallow to mid-level layers. The internal structure of each module is designed as follows: First, spatial feature extraction is performed through a 3×3 convolutional layer to capture local details; then, a Squeeze-and-Excitation (SE) attention mechanism is introduced to recalibrate the channel dimensions, enhancing the response of important feature channels and suppressing invalid channels; finally, channel transformation is performed through a 1×1 convolutional layer to achieve feature fusion and dimensionality increase / decrease. This design effectively balances computational cost and feature representation capability.

[0037] Deep Semantic Feature Extraction (MBConv Module): The last three MBConv modules are used to extract deep semantic features. Internally, they employ a depthwise separable convolutional structure: first, dimensionality is increased through a 1×1 convolutional layer to expand the feature representation space; then, a depthwise 3×3 convolutional layer performs independent spatial filtering in each channel, significantly reducing the number of parameters; subsequently, an SE attention layer is introduced again to recalibrate the channel weights; finally, dimensionality is reduced through a 1×1 convolutional layer, outputting a compact and semantically rich deep convolutional feature map.

[0038] After layers of processing by the backbone, the network outputs three feature maps at different resolutions (corresponding to...). Figure 2 The three branches output from the right side of the Backbone provide feature information at different levels of abstraction.

[0039] Second, after the Backbone outputs the feature map, this invention introduces a CBAM attention module (such as...). Figure 2 (The connection between the Backbone and Head is shown in the diagram). This module cascades channel attention and spatial attention.

[0040] The channel attention module obtains global information through global max pooling and global average pooling, generates channel attention weights, and tells the network "which feature channels to focus on".

[0041] The spatial attention module generates spatial attention weights by concatenating pooling results and performing convolution, telling the network "what location in the feature map to focus on".

[0042] The feature map processed by CBAM is fed into the Head section, which effectively suppresses background noise and highlights the target features.

[0043] Third, the Head part adopts a structure similar to PANet (Path Aggregation Network), which aims to fuse multi-scale features output by the Backbone to detect targets of different sizes.

[0044] Top-down and bottom-up paths: such as Figure 2 As shown, the Head section includes upsampling (UPSample) and downsampling (MPConv2), typically referring to max pooling or strided convolution operations. For the top-down path, the high-level semantic feature map, after upsampling and processing by the CBS module, is concatenated with the same-scale feature map output from the Backbone, achieving the transfer of high-level semantic information to the shallower layers. For the bottom-up path, the fused feature map is then downsampled using MPConv2 and concatenated with deeper feature maps, achieving the transfer of shallow localization information to deeper layers.

[0045] After multiple feature fusions and further feature extraction by the E-ELAN (Efficient ELAN) module, the feature map enters the final prediction branch. In each prediction branch, the feature map first undergoes efficient feature transformation via the REPConv (Reparameterized Convolution) module, followed by the CBM (Conv-BN-Mish) module for final feature decoupling and classification prediction. The outputs of these three processes correspond to the detection tasks for small, medium, and large targets, respectively, ultimately generating a prediction tensor containing the bounding box coordinates, confidence scores, and class probabilities.

[0046] Fourth, the model's Head section ultimately outputs three tensors, each containing the coordinates of the predicted bounding box. The algorithm considers confidence levels and class probabilities. Since the anchor mechanism generates a large number of overlapping predicted boxes, this invention employs a non-maximum suppression algorithm for post-processing: predictive boxes with confidence levels higher than a threshold are selected; the intersection-over-union (IoU) ratio between the predicted boxes is calculated; the box with the highest confidence level is retained, while other overlapping boxes with excessively large IoU are suppressed; finally, the target detection boxes and their class labels after NMS filtering are output, completing the detection of targets in the image.

[0047] In one optional implementation, the total loss function of the ECNP-YOLOv7 detection model during the training phase includes: bounding box regression loss term, target confidence loss term, and classification loss term; The total loss function is calculated as follows: ; in, For bounding box regression loss term; For target confidence loss term; For classification loss terms; , , These are the weight coefficients corresponding to the bounding box regression loss term, the target confidence loss term, and the classification loss term, respectively.

[0048] The bounding box regression loss term uses the normalized Wasserstein distance instead of the original YOLOv7 network's CIoU (Complete IoU) loss function.

[0049] Specifically, the original YOLOv7 network uses CIoU as the bounding box regression loss function. However, from the aerial perspective of a UAV, "initial thermal melting sinkholes" and "small neighboring snowdrifts" often appear as extremely small targets. When the predicted bounding box does not overlap with the ground truth bounding box, the CIoU loss function cannot provide effective gradient information, making it difficult for small target detection to converge. Therefore, this invention uses normalized Wasserstein distance instead of the CIoU loss function.

[0050] In one alternative implementation, the process of forming the bounding box regression loss term is as follows: step a The predicted bounding boxes output by the ECNP-YOLOv7 detection model and the corresponding real frame labels Represented as two-dimensional Gaussian distributions and ;in, , These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. , These are the widths of the predicted bounding box and the ground truth bounding box, respectively; , These are the lengths of the predicted bounding box and the ground truth bounding box, respectively; , These are the mean vectors of the predicted bounding box and the ground truth bounding box, respectively; , These are the covariance matrices of the predicted bounding box and the ground truth bounding box, respectively.

[0051] Specifically, NWD is a target detection bounding box similarity measurement method based on optimal transport theory. From the perspective of probability distribution, it transforms the bounding box matching problem into a distance metric between distributions. A bounding box can be represented as... ,in, Indicates the center coordinates; Indicates the width of the bounding box; This represents the length of the bounding box. Further calculations require converting it into two Gaussian distribution parameters: (1) the mean vector. (2) Covariance matrix This describes the scale features of the bounding box. The calculations for both are as follows: ; ; Based on this, a prediction box is now given. and real frame The corresponding two-dimensional Gaussian distributions are as follows: and .

[0052] step b Calculate the second-order Wasserstein distance between the two-dimensional Gaussian distributions corresponding to the predicted bounding box and the ground truth bounding box. Its expression is: ; Specifically, the two-dimensional Gaussian distributions corresponding to the predicted bounding boxes and the ground truth bounding boxes. and second-order Wasserstein distance It can be represented as: ; in, The squared Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; This is the trace operation of a matrix.

[0053] In object detection tasks, both the predicted bounding box and the ground truth bounding box are horizontal rectangles, so the above formula can be simplified to: .

[0054] step c The similarity measure is obtained by exponentially normalizing the second-order Wasserstein distance and mapping it to the [0,1] interval. Its expression is: ; in, This is the scale normalization factor.

[0055] Specifically, due to the second-order Wasserstein distance As a distance metric, it cannot be directly used as a similarity metric; it needs to be mapped using exponential normalization. Similarity measure is obtained after intervals. . It is a scale normalization factor, which is reflected as a hyperparameter during training and needs to be set manually in advance.

[0056] step d Calculate the bounding box regression loss term based on the similarity metric. Its calculation expression is: .

[0057] In an optional implementation, the method of this application further includes performing structured model pruning during the training phase of the ECNP-YOLOv7 detection model, specifically: To further reduce the floating-point computation and memory usage of the ECNP-YOLOv7 detection model while ensuring detection accuracy, and to make it suitable for UAV-borne edge computing devices with limited computing power, this invention introduces a structured channel pruning technique based on batch normalized layer scaling factor during the model training stage.

[0058] Step A: Scaling factor for the batch normalization layer after all convolutional layers in the ECNP-YOLOv7 detection model. Applying L1 regularization constraints reduces the number of redundant channels that contribute less to feature extraction. The values ​​gradually approach 0 during backpropagation, thereby automatically identifying and sparsifying the importance of model channels, and ultimately obtaining the corrected total loss function. Its expression is: ; in, The sparsity penalty coefficient; Scaling factor for all batch normalization layers with L1 regularization constraints applied The set of absolute values; Step B: Based on the target pruning rate and the set of absolute values Calculate the global pruning threshold Global pruning threshold For the set of absolute values The number after sorting according to the preset order The numerical values ​​corresponding to quantiles.

[0059] Step C: For any convolutional layer, the first... iEach channel is marked as retained if its scaling factor absolute value is greater than or equal to the global pruning threshold, otherwise it is marked as discarded, thereby generating a binary pruning mask; The expression for calculating the binary clipping mask is: ; in, For binary cropping mask; The first convolutional layer Scaling factor corresponding to each channel; This is the global pruning threshold.

[0060] Step D: For the backbone network, set all channel masks of the input and output layers of the Fused-MBConv module to 1, and perform masking operations only on the Conv3×3 layer; set all masks of all layers of the MBConv module to 1; for the head network, perform masking operations directly; for the last convolutional layer of the detection head, set all masks to 1.

[0061] Specifically, to prevent pruning from disrupting the connectivity of the network structure, especially the residual structure in the EfficientNetV2-S backbone, this invention divides the ECNP-YOLOv7 detection model into three regions and applies different strategies accordingly: First, the backbone network: Because the Fused-MBConv module contains residual connections, direct pruning will lead to dimension mismatch. This invention employs a "bottleneck-internal pruning" strategy: it forcibly locks the channel masks of the input and output layers of the Fused-MBConv module to all 1s, preventing pruning operations, and only sets the global pruning threshold. This applies to the Conv3×3 layer within the module. If the scaling factor of a channel in the Conv3×3 layer is lower than the global pruning threshold... If the expansion channel and its associated weights are not found, the expansion channel is removed. This dynamically reduces the expansion ratio of the module, significantly compressing parameters while strictly maintaining the consistency of input and output dimensions required for the residual addition operation.

[0062] The EfficientNetV2-S backbone network employs the traditional MBConv module in its mid-to-late stages. Since the core operator within the MBConv module is a depthwise separable convolution, it already possesses extremely high parameter sparsity and computational efficiency. Forcibly pruning it not only results in minimal marginal gains in reducing overall model parameters but also easily leads to dimensionality conflicts in residual connections. Therefore, this invention adopts a strategy of fully retaining the MBConv module, forcibly locking the pruning mask of all layers in this stage to 1. This achieves an optimal engineering balance between target detection accuracy and onboard inference speed while completely eliminating underlying hardware adaptation risks and compilation crashes.

[0063] Second, the head network: For regular convolutional layers in the head network, a global pruning threshold is directly applied. Generate a mask. If M=0, physically remove the channel and its associated convolutional kernel.

[0064] Third, the output layer: For the last convolutional layer of the detection head, its output channel number corresponds to a fixed number of detection categories and anchor boxes. Its mask M is forcibly set to all 1s, and no pruning is performed to ensure that the model output format is compatible with the subsequent statistical analysis module.

[0065] Furthermore, it should be noted that after pruning the structured model, because the physical reconstruction alters some feature responses, a small amount of fine-tuning training is needed on the reconstructed ECNP-YOLOv7 detection model using a smaller learning rate. This is to recover from the accuracy fluctuations that may be caused by the reduction in channels, ultimately achieving the optimal balance between accuracy and inference speed.

[0066] Step S3: Generate the target statistical vector for the current frame using the lightweight statistical parsing module integrated into the output of the ECNP-YOLOv7 detection model. and the target statistical vector Perform binarization mapping to generate a binary state vector. Then, based on preset asymmetric weights, the binarized state vector is... Weighted fusion was performed to calculate the risk score of glacial moraine dams. .

[0067] In one optional implementation, step S3 specifically includes: Step S31: Generate a four-dimensional target statistical vector using the lightweight statistical parsing module. ;in, This represents the total number of thermal cavities detected in the current frame; This represents the total number of creep folds detected in the current frame; This represents the total number of hill-shaped depressions detected in the current frame; This represents the total number of neighboring snowdrifts detected in the current frame.

[0068] Step S3 employs an "end-to-end statistical analysis-weighted fusion" decision-making logic. This step aims to directly convert the results of visual inspection into a glacial moraine dam risk score.

[0069] Specifically, this invention integrates a lightweight statistical analysis module at the output of the ECNP-YOLOv7 detection model. This module can output a four-dimensional target statistical vector without affecting the original output. .

[0070] In addition, it should be noted that step S31 utilizes the parallel computing capabilities of the onboard edge computing device to directly complete the statistics, avoiding the additional delay caused by the subsequent secondary traversal of the massive detection boxes.

[0071] Step S32: Use a non-zero indicator function to set the target statistical vector. Each element in , , , Perform binarization mapping to generate a binary state vector. When an element is greater than 0, it means that the target of the element's class exists, and the mapping is 1; when an element is equal to 0, it means that the target of the element's class does not exist, and the mapping is 0. ; in, for The value of an element after binarization mapping; for The value of an element after binarization mapping; for The value of an element after binarization mapping; for The value of an element after binarization mapping; Specifically, considering that the identification logic of glacial moraine dams is mainly based on the "existence" of geomorphic features rather than their "absolute quantity", the system introduces a non-zero indicator function. For target statistical vector Perform state mapping (i.e., binarization mapping): This mapping generates a binary state vector. .

[0072] Step S33: Based on prior knowledge of glaciology, assign risk weight coefficients to each geomorphic feature. Calculate the risk score of glacial moraine dams ;in, The output of the non-zero indicator function, i.e., the target statistical vector. The value of each element after binarization mapping.

[0073] Step S4: The airborne edge computing device monitors the quality of the communication link between the UAV and the ground station in real time, and adaptively selects to transmit the video stream rendered by OSD through the digital image transmission link or transmit the narrowband data containing the detection results through the telemetry data link based on the quality of the communication link. The video stream includes real-time footage, target detection boxes, the confidence scores of the target detection boxes, target statistical vectors, glacial moraine risk scores, and a determination of whether glacial moraine exists based on a comparison of the glacial moraine risk scores with preset score thresholds. Figure 7 As shown; narrowband data includes target statistical vectors and glacial moraine dam risk scores.

[0074] This application overcomes the traditional reliance on broadband image transmission for UAV monitoring by proposing a dual-mode adaptive backhaul mechanism based on link status awareness, with unified logical scheduling performed by an onboard edge computing device. Specifically, two transmission channels with different priorities are abstracted at the logical layer: a "digital image transmission link" and a "telemetry data link". When the UAV image transmission signal is good, the digital image transmission link is activated to transmit the video stream rendered by the OSD of the edge computing device, including real-time images, target detection boxes, the confidence scores of the target detection boxes, and target statistical vectors. Glacial moraine dam risk assessment S and risk assessment based on glacial moraine dams S The determination of whether an ice core / moat dam exists is based on a comparison with a preset scoring threshold. When the UAV image transmission signal is poor, the telemetry data link is activated to transmit narrowband data with extremely low bandwidth usage, including target statistical vectors. and glacial moraine dam risk assessment S .

[0075] In one optional implementation, step S4 specifically includes: Step S41: The airborne edge computing device establishes a link status monitoring thread through the UAV SDK (Software Development Kit), reads the underlying status parameters of the communication module at a fixed frequency, and calculates a comprehensive link quality score based on the Received Signal Strength Indicator (RSSI) and Signal-to-Noise Ratio (SNR). The calculation formula is as follows: ; in, The weighting coefficients for the received signal strength indication; These are the weighting coefficients for the signal-to-noise ratio; Step S42: Set the upper limit threshold for communication quality. and lower threshold And execute the dual threshold hysteresis switching logic: when At this time, the digital image transmission link is activated to transmit the video stream back, and the narrowband data transmission thread is suspended. when At this time, the telemetry data link is activated to transmit narrowband data back; when When the system starts, the current transmission state remains unchanged; if the initial comprehensive link quality score is in the range specified in the original text, the transmission state will be maintained. and In between, the digital video transmission link is enabled by default to prioritize video data backhaul.

[0076] Specifically, to avoid frequent transmission mode switching in critical signal regions, this application employs a dual-threshold hysteresis switching strategy. The upper limit threshold for communication quality is set as follows: The lower threshold is The airborne edge computing device executes the following state logic: when At this time, the digital image transmission link is enabled, and the telemetry data link is disabled. The system activates the video encoder and transmits a full-resolution video stream rendered by OSD. Simultaneously, the narrowband data transmission thread is suspended to conserve flight control serial port resources. when At this time, disable the digital image transmission link and enable the telemetry data link. Immediately stop sending video data to the image transmission channel to prevent congestion from causing flight control command delays. Start data encapsulation, package the latest detection results into a custom telemetry frame, send it to the UAV flight controller via UART serial port, and then transmit it back to the ground terminal via the flight controller. when At this time, the current transmission state remains unchanged, and the system utilizes the hysteresis interval to absorb minor fluctuations in signal strength, ensuring system stability. When the system starts up, if the initial integrated link quality score is within a certain range... and In between, the digital video transmission link is enabled by default to prioritize video data backhaul.

[0077] To verify that the ECNP-YOLOv7 detection model pruned by the structured model in this application performs best in terms of a comprehensive trade-off between accuracy and speed, the ECNP-YOLOv7 detection model in this application was compared with models such as Faster R-CNN, YOLOv5s, YOLOv7, YOLOv12n, YOLOv26n, and ECN-YOLOv7 in terms of precision, recall, F1 score, mean precision, and frame rate. The results of the comparative experiments are shown in Table 1.

[0078] Table 1

[0079] The comparison results above show that, in terms of detection accuracy, even after lightweighting, its precision, recall, F1 score, and mean precision still reach 84.3%, 88.7%, 0.86%, and 85.4%, respectively. These figures represent improvements of 1.8%, 6.9%, 0.04%, and 4.1% compared to the high-performance lightweight comparison model YOLOv12n, demonstrating its significant advantage in feature extraction. Regarding inference speed, thanks to the effective reduction of redundant parameters in the model through pruning strategies, ECNP-YOLOv7 can process up to 237 frames per second, ranking first among all mainstream networks compared. In particular, compared to the unpruned ECN-YOLOv7 model, ECNP-YOLOv7 significantly increases the inference frame rate by approximately 30.2% at the cost of only a 1.1% decrease in mean precision, greatly alleviating the computational bottleneck.

[0080] In summary, ECNP-YOLOv7 perfectly balances high precision and extreme lightweight design, demonstrating strong engineering deployment potential and verifying its application value in meeting stringent real-time requirements for ice core and moraine dam detection tasks.

[0081] Meanwhile, to verify the performance of the dual threshold hysteresis switching logic of this application under harsh conditions, the method of this application was used to conduct tests in open glacier environment and extreme canyon environment, respectively. The test results of the UAV hardware performance under normal and harsh conditions are shown in Table 2.

[0082] Table 2

[0083] As shown in Table 2, the effective flight time of the UAV under the load conditions of carrying a computing platform and a visible light camera and continuously performing edge inference is about 45 minutes, which fully meets the requirements of high-altitude and long-distance operations.

[0084] In unobstructed, open glacial areas, the system operates stably in "digital image transmission mode," with the return latency of high-definition OSD enhanced images remaining stable between 150-180 ms, achieving ultra-low latency response in the closed-loop system at the end. When the drone ventures deep into glacial canyons or other harsh conditions, the link signal-to-noise ratio deteriorates drastically. Traditional image transmission systems typically experience disconnections or even crashes lasting several seconds in such environments; however, this solution, with its "adaptive dual-mode link" control strategy, automatically blocks the video stream and switches to "telemetry data mode" within <50 ms when the quality score drops to the trigger threshold. Tests show that even under these extreme conditions, the average packet loss rate of key alarm frames is still controlled at 1.3%, ensuring stable transmission of core early warning information and effectively solving the data interruption problem of traditional solutions in network-free canyon areas.

[0085] In addition, the hardware platform and software environment used to implement the method of this application are described below: 1. Hardware platform: Flight platform: DJI Matrice 400; Airborne edge computing device: DJI Manifold 3; Model training CPU: Intel(R) Xeon(R) Gold 6430; Model training GPU: NVIDIA RTX4090 (24GB).

[0086] 2. Software environment: 2.1 Model Training and Optimization Environment: Operating system: Ubuntu 20.04 LTS; Environment management tool: Anaconda3-4.12.0; Parallel computing architecture: CUDA 12.4 and its corresponding cuDNN acceleration library; Deep learning framework: PyTorch 2.5.1; Basic algorithm dependencies: NumPy, OpenCV, etc.

[0087] 2.2 Airborne Edge Inference and Control Environment: Operating system: Ubuntu 20.04 LTS; Edge hardware acceleration platform: NVIDIA JetPack SDK; Edge inference engine: NVIDIA TensorRT (C++ API); Drone communication control stack: DJI PSDK V3; Visual post-processing and concurrency frameworks: OpenCV library, GStreamer multimedia framework.

[0088] 3. The specific software environment configuration is as follows: Phase 1: Setting up the model training environment.

[0089] System and driver initialization: On a workstation equipped with an Intel Xeon Gold 6430 and an NVIDIA RTX 4090, a bare-metal Ubuntu 20.04 LTS operating system was installed. Subsequently, the matching NVIDIA graphics card driver was compiled and installed for the graphics card architecture. Parallel computing underlying deployment: Configure the CUDA 12.4 runtime environment and inject the corresponding cuDNN deep neural network acceleration primitive library into the system, so that the GPU computing power can be fully utilized by upper-layer applications; Algorithm Framework and Dependency Setup: An isolated Python virtual environment was created using Anaconda3-4.12.0. Within this environment, the PyTorch 2.5.1 deep learning framework was compiled and installed, and image processing and matrix operation dependencies such as NumPy and OpenCV were configured simultaneously to complete the training environment setup.

[0090] Phase 2: Airborne edge environment configuration.

[0091] Edge-based system flashing: Connect the factory-configured DJI Manifold 3 via the host cable, flash the customized Ubuntu 20.04 firmware using NVIDIA SDK Manager, and simultaneously install the NVIDIA JetPackSDK to activate the sparse computing scheduling permissions of the Jetson Orin core. Inference engine and media framework deployment: Configure the NVIDIA TensorRT (C++ API) inference acceleration stack on top of the JetPack environment. At the same time, compile and link the OpenCV source code and GStreamer multimedia framework to establish the underlying video hardware encoding and decoding (NVENC / NVDEC) channel.

[0092] Drone communication stack integration: Import and compile the official DJI PSDK V3 runtime library. Configure the serial port (UART) and high-speed network port (Ethernet) permissions of Manifold 3 to ensure that the application can call the underlying E-Port V2 data interface for communication.

[0093] In summary, this invention provides a method for detecting ice cores and moraines on UAV platforms. This method improves the deep learning network structure to adapt to the computing power of airborne edge devices, while achieving accurate capture of minute topographic features. Furthermore, it utilizes vectorized decision logic and a signal quality-driven adaptive transmission strategy to ensure that key results can still be obtained in real time under extreme communication conditions.

[0094] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for detecting ice core and moraine dams on an unmanned aerial vehicle (UAV) platform, characterized in that, Includes the following steps: S1. Real-time images of the glacial moraine area are acquired using imaging equipment mounted on a drone, and the real-time images are filtered and enhanced. S2. Input the preprocessed image into the ECNP-YOLOv7 detection model deployed on an airborne edge computing device to detect and identify the geomorphic features of ice cores and moraines in the image; the geomorphic features include thermomelt sinkholes, creep folds, mound subsidence, and adjacent snowdrifts; The ECNP-YOLOv7 detection model uses the YOLOv7 network as its basic architecture. The backbone network in the YOLOv7 network model is replaced with the first 7 stages of EfficientNetV2-S as the feature extraction backbone network, and the CBAM attention module is embedded between the backbone network and the head network. S3. Generate the target statistical vector of the current frame by integrating the lightweight statistical parsing module at the output end of the ECNP-YOLOv7 detection model, and perform binarization mapping on the target statistical vector to generate a binarized state vector; then perform weighted fusion on the binarized state vector based on preset asymmetric weights to calculate the glacial moraine dam risk score. S3 includes: S31. A four-dimensional target statistical vector is generated through the lightweight statistical parsing module. ;in, This represents the total number of thermal cavities detected in the current frame; Represents the total number of creep folds detected in the current frame; This represents the total number of hill-shaped depressions detected in the current frame; This represents the total number of neighboring snowdrifts detected in the current frame; S4. The airborne edge computing device monitors the quality of the communication link between the UAV and the ground station in real time, and adaptively selects to transmit the video stream rendered by OSD through the digital image transmission link or to transmit narrowband data containing the detection results through the telemetry data link based on the quality of the communication link. The video stream includes real-time footage, target detection boxes, confidence scores corresponding to the target detection boxes, target statistical vectors, glacial moraine risk scores, and a determination of whether glacial moraine exists based on a comparison of the glacial moraine risk scores with a preset scoring threshold; the narrowband data includes target statistical vectors and glacial moraine risk scores. The method also includes performing structured model pruning during the training phase of the ECNP-YOLOv7 detection model, specifically: Scaling factor of the batch normalization layer after all convolutional layers in the ECNP-YOLOv7 detection model By applying L1 regularization constraints, the modified total loss function is obtained. Its expression is: ； in, The sparsity penalty coefficient; Scaling factor for all batch normalization layers with L1 regularization constraints applied The set of absolute values; The global pruning threshold is calculated based on the target pruning rate and the set of absolute values; the global pruning threshold is the value corresponding to the quantile of the target pruning rate after the set of absolute values ​​is sorted in a preset order. For any convolutional layer, the th i Each channel is marked as retained if its scaling factor absolute value is greater than or equal to the global pruning threshold, otherwise it is marked as discarded, thereby generating a binary pruning mask; For the backbone network, set all channel masks of the input and output layers of the Fused-MBConv module to 1, and perform masking operations only on the Conv3×3 layer; set all masks of all layers of the MBConv module to 1; perform masking operations directly on the head network; and set all masks to 1 for the last convolutional layer of the detection head.

2. The method according to claim 1, characterized in that, The backbone network includes one CBS module, three Fused-MBConv modules, and three MBConv modules cascaded in sequence. The CBS module is used to perform convolution, batch normalization and SiLU activation operations on the preprocessed image in sequence, and output the initial feature map. Each of the Fused-MBConv modules contains a Conv3×3 layer, an SE attention layer, and a Conv1×1 layer connected sequentially from the input to the output. These layers are used to extract spatial features from the feature map through the Conv3×3 layer, then perform channel recalibration through the SE attention layer, and finally perform channel transformation through the Conv1×1 layer to output a fused convolutional feature map. Each MBConv module contains, from input to output, a Conv1×1 dimensionality-upgrading layer, a DepthwiseConv3×3 layer, an SE attention layer, and a Conv1×1 dimensionality-reducing layer. This is used to first upgrade the feature map using Conv1×1, then perform depthwise separable convolution operations using Depthwise Conv3×3, then recalibrate the channel weights using the SE attention layer, and finally reduce the dimensionality using Conv1×1 to output a depthwise convolutional feature map. The backbone network is processed sequentially by the CBS module, three Fused-MBConv modules, and three MBConv modules, and finally outputs a multi-scale input feature map.

3. The method according to claim 1, characterized in that, The CBAM attention module includes a channel attention module and a spatial attention module cascaded in sequence; The channel attention module is used to perform global max pooling and global average pooling on the input feature map respectively. After inputting the two pooling results into the shared neural network, the two feature vectors are fused element by element. Then, the channel attention weights are generated by the Sigmoid activation function. The channel attention weights are multiplied element by element with the input feature map, and the intermediate feature map is output to the spatial attention module. The spatial attention module is used to receive the intermediate feature map, and perform global max pooling and global average pooling on the intermediate feature map respectively. The two pooling results are concatenated along the channel dimension, and then a 7×7 convolutional layer and a sigmoid activation function are used to generate spatial attention weights. The spatial attention weights are then multiplied element-wise with the intermediate feature map to output the final enhanced feature map.

4. The method according to claim 1, characterized in that, The total loss function of the ECNP-YOLOv7 detection model during the training phase includes: bounding box regression loss term, target confidence loss term, and classification loss term; The total loss function is calculated as follows: ； in, For bounding box regression loss term; For target confidence loss term; For classification loss items; , , These are the weight coefficients for the bounding box regression loss term, the target confidence loss term, and the classification loss term, respectively. The bounding box regression loss term uses the normalized Wasserstein distance instead of the original YOLOv7 network's CIoU loss function.

5. The method according to claim 1, characterized in that, The filtering enhancement in S1 employs a bilateral filtering algorithm, the mathematical expression of which is: ； in, The coordinates of the target pixel; Indicates the target pixel point The neighborhood centered on; For the neighborhood The coordinates of any pixel within the range; For point The corresponding pixel value; The pixel value corresponding to the target pixel; For point The corresponding weighting coefficients.

6. The method according to claim 3, characterized in that, The calculation formula for the channel attention module is as follows: ； ； The calculation formula for the spatial attention module is as follows: ； ； in, F Input feature map; This is an intermediate feature map; To enhance the feature map; Channel attention weights; Spatial attention weights; This represents element-wise multiplication; Indicates the Sigmoid activation operation; This indicates a global average pooling operation; This represents the global max pooling operation; Indicates multilayer sensor operation; This represents a 7×7 convolution operation.

7. The method according to claim 4, characterized in that, The process of forming the bounding box regression loss term is as follows: The predicted bounding boxes output by the ECNP-YOLOv7 detection model and the corresponding real frame labels Represented as two-dimensional Gaussian distributions and ;in, , These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. , These are the widths of the predicted bounding box and the ground truth bounding box, respectively; , These are the lengths of the predicted bounding box and the ground truth bounding box, respectively; , These are the mean vectors of the predicted bounding box and the ground truth bounding box, respectively; , These are the covariance matrices of the predicted bounding box and the ground truth bounding box, respectively. Calculate the second-order Wasserstein distance between the two-dimensional Gaussian distributions corresponding to the predicted bounding box and the ground truth bounding box. Its expression is: ； The similarity metric is obtained by exponentially normalizing the second-order Wasserstein distance and mapping it to the [0, 1] interval. Its expression is: ； in, This is the scale normalization factor; Calculate the bounding box regression loss term based on the similarity metric. Its calculation expression is: 。 8. The method according to claim 1, characterized in that, S3 further includes: S32. Binarize each element in the target statistical vector using a non-zero indicator function to generate a binary state vector. When an element is greater than 0, it indicates that the target of the element's class exists, and the mapping is 1; when an element is equal to 0, it indicates that the target of the element's class does not exist, and the mapping is 0. ； in, for The value of an element after binarization mapping; for The value of an element after binarization mapping; for The value of an element after binarization mapping; for The value of an element after binarization mapping; S33. Based on prior knowledge of glaciology, assign risk weight coefficients to each geomorphic feature. Calculate the risk score of glacial moraine dams ;in, The output of the non-zero indicator function.

9. The method according to claim 1, characterized in that, S4 specifically includes: S41. The airborne edge computing device establishes a link status monitoring thread through the UAV SDK, reads the underlying status parameters of the communication module at a fixed frequency, and calculates a comprehensive link quality score based on the Received Signal Strength Indicator (RSSI) and Signal-to-Noise Ratio (SNR). The calculation formula is as follows: ； in, The weighting coefficients for the received signal strength indication; These are the weighting coefficients for the signal-to-noise ratio; S42. Set the upper limit threshold for communication quality. and lower threshold And execute the dual threshold hysteresis switching logic: when At this time, the digital image transmission link is activated to transmit the video stream back, and the narrowband data transmission thread is suspended. when At this time, the telemetry data link is activated to transmit narrowband data back; when When the system starts, the current transmission state remains unchanged; if the initial comprehensive link quality score is in the range specified in the original text, the transmission state will be maintained. and In between, the digital video transmission link is enabled by default to prioritize video data backhaul.

Images