An artificial intelligence-based unmanned aerial vehicle river edge computing method

Abstract

The application provides a kind of unmanned aerial vehicle riverway edge calculation method based on artificial intelligence, belongs to riverway edge detection technical field, the application is extracted basic feature vector by real-time processing riverway image and extracting basic feature vector to unmanned aerial vehicle carrying lightweight basic feature extraction model, ground server calls depth edge detection model after receiving feature vector and outputs riverway edge position matrix and change trend matrix, when detection confidence is lower than 85%, feature enhancement parameter is adjusted by gain amplification matrix to improve detection precision, when edge change amplitude exceeds standard range, start to slow down filter processing to reduce noise interference, based on edge position and change trend, erosion prediction model is constructed to generate early warning signal, feedback optimization mechanism is established to dynamically update granularity adjustment matrix and depth feature mining matrix parameters, solve the technical problems of too large data transmission and insufficient real-time processing precision in unmanned aerial vehicle riverway edge detection.

Description

Technical Field

[0001] This invention belongs to the field of river edge detection technology, and more specifically, relates to an artificial intelligence-based method for calculating river edges using unmanned aerial vehicles (UAVs). Background Technology

[0002] River edge monitoring is a crucial technology in water conservancy projects and environmental protection. Traditional river edge detection primarily relies on ground-based surveying equipment or satellite remote sensing for data collection, extracting river boundary information through manual observation or simple image processing algorithms. It plays a vital role in applications such as river monitoring, flood warning, and soil and water conservation. With the development of unmanned aerial vehicle (UAV) technology, UAV-based river monitoring systems are gradually becoming the mainstream solution, enabling large-scale, high-frequency river condition monitoring. However, traditional UAV-based river monitoring technology has significant drawbacks, mainly in data transmission and processing efficiency. Existing solutions typically require the complete transmission of high-resolution images collected by the UAV to a ground processing center, resulting in massive data transmission volumes and severe latency. Simultaneously, the computational load on the ground processing center is excessive, making it difficult to meet the demands of real-time monitoring. Furthermore, traditional edge detection algorithms have limited accuracy in identifying river boundaries in complex environments and are easily affected by factors such as changes in lighting, water surface reflection, and vegetation obstruction. In existing technologies, UAV river monitoring systems require the transmission of large amounts of raw image data to the ground for processing, resulting in huge data transmission bandwidth requirements and poor real-time performance. At the same time, traditional edge detection algorithms are not accurate enough under complex environmental conditions and are difficult to accurately identify subtle changes in river boundaries. In other words, existing technologies suffer from the technical problems of excessive data transmission and insufficient real-time processing accuracy in UAV river edge detection. Summary of the Invention

[0003] In view of this, the present invention provides an artificial intelligence-based method for calculating the river edge of a drone, which can solve the technical problems of excessive data transmission and insufficient real-time processing accuracy in the existing technology of drone river edge detection.

[0004] This invention is implemented as follows: It provides an AI-based method for calculating river edges using unmanned aerial vehicles (UAVs). The UAV carries a high-resolution camera to collect river image data, obtaining raw river images. A lightweight basic feature extraction model is deployed in the onboard processor to process the raw river images in real time, extracting basic feature vectors. These basic feature vectors are transmitted to a ground server via a wireless link. The ground server calls a deep edge detection model to process the basic feature vectors, outputting a river edge position matrix and a river edge change trend matrix. The detection confidence of the river edge position matrix is ​​calculated. When the detection confidence is lower than a preset threshold, an adjustment command is sent to the UAV. Abnormal fluctuations in the river edge change trend matrix are analyzed. When the change amplitude exceeds the set range of the standard change amplitude matrix, mitigation filtering is initiated. Based on the river edge position matrix and the river edge change trend matrix, a river erosion prediction model is constructed to generate early warning signals.

[0005] The lightweight basic feature extraction model adopts an improved MobileNet architecture, which includes three depthwise separable convolutional layers, two residual connection blocks and one feature fusion layer. The model parameter size is controlled within 8MB. The input layer receives the original river channel image. The first layer uses depthwise separable convolutional kernels to extract edge contour features, the second layer extracts texture detail features, and the third layer performs feature dimensionality reduction through convolution.

[0006] The dimension of the basic feature vector is dynamically adjusted according to the granularity adjustment matrix, which is a weight matrix that controls the fineness of feature extraction. The adjustment process is achieved by modifying the stride parameter of the convolution kernel and the size of the pooling window. When it is necessary to increase the granularity, the convolution stride is reduced, the pooling window is reduced, and the spatial resolution of the feature map is increased.

[0007] The deep edge detection model uses a deep feature mining matrix to control the depth of feature extraction. The deep feature mining matrix is ​​a parameter matrix that controls the number of feature extraction layers and the size of the receptive field in the deep learning model. The deep edge detection model adopts a UNet encoding and decoding structure. The encoder contains four downsampling blocks and the decoder contains four upsampling blocks. The deep feature mining matrix controls the activation threshold and feature weight allocation of each layer.

[0008] The adjustment instruction adjusts the feature enhancement parameters of the lightweight basic feature extraction model through the gain amplification matrix. The gain amplification matrix is ​​a coefficient matrix used to amplify the intensity of the feature signal. The adjustment process involves the scaling and bias parameters of the batch normalization layer in the lightweight basic feature extraction model. When the detection confidence is lower than the threshold, the scaling coefficient of the batch normalization is increased, the bias value is adjusted, and the learning rate of the convolutional layer is increased.

[0009] Among them, the mitigation filtering process reduces the impact of noise interference on edge detection accuracy by mitigation suppression matrix. The mitigation suppression matrix is ​​a filtering matrix used to suppress image noise and abnormal fluctuations. The mitigation filtering process is achieved by adding a Gaussian filter during the feature extraction process. The Gaussian filter kernel size is standard size, and the standard deviation parameter is adjusted according to the element values ​​of the mitigation suppression matrix, while reducing the weight coefficient of high-frequency features.

[0010] The standard variation range matrix is ​​a reference matrix obtained by statistical analysis of historical river channel change data. Each matrix element records the normal variation range at the corresponding spatial location. The standard variation range matrix is ​​established using collected river channel monitoring data, including edge change records under different water level conditions. The mean and standard deviation of the variation at each location are obtained through statistical analysis. The normal range is set as the interval between the mean and two standard deviations above and below the mean.

[0011] Among them, the river edge position matrix is ​​a numerical matrix that describes the precise position of the river boundary line in a two-dimensional coordinate system. The matrix element values ​​represent the coordinate information of the edge points, and the matrix size corresponds to the resolution of the original river image. The river edge change trend matrix is ​​a prediction matrix that describes the change law of the river boundary line over time. The matrix element values ​​represent the speed and direction of edge change at different locations, and are calculated by the river edge position matrix of multiple consecutive frames.

[0012] Among them, the river erosion prediction model is a prediction algorithm based on time series analysis. It uses the river edge location matrix and river edge change trend matrix in a continuous time period as input, and outputs the predicted location of the river edge in the future time period. When the predicted change exceeds the safe range, an early warning signal is generated. The early warning signal includes the coordinates of the changed location, the change magnitude, and the expected time of occurrence.

[0013] This also includes: establishing a feedback optimization mechanism to dynamically update parameter values; using a separate architecture for airborne lightweight feature extraction and ground depth processing to avoid transmitting raw image data and significantly reduce data transmission volume; and using an adaptive parameter adjustment mechanism to dynamically optimize feature extraction accuracy based on detection results.

[0014] Furthermore, the feedback optimization mechanism verifies the accuracy of edge detection results based on ground-based measured data, dynamically updates the parameter values ​​of the granularity adjustment matrix and the deep feature mining matrix, continuously optimizes the performance of the entire detection system, and adjusts various parameter matrices by comparing the deviation between measured data and detection results, thereby achieving self-learning and performance improvement of the system.

[0015] The original river channel image has a resolution of 1024×768 pixels. The lightweight basic feature extraction model ultimately outputs a 256-dimensional basic feature vector. The training dataset contains 30,000 river channel images under different seasons and lighting conditions. Manually labeled edge positions are used as supervision signals, and the gradient descent algorithm is used for training for 300 cycles.

[0016] The granularity adjustment matrix is ​​a 32×32 weight matrix. Each element in the granularity adjustment matrix corresponds to the sensitivity weight of a local region. The weight values ​​are adjusted within a preset range. The deep feature mining matrix is ​​a 16×16 parameter matrix. By adjusting the element values ​​of the deep feature mining matrix, the effective depth of the network can be changed within a preset layer range.

[0017] The gain amplification matrix is ​​a 64×64 coefficient matrix. The element values ​​of the gain amplification matrix control the amplification factor of each feature channel. The amplification factor range is adjusted within a preset interval. The mitigation and suppression matrix is ​​a 128×128 filter matrix. The Gaussian filter kernel size is 5×5. At the same time, a moving average window is applied to the time series for smoothing.

[0018] The standard variation range matrix is ​​a 100×100 reference matrix. The standard variation range matrix is ​​established using river monitoring data collected over the past 5 years. The preset threshold for the detection confidence level is 85%. When the detection confidence level is lower than 85%, an adjustment command is sent to the UAV to adjust the feature enhancement parameters of the lightweight basic feature extraction model.

[0019] The first layer uses a 3×3 depthwise separable convolution kernel to extract edge contour features, the second layer uses a 5×5 convolution kernel to extract texture detail features, and the third layer uses a 1×1 convolution to reduce feature dimensionality. Each downsampling block contains two 3×3 convolutional layers and one 2×2 max pooling layer, and each upsampling block contains a deconvolutional layer and skip connections.

[0020] Specifically, when it is necessary to increase the granularity, the convolution stride is reduced from 2 to 1, the pooling window is reduced from 4×4 to 2×2, the spatial resolution of the feature map is increased from 64×64 to 128×128, the moving average window size is increased from 3 frames to 7 frames for smoothing, the weight values ​​are adjusted within the range, and the magnification factor is adjusted within the preset range.

[0021] This invention significantly reduces data transmission volume and improves the system's real-time response capability by deploying a lightweight basic feature extraction model on the UAV end, transmitting only basic feature vectors instead of the original image. Simultaneously, combined with a ground-based deep edge detection model, it achieves high-precision river edge recognition. This invention employs a hierarchical processing architecture: the lightweight model on the UAV end is responsible for extracting basic feature information, while the ground-based deep model handles precise edge detection. This design effectively solves the problem of large data transmission volume in traditional technologies. Through parameter control mechanisms such as granularity adjustment matrices and deep feature mining matrices, the system can dynamically adjust the feature extraction accuracy based on detection confidence. When the detection effect is poor, the processing granularity is automatically increased; when environmental noise is high, a mitigation filtering process is initiated, ensuring high detection accuracy under various complex environmental conditions. In summary, this invention, through an edge computing architecture and adaptive parameter adjustment mechanisms, solves the technical problems of excessive data transmission volume and insufficient real-time processing accuracy in existing UAV river edge detection technologies. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a graph showing the change in confidence level of river edge detection over time in Example 2.

[0024] Figure 3 This is a prediction and analysis diagram of the riverbed edge change trend in Example 2. Detailed Implementation

[0025] 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.

[0026] like Figure 1 The diagram shown is a flowchart of an artificial intelligence-based method for calculating river edges using unmanned aerial vehicles (UAVs), provided by this invention. This method includes the following steps:

[0027] S01. The UAV is equipped with a high-resolution camera to collect river image data and obtain an original river image with a resolution of 1024×768 pixels. At the same time, a lightweight basic feature extraction model is deployed in the airborne processor to process the original river image in real time and extract a basic feature vector containing the shape information of the river edge. The dimension of the basic feature vector is dynamically adjusted according to the granularity adjustment matrix.

[0028] S02. The basic feature vector is transmitted to the ground server via a wireless transmission link, avoiding the transmission of the original river channel image, which significantly reduces the amount of data transmitted and the transmission delay time.

[0029] S03. After receiving the basic feature vector, the ground server calls the pre-trained deep edge detection model for processing. The deep edge detection model uses a deep feature mining matrix to control the depth of feature extraction and outputs the river edge position matrix and the river edge change trend matrix.

[0030] S04. Calculate the detection confidence of the river edge position matrix. When the detection confidence is lower than a preset threshold of 85%, send an adjustment command to the UAV to adjust the feature enhancement parameters of the lightweight basic feature extraction model through the gain amplification matrix to improve the detail granularity of feature extraction.

[0031] S05. Analyze the abnormal fluctuations in the river edge change trend matrix. When the change amplitude exceeds the set range of the standard change amplitude matrix, start the mitigation filtering process to reduce the impact of noise interference on the edge detection accuracy through the mitigation suppression matrix.

[0032] S06. Construct a river erosion prediction model based on the river edge location matrix and the river edge change trend matrix, calculate the edge change amplitude in the future time period, and generate an early warning signal when the predicted change amplitude exceeds the safe range.

[0033] S07. Optionally, it also includes establishing a feedback optimization mechanism to verify the accuracy of the edge detection results based on ground-based measured data, dynamically update the parameter values ​​of the granularity adjustment matrix and the depth feature mining matrix, and continuously optimize the performance of the entire detection system.

[0034] The lightweight basic feature extraction model adopts an improved MobileNet architecture, which includes three depthwise separable convolutional layers, two residual connection blocks, and one feature fusion layer. The model parameter size is controlled within 8MB. The model structure is as follows: the input layer receives the original river channel image; the first layer uses a 3×3 depthwise separable convolutional kernel to extract edge contour features; the second layer uses a 5×5 convolutional kernel to extract texture detail features; and the third layer performs feature dimensionality reduction through a 1×1 convolution, finally outputting the 256-dimensional basic feature vector. The training dataset contains 30,000 river channel images under different seasons and lighting conditions, and manually labeled edge positions are used as supervision signals. The gradient descent algorithm is used for training for 300 epochs.

[0035] The granularity adjustment matrix is ​​a 32×32 weight matrix that controls the fineness of feature extraction. The adjustment process is achieved by modifying the stride parameter of the convolution kernel and the size of the pooling window. When it is necessary to increase the granularity, the convolution stride is reduced from 2 to 1, the pooling window is reduced from 4×4 to 2×2, and the spatial resolution of the feature map is increased from 64×64 to 128×128. Each element in the granularity adjustment matrix corresponds to the sensitivity weight of a local region, and the weight value is adjusted within the range [0.1, 2.0].

[0036] The deep feature mining matrix is ​​a 16×16 parameter matrix that controls the number of feature extraction layers and the receptive field size of the deep learning model. The deep edge detection model adopts a UNet encoder-decoder structure. The encoder contains four downsampling blocks, each containing two 3×3 convolutional layers and one 2×2 max pooling layer. The decoder contains four upsampling blocks, each containing a deconvolutional layer and a skip connection. The deep feature mining matrix controls the activation threshold and feature weight allocation of each layer. By adjusting the element values ​​of the deep feature mining matrix, the effective depth of the network can be changed within the range of [12, 20] layers.

[0037] The gain amplification matrix is ​​a 64×64 coefficient matrix used to amplify the intensity of the feature signal. The adjustment process involves the scaling and bias parameters of the batch normalization layer in the lightweight basic feature extraction model. When the detection confidence is lower than 85%, the scaling coefficient of the batch normalization is increased from 1.0 to 1.5, the bias value is adjusted from 0.0 to 0.2, and the learning rate of the convolutional layer is increased from 0.001 to 0.003. The element values ​​of the gain amplification matrix control the amplification factor of each feature channel, and the amplification factor is adjusted within the range [1.0, 3.0].

[0038] The mitigation and suppression matrix is ​​a 128×128 filter matrix used to suppress image noise and abnormal fluctuations. The mitigation filtering process is achieved by adding a Gaussian filter during feature extraction. The Gaussian filter kernel size is 5×5, and the standard deviation parameter is adjusted within the range [0.5, 2.0] according to the element values ​​of the mitigation and suppression matrix. At the same time, the weight coefficient of high-frequency features is reduced from 1.0 to 0.3, and a moving average window is applied to the time series. The size of the moving average window is increased from 3 frames to 7 frames for smoothing.

[0039] The standard variation range matrix is ​​a reference matrix obtained by statistical analysis of historical river channel change data. The matrix size is 100×100, and each matrix element records the normal variation range at the corresponding spatial location. The standard variation range matrix is ​​established using river channel monitoring data collected over the past 5 years, including edge change records under different water level conditions. The mean and standard deviation of the variation at each location are obtained through statistical analysis, and the normal range is set as the interval between the mean and two standard deviations above and below the mean.

[0040] The river edge position matrix is ​​a numerical matrix that describes the precise position of the river boundary line in a two-dimensional coordinate system. The matrix element values ​​represent the coordinate information of the edge points, and the size of the matrix corresponds to the resolution of the original river image.

[0041] The riverbed edge change trend matrix is ​​a prediction matrix that describes the change pattern of the riverbed boundary line over time. The matrix element values ​​represent the speed and direction of edge change at different locations and are calculated from the riverbed edge position matrix of multiple consecutive frames.

[0042] The river erosion prediction model is a prediction algorithm based on time series analysis. It uses the river edge location matrix and the river edge change trend matrix over a continuous time period as input to output the predicted location of the river edge in the future time period.

[0043] The early warning signal is an alarm message generated when the predicted changes in the riverbed edge exceed a preset safety threshold. It includes the coordinates of the change location, the magnitude of the change, and the expected time of occurrence.

[0044] The specific implementation methods of the above steps are described in detail below.

[0045] The specific implementation of step S01 involves continuously acquiring river channel image data using a high-resolution camera mounted on a UAV. The image sensor is set to a resolution of 1024×768 pixels, and the frame rate is maintained at 30fps to ensure data continuity. The onboard processor first preprocesses the raw river channel images, including image denoising, brightness normalization, and contrast enhancement. The denoising process uses a median filtering algorithm to remove salt-and-pepper noise, with a filter window size of 3×3. Brightness normalization maps pixel values ​​to the range of 0 to 255, and contrast enhancement uses a histogram equalization algorithm to improve image clarity. The preprocessed image is then input into a lightweight basic feature extraction model. This model is based on an improved MobileNet architecture and reduces computational complexity through depthwise separable convolution. During feature extraction, a granularity adjustment matrix dynamically controls the dimension of the feature vectors, initially set to 256 dimensions, which can be expanded to 512 dimensions when detection accuracy requirements increase. The purpose of this step is to achieve efficient feature extraction with the limited computing resources of the UAV, laying the foundation for subsequent transmission and processing.

[0046] The specific implementation of step S02 involves establishing a wireless transmission link between the UAV and the ground server, using 5G or WiFi 6 protocols to ensure transmission stability and speed. Before transmission, the basic feature vectors are compressed using a lossless compression algorithm, reducing the data size of the 256-dimensional feature vectors from the original 1KB to approximately 300 bytes. The transmission protocol uses TCP to ensure data integrity, and a transmission priority queue is set up to prioritize feature vector data for real-time performance. Timestamps and sequence numbers are added during data encapsulation to facilitate data synchronization and integrity verification by the ground server. Network latency is monitored during transmission, and a data caching mechanism is activated when the latency exceeds 100 milliseconds to prevent data loss. The purpose of this step is to significantly reduce the amount of data transmitted, reducing the original image data size from approximately 2.4MB to 300 bytes, improving transmission efficiency by approximately 8000 times, while ensuring that the critical information required for river edge detection is not lost.

[0047] The specific implementation of step S03 involves the ground server receiving the basic feature vector and first performing data decompression and integrity verification. After successful verification, the data is input into a pre-trained deep edge detection model. This model employs a UNet encoder-decoder architecture. The encoder extracts multi-scale features progressively through four downsampling blocks, each containing a batch normalization layer, an activation function layer, and a pooling layer. The decoder reconstructs a high-resolution feature map through four upsampling blocks. A skip connection mechanism directly passes the encoder's features to the corresponding decoder layer, preserving detailed information. A deep feature mining matrix controls the effective depth of the network, optimizing feature extraction by adjusting the activation threshold and feature weight allocation. The model output includes a river edge position matrix and a river edge change trend matrix. The position matrix records the pixel coordinates of edge points, and the trend matrix calculates the edge change rate through time series analysis. The purpose of this step is to leverage the powerful computing capabilities of the ground server for precise edge detection analysis, compensating for the insufficient computing resources of the UAV.

[0048] The specific implementation of step S04 involves calculating the detection confidence score of the river channel edge position matrix and using a probabilistic statistical method to evaluate the reliability of the detection results. The confidence score calculation is based on three indicators: gradient strength, continuity, and consistency of the edge points. Gradient strength reflects the clarity of the edge, continuity measures the integrity of the edge line, and consistency assesses the stability of the edge positions between adjacent frames. The weights of the three indicators are set to 0.4, 0.3, and 0.3, respectively, and a weighted average is used to obtain the comprehensive confidence score. When the confidence score is lower than a preset threshold of 85%, the system automatically sends an adjustment command to the UAV, which includes the parameter update values ​​of the gain amplification matrix. The adjustment process is achieved by modifying the batch normalization parameters in the lightweight basic feature extraction model, increasing the scaling factor from 1.0 to 1.5 and adjusting the bias value from 0.0 to 0.2. Simultaneously, the learning rate of the convolutional layer is increased from 0.001 to 0.003 to enhance the sensitivity of feature extraction. The purpose of this step is to establish a closed-loop feedback mechanism, adaptively adjusting parameters according to the detection effect to ensure that the system maintains high-precision detection under various environmental conditions.

[0049] The specific implementation of step S05 involves analyzing abnormal fluctuations in the river channel edge change trend matrix and using a statistical anomaly detection algorithm to identify changes exceeding the normal range. Anomaly detection is based on a sliding window statistical method, with a window size set to 10 frames. The mean and standard deviation of edge changes within the window are calculated. When the change amplitude at a certain location exceeds the mean plus twice the standard deviation, it is determined to be an abnormal fluctuation. A standard change amplitude matrix provides a reference benchmark, established based on 5 years of historical data, recording the normal change range at different locations. After anomaly detection, a mitigation filtering process is initiated, using a Gaussian filter during feature extraction to suppress noise. The Gaussian filter kernel size is set to 5×5, and the standard deviation parameter is adjusted within the range of 0.5 to 2.0 based on the element values ​​of the mitigation suppression matrix. Simultaneously, the weight coefficient of high-frequency features is reduced from 1.0 to 0.3 to reduce noise interference. A moving average window is applied to smooth the time series, with the window size increased from 3 frames to 7 frames. The purpose of this step is to improve the robustness of edge detection and reduce the impact of environmental noise and sudden interference on detection accuracy.

[0050] Step S06 is implemented by constructing a time-series prediction model based on the riverbed edge location matrix and edge change trend matrix, and using a Long Short-Term Memory (LSTM) network algorithm to predict river erosion. The model input consists of a location matrix and a trend matrix over a continuous time period, with a time window length of 30 days and data collected four times daily. The network structure includes an input layer, three LSTM hidden layers, and an output layer, with the number of neurons in the hidden layers set to 128, 64, and 32, respectively. The training process uses a mean squared error loss function, the optimizer uses the Adam algorithm, and the learning rate is set to 0.001. The model outputs predicted edge locations for the next 7 days, with prediction accuracy exceeding 90% as verified by historical data. An early warning mechanism assesses risk based on the prediction results, generating an early warning signal when the predicted edge change exceeds a safety threshold of 3 meters. The warning signal includes the specific coordinates of the changed location, the expected change magnitude, and the time of occurrence, facilitating timely protective measures by relevant departments. The purpose of this step is to achieve early warning of river changes, providing a scientific basis for river management and disaster prevention and mitigation.

[0051] Step S07 is optional. Its specific implementation involves establishing a feedback mechanism for continuous system performance optimization and periodically collecting ground-based measurement data to verify the accuracy of edge detection results. The verification process uses manual field measurements and high-precision GPS equipment to obtain the true location coordinates of the riverbed edge, with measurement accuracy controlled at the centimeter level. The measured data is compared and analyzed with the system detection results to calculate indicators such as location error, trend prediction error, and time delay. Based on the error analysis results, a gradient descent algorithm is used to dynamically update the parameter values ​​of the granularity adjustment matrix and the deep feature mining matrix. A learning rate decay mechanism is set during the update process, with an initial learning rate of 0.01, decaying by 10% every 100 iterations to prevent over-updating of parameters. Simultaneously, a parameter version management mechanism is established to save historically optimal parameter configurations, allowing for rapid rollback when new parameters degrade performance. The optimization frequency is set to once a month to ensure the system can adapt to seasonal and environmental changes. The purpose of this step is to establish self-learning and continuous improvement capabilities, ensuring long-term stable operation and maintaining optimal performance.

[0052] The detailed structure of the lightweight basic feature extraction model includes an input layer that receives a raw river channel image of 1024×768 pixels. The first layer uses a 3×3 depthwise separable convolution kernel for edge contour feature extraction, with an output feature map size of 512×384 and 32 channels. Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. The depthwise convolution performs spatial filtering for each input channel individually, while the pointwise convolution fuses information between channels using a 1×1 convolution kernel, reducing computational complexity by 8 times compared to standard convolution. The second layer uses a 5×5 convolution kernel to extract texture detail features and employs dilated convolution to expand the receptive field, resulting in an output feature map size of 256×192 and an increased number of channels to 64. The third layer uses a 1×1 convolution for feature dimensionality reduction and channel fusion, with an output feature map size of 128×96 and 128 channels. Two residual connection blocks are located after the first and second layers, respectively, employing a bottleneck structure to reduce the number of parameters. Each residual block contains a 1×1 dimensionality-reduced convolution, a 3×3 depthwise convolution, and a 1×1 dimensionality-upgrading convolution. The feature fusion layer adaptively weights and fuses multi-scale features, ultimately outputting a 256-dimensional basic feature vector. The training dataset was built by collecting 30,000 river images under different seasons and lighting conditions, covering four typical periods: spring (high water season), summer (flood season), autumn (dry season), and winter (freezing season). Industrial-grade cameras with a resolution of at least 1024×768 were used for image acquisition to ensure stable image quality. Manual annotation was performed by experienced hydraulic engineers, using image annotation tools to accurately mark the riverbed edges with pixel-level precision. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio, and data augmentation techniques, including rotation, flipping, scaling, and brightness adjustment, were used to expand the training samples to 50,000 images. The training process uses the gradient descent algorithm, with a batch size of 32, 300 training cycles, an initial learning rate of 0.001, and a decay of 20% every 50 cycles.

[0053] The detailed structure of the deep edge detection model is based on the UNet encoder-decoder architecture. The encoder contains four downsampling blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, a ReLU activation function, and a 2×2 max-pooling layer. The first downsampling block has 256 input channels and 64 output channels, with the feature map size gradually decreasing from the original input to 1 / 16 of its original size. The decoder contains four upsampling blocks, each consisting of a transposed convolutional layer, a concatenation layer, and two 3×3 convolutional layers. Skip connections directly pass the feature maps of the corresponding layers in the encoder to the decoder, preserving spatial detail. The network has a total depth of 20 layers and contains approximately 23 million trainable parameters. The loss function uses a weighted combination of binary cross-entropy and Dice coefficients with a weight ratio of 1:1, and the optimizer uses the Adam algorithm. The training dataset is based on historical river monitoring data, collecting 100,000 river images under different hydrological conditions over five years, including corresponding edge annotation information. Data preprocessing includes image size normalization, pixel value normalization, and edge label binarization. An early stopping strategy is employed during training to prevent overfitting; training stops when the validation set loss does not decrease for 10 consecutive epochs. The model achieves edge detection accuracy exceeding 95% on the independent test set, meeting the requirements for practical applications.

[0054] It should be noted that the key technical ideas of this invention include an edge-cloud collaborative computing architecture design, an adaptive parameter adjustment mechanism, and a multimodal feature fusion strategy. The edge-cloud collaborative computing architecture optimizes the allocation of computing resources by deploying lightweight feature extraction on the UAV and complex edge detection analysis on the ground server. Compared to traditional pure cloud processing methods, this architecture significantly reduces data transmission volume, mitigates the impact of network latency on real-time performance, and avoids the accuracy issues caused by resource limitations in pure edge computing. The adaptive parameter adjustment mechanism, based on feedback control of detection confidence, dynamically optimizes model parameters according to environmental changes and detection results. Compared to traditional methods with fixed parameters, this mechanism improves the system's adaptability to complex environments, ensuring stable detection performance under different lighting, weather, and hydrological conditions. The multimodal feature fusion strategy combines spatial, temporal, and statistical features, achieving multi-level, multi-scale feature expression through the coordinated control of granularity adjustment matrices and deep feature mining matrices. Compared to single feature extraction methods, this strategy enhances the discriminative power and robustness of features. The synergistic effect of these three key technical ideas forms a complete intelligent river monitoring solution. The edge-cloud collaborative architecture ensures the system's real-time performance and scalability, the adaptive adjustment mechanism provides dynamic optimization capabilities, and the multimodal fusion strategy ensures the accuracy and stability of detection. These three elements complement and promote each other, enabling the system to achieve high-precision and high-efficiency river edge detection and change prediction in complex and ever-changing natural environments, providing solid technical support for the intelligent and automated management of rivers.

[0055] It should be noted that this invention also solves the following technical problem: the lack of adaptive environmental change capability in existing river edge detection systems. Traditional river monitoring systems typically use fixed detection parameters and algorithm models, which cannot automatically adjust detection strategies according to different environmental conditions. This leads to a significant decrease in detection accuracy under the influence of environmental factors such as changes in light intensity, seasonal changes, and weather conditions. In particular, under special environmental conditions such as drastic water level changes during the rainy season, freezing in winter, and lush vegetation in summer, traditional systems often struggle to maintain stable detection results.

[0056] This invention effectively solves this technical problem by establishing a multi-level parameter adjustment matrix system. The granularity adjustment matrix dynamically adjusts the convolution kernel stride and pooling window size according to detection requirements, expanding the feature map spatial resolution from 64×64 to 128×128 when fine detection is required. The deep feature mining matrix adapts to different complexity scenarios by controlling the effective network depth to vary within the range of 12 to 20 layers. The gain amplification matrix automatically increases the feature signal intensity when the detection confidence is insufficient. The mitigation and suppression matrix suppresses environmental noise by adjusting the standard deviation parameter of the Gaussian filter and the moving average window size. This multi-dimensional adaptive adjustment mechanism enables the system to automatically optimize detection parameters according to actual environmental conditions, ensuring high detection accuracy and stability in various complex environments.

[0057] Furthermore, this invention addresses the technical problem of the lack of long-term trend analysis capabilities in river channel change prediction. Existing river monitoring technologies primarily focus on detecting the current state, lacking the ability to analyze and predict long-term river channel evolution trends, making it difficult to provide forward-looking technical support for water conservancy project planning and flood control decisions. This invention establishes a prediction algorithm based on time series analysis through a river erosion prediction model. Using the river edge location matrix and change trend matrix over a continuous time period as input, combined with historical benchmark data established by a standard change amplitude matrix, it can predict the location and magnitude of changes in the river edge over future time periods. When the predicted changes exceed a safety threshold, it promptly generates early warning signals, providing important decision-making basis for flood control early warning and water conservancy project maintenance, significantly improving the predictive analysis capabilities and practical value of the river monitoring system.

[0058] Specifically, the principle of this invention is as follows: The fundamental principle that enables this invention to solve the problems of the prior art lies in the distributed architecture design of edge computing and cloud collaborative processing. By deploying a lightweight basic feature extraction model on the drone, the high-resolution image data that originally needed to be transmitted is converted into low-dimensional feature vectors, thereby significantly reducing the amount of data transmitted. This design makes full use of the computing power of the drone's onboard processor and avoids redundant transmission of the original image data in traditional solutions. At the same time, the lightweight model adopts an improved MobileNet architecture, which controls the model parameter size to within 8MB while ensuring the feature extraction effect, thus ensuring efficient operation under the limited computing resources of the drone. The technical solution of this invention possesses a logically complete adaptive optimization mechanism. It dynamically controls the fineness of feature extraction through a granularity adjustment matrix. When the ground-end detection confidence level falls below a preset threshold, the system automatically sends adjustment commands to the UAV, increasing the feature enhancement parameters through a gain amplification matrix and enhancing the detail granularity of feature extraction. This closed-loop feedback mechanism ensures that the system maintains optimal detection performance under different environmental conditions. Simultaneously, a deep feature mining matrix controls the feature extraction level and receptive field size of the ground-end deep edge detection model, allowing the model to adjust the effective depth of the network according to specific detection needs. In complex scenarios, increasing network depth extracts richer feature information, while in simple scenarios, reducing computational complexity improves processing efficiency. Furthermore, this invention establishes a comprehensive noise suppression and anomaly detection mechanism. A change benchmark based on historical data is established through a standard change amplitude matrix. When abnormal fluctuations are detected, a mitigation and suppression matrix is ​​activated for filtering, effectively reducing the impact of environmental noise on edge detection accuracy. The river erosion prediction model, based on a time series analysis algorithm, can predict the trend of river edge changes in future time periods, providing crucial decision support for flood prevention and early warning. The entire system continuously updates various parameter matrices through a feedback optimization mechanism, achieving continuous improvement and optimization of system performance.

[0059] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0060] The specific implementation of step S01 involves acquiring river channel image data using a high-resolution camera mounted on a drone, and then processing the original river channel images in real time using a lightweight basic feature extraction model. The convolution operation function is defined as follows:

[0061] ;

[0062] In the formula, This is the convolution operation function; Input image; For convolution kernel; Output position coordinates; These are the offset coordinates of the convolution kernel.

[0063] The dynamic adjustment process of the granularity adjustment matrix is ​​specifically represented as follows:

[0064] ;

[0065] In the formula, The basic feature vector of the output; This is a granularity adjustment matrix with a size of 32×32; The input is the original river channel image; These are the adjusted convolution kernel parameters; This is for the Hadamard product operation.

[0066] The formula for calculating the elements of the granularity adjustment matrix is:

[0067] ;

[0068] In the formula, For the granularity adjustment matrix, the first... Line number The element values ​​of the column; The step size adjustment parameter for the corresponding position; Adjust the pooling window parameters for the corresponding positions; These are the weighting coefficients, with default values ​​of 0.6, 0.3, and 0.1 respectively.

[0069] The parameter acquisition method is as follows: The calculation steps, obtained by analyzing the local complexity of the image, include: first, calculating the local gradient intensity, and then determining the step size requirement based on the gradient distribution. The pooling window size is determined by statistically analyzing the number of local feature points.

[0070] The specific implementation of step S02 is the same as described above, and will not be repeated in detail here.

[0071] The specific implementation of step S03 involves the ground server calling the deep edge detection model for processing, with the deep feature mining matrix controlling the depth of feature extraction. The calculation process of the deep feature mining matrix is ​​specifically represented as follows:

[0072] ;

[0073] In the formula, For the first Layer feature output; For the first The elements of the deep feature mining matrix corresponding to the layer; For activation functions; For the first Layer weight matrix; The feature input for the previous layer; For the first The bias vector of the layer.

[0074] The formula for calculating the river channel edge location matrix is:

[0075] ;

[0076] In the formula, This is a matrix representing the positions of the riverbed edges; For the first The coordinates of the edge points; This represents the total number of edge points detected.

[0077] The formula for calculating the trend matrix of river channel edge changes is:

[0078] ;

[0079] In the formula, For position Edge change trend value; For the current moment In position The edge position; For the previous moment At the edge of the same location; This represents the inter-frame time interval.

[0080] The parameter acquisition method is as follows: The value was obtained through adaptive learning during the network training process, with the initial value set to 1.0. and Obtained through training using the backpropagation algorithm. The frequency of data collection is determined by the drone, with a typical value of 0.1 to 1 second.

[0081] The specific implementation of step S04 involves calculating the detection confidence level of the river channel edge location matrix, and adjusting parameters when the confidence level is lower than a preset threshold. The formula for calculating the detection confidence level is as follows:

[0082] ;

[0083] In the formula, To comprehensively test the confidence level; This is an index of gradient strength at edge points; As an indicator of edge continuity; This serves as a consistency indicator between adjacent frames. These are weighting coefficients, which are 0.4, 0.3, and 0.3, respectively.

[0084] The adjustment formula for the gain amplification matrix is:

[0085] ;

[0086] In the formula, For the gain amplification matrix, the first Line number Column elements; This is the base gain coefficient, with a default value of 1.0. To adjust the sensitivity parameter, the default value is 0.8; The confidence threshold is set to 0.85.

[0087] The parameter acquisition method is as follows: Image gradients are obtained by calculating the Sobel operator. It is obtained by analyzing the spatial distance between edge points. It is obtained by calculating the Euclidean distance between the edge positions of adjacent frames.

[0088] The specific implementation of step S05 involves analyzing abnormal fluctuations in the river channel edge change trend matrix. When the change amplitude exceeds the set range of the standard change amplitude matrix, mitigation filtering is initiated. The formula for calculating the standard change amplitude matrix is:

[0089] ;

[0090] In the formula, For position The range of standard variations at the location; This represents the average change in historical data at this location. This represents the standard deviation of the historical data at this position.

[0091] The specific formula for calculating abnormal fluctuation detection is as follows:

[0092] ;

[0093] In the formula, For position The degree of abnormality at the location; Position of the current frame The change value at; The mean value within the sliding window; This represents the standard deviation within the sliding window.

[0094] The Gaussian filtering formula for mitigating the suppression matrix is:

[0095] ;

[0096] In the formula, This is the filtered result; This is the original data; It is a Gaussian filter kernel; The filter radius is set to 2.

[0097] The formula for calculating the Gaussian filter kernel is:

[0098] ;

[0099] In the formula, This is the spatial offset of the filter kernel; The Gaussian standard deviation parameter is adjusted within the range of 0.5 to 2.0 based on the element values ​​of the mitigation and suppression matrix.

[0100] The parameter acquisition method is as follows: and It was obtained by statistical calculation of 10 frames of historical data. and This was obtained through statistical analysis of five years of historical monitoring data.

[0101] The specific implementation of step S06 involves constructing a river erosion prediction model based on the riverbed edge location matrix and the riverbed edge change trend matrix. The calculation formula for the time series prediction model is specifically expressed as follows:

[0102] ;

[0103] In the formula, The predicted future location of the riverbank; Input the multidimensional features at the current time. For Long Short-Term Memory (LSTM) network mapping functions; For the prediction time interval; This is the prediction error term.

[0104] The formula for calculating the magnitude of change is:

[0105] ;

[0106] In the formula, To predict the magnitude of change; This is the predicted vector of the riverbed edge location; This is the current position vector of the riverbed edge; It is the Euclidean norm.

[0107] The parameter acquisition method is as follows: Extract the location matrix and trend matrix obtained through the aforementioned steps. The feature dimension is set to 256. The range was obtained through historical prediction error statistics, with a typical value of 0.1 to 0.3 meters. Based on the predicted demand, the typical value ranges from 1 hour to 7 days.

[0108] The specific implementation of step S07 involves establishing a feedback optimization mechanism to verify the accuracy of the edge detection results based on ground-based measured data. The gradient descent algorithm for parameter optimization is specifically represented as follows:

[0109] ;

[0110] In the formula, The updated parameter vector; This is the current parameter vector; The learning rate; This represents the gradient of the loss function with respect to the parameters. This is the loss function.

[0111] The formula for calculating the loss function is:

[0112] ;

[0113] In the formula, To verify the sample size; For the first The test results of each sample; For the first The ground truth value of a sample; For regularization terms; This is the regularization coefficient, set to 0.01.

[0114] The parameter acquisition method is as follows: Data was obtained through on-site measurements using high-precision GPS equipment, with measurement accuracy controlled at the centimeter level. An adaptive decay strategy is adopted, with an initial value of 0.01 and a decay of 10% every 100 iterations.

[0115] Further explanation is needed regarding the principle behind the granularity adjustment matrix calculation formula, which is based on adaptive feature extraction theory. It achieves feature extraction under different accuracy requirements by dynamically adjusting convolution parameters. This formula considers three core factors: stride adjustment, pooling adjustment, and baseline bias. The stride adjustment term controls the spatial sampling density, the pooling adjustment term controls the degree of feature aggregation, and the baseline bias term provides a stable benchmark value. Compared to traditional methods with fixed parameters, this formula achieves dynamic parameter optimization, significantly improving the accuracy and robustness of feature extraction in complex environments.

[0116] The principle behind the deep feature mining matrix operation formula is based on the attention mechanism in deep learning, achieving selective enhancement of features at different layers through element-wise multiplication. This formula uses the depth mining matrix as a weighting factor to weight and adjust the feature activation of each layer, thereby controlling the effective depth and feature representation capability of the network. Compared to fixed-depth network structures, this formula provides dynamic depth adjustment capability, improving edge detection accuracy while maintaining computational efficiency.

[0117] The principle behind the confidence score calculation formula is based on a multi-indicator fusion reliability assessment theory, comprehensively considering three key indicators: gradient strength, continuity, and consistency. The gradient strength term reflects the clarity of the edge, the continuity term measures the integrity of the edge line, and the consistency term assesses temporal stability. The weighted combination of these three indicators provides a comprehensive quality assessment. Compared to single-indicator assessments, this formula significantly improves the accuracy and reliability of confidence score calculation.

[0118] The principle behind the formula for calculating the trend matrix of river channel edge changes is based on the finite difference method, which obtains the rate of change by calculating the difference between the edge positions at adjacent times:

[0119] ;

[0120] This formula discretizes the continuous time derivative into a difference form between adjacent frames, which can effectively capture the dynamic changes of the river edge.

[0121] The principle behind the formula for calculating the standard variation range matrix is ​​based on the normal distribution assumption in statistics, and uses the two-standard-deviation criterion to establish the normal variation range:

[0122] ;

[0123] This formula ensures that 95% of normal variations are within the set range, providing a reliable statistical benchmark for anomaly detection.

[0124] The principle behind the abnormal fluctuation detection formula is based on the standardized anomaly detection theory in statistics. It identifies anomalies by calculating the standardized distance of the current value from the statistical mean. This formula uses a sliding window statistical method to establish a dynamic benchmark, which can adapt to the seasonal and trend characteristics of river channel changes. Compared to fixed-threshold detection methods, this formula provides adaptive anomaly detection capabilities, effectively reducing false alarms and false negatives.

[0125] The principle of Gaussian filtering is based on low-pass filtering theory in signal processing, achieving noise suppression and signal smoothing through convolution operations. This formula uses a Gaussian kernel function as the filter, which can suppress noise while preserving the integrity of edge information. The adaptive adjustment mechanism of the standard deviation parameter allows the filter intensity to dynamically change according to the noise level. Compared to fixed-parameter filtering methods, this formula achieves a better balance between noise suppression and detail preservation.

[0126] The principle behind the time series forecasting formula is based on sequence modeling theory in deep learning, capturing the temporal dependencies of river channel changes through a long short-term memory network. This formula takes multidimensional features as input and outputs predicted values ​​for future times, while also considering a prediction error term to improve the model's robustness. Compared to traditional linear forecasting methods, this formula can model complex nonlinear time relationships, significantly improving the accuracy and reliability of river channel change predictions.

[0127] The gradient descent formula for parameter optimization is based on the gradient optimization method in optimization theory, using the gradient of the loss function to guide the parameter update direction. This formula combines prediction error and regularization constraints, improving detection accuracy while preventing overfitting. The adaptive learning rate mechanism allows the optimization process to dynamically adjust the update step size based on convergence, providing more stable and efficient parameter optimization capabilities compared to fixed learning rate optimization methods.

[0128] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: A technical team needs to continuously monitor a 15-kilometer-long section of a river in a certain watershed. In recent years, due to increased rainfall and the impact of human activities, the erosion of the riverbed edge in this section has become increasingly serious, and the traditional manual inspection method is inefficient.

[0129] The technical team first configured a six-axis drone equipped with a high-resolution camera (1024×768 pixels), an NVIDIA Jetson Nano processor, and 4GB of RAM. A lightweight basic feature extraction model was deployed on the onboard processor. This model uses a modified MobileNet architecture, comprising three depthwise separable convolutional layers, two residual connective blocks, and a feature fusion layer, with the model parameters controlled to 7.8MB. The model structure is designed so that the input layer receives a 1024×768 pixel raw river image. The first layer uses a 3×3 depthwise separable convolutional kernel with a stride of 2 and padding of 1 to extract river edge contour features, outputting a feature map of 512×384×32. The second layer uses a 5×5 convolutional kernel with a stride of 1 and padding of 2 to extract texture detail features, outputting a feature map of 512×384×64. The third layer performs feature dimensionality reduction using a 1×1 convolution, outputting a 256-dimensional basic feature vector.

[0130] The granularity adjustment matrix is ​​set to a 32×32 weight matrix, with initial weight values ​​all set to 1.0, and the weight adjustment range is between 0.1 and 2.0. When it is necessary to improve the fineness of feature extraction, the convolution stride is reduced from 2 to 1, the pooling window is reduced from 4×4 to 2×2, and the feature map spatial resolution is expanded from 64×64 to 128×128. Each element in the matrix corresponds to the sensitivity weight of a local region, and the feature extraction accuracy of different regions is controlled by adjusting these weight values ​​in real time.

[0131] The drone flies at a pre-set route, 50 meters above the river channel, at a speed of 8 meters per second, capturing 10 frames of river channel images per second. The onboard processor processes each frame in real time, extracting basic feature vectors, which are then transmitted to the ground server via a 5G wireless link. Compared to transmitting the original image data, the feature vector data volume is only 1.2% of the original image data, significantly reducing data transmission volume and latency.

[0132] The ground server is equipped with a high-performance computing server featuring an NVIDIA RTX 4090 graphics card and 64GB of memory. After receiving the basic feature vectors, the server calls a pre-trained deep edge detection model for processing. This model uses a UNet encoder-decoder architecture. The encoder contains four downsampling blocks, each with two 3×3 convolutional layers and one 2×2 max-pooling layer. The decoder contains four upsampling blocks, each with a deconvolutional layer and skip connections. The deep feature mining matrix is ​​set to a 16×16 parameter matrix, controlling the effective depth of the network to vary between 12 and 20 layers. The activation threshold and feature weight distribution of each layer are changed by adjusting the matrix element values.

[0133] like Figure 2 As shown, the system monitors the detection confidence of the river edge location matrix in real time during operation. During a certain period, the detection confidence fluctuates. When the confidence drops to 82%, below the preset threshold of 85%, the system automatically sends adjustment commands to the drone. The gain amplification matrix, a 64×64 coefficient matrix, adjusts the parameters of the batch normalization layer in the lightweight basic feature extraction model, increasing the scaling factor from 1.0 to 1.5, adjusting the bias value from 0.0 to 0.2, and increasing the learning rate of the convolutional layer from 0.001 to 0.003. The matrix element values ​​control the amplification factor of each feature channel, which is adjusted within the range of 1.0 to 3.0. After adjustment, the detection confidence increases to 89%.

[0134] The system continuously analyzes abnormal fluctuations in the river channel edge change trend matrix. The standard variation amplitude matrix was established using river monitoring data collected over five years, including edge change records under different water level conditions. The matrix size is 100×100, and each matrix element records the normal variation range at its corresponding spatial location. Statistical analysis yields the mean and standard deviation of the variation at each location, with the normal range defined as the interval between the mean and two standard deviations above and below. See Table 1 for details.

[0135] Table 1. Standard variation data for different river sections

[0136]

[0137] When the change in a monitoring point exceeds the range set by the standard change amplitude matrix, the system initiates a mitigation filtering process. The mitigation and suppression matrix is ​​set to a 128×128 filter matrix. Noise suppression is achieved by adding a Gaussian filter during feature extraction. The Gaussian filter kernel size is 5×5, and the standard deviation parameter is adjusted within the range of 0.5 to 2.0 based on the element values ​​of the mitigation and suppression matrix. Simultaneously, the weighting coefficient of high-frequency features is reduced from 1.0 to 0.3, and a moving average window is applied to the time series, with the moving average window size increased from 3 frames to 7 frames for smoothing.

[0138] like Figure 3 As shown, the river erosion prediction model is based on a time series analysis algorithm, using a riverbed edge location matrix and a riverbed edge change trend matrix over 15 consecutive days as input data. The riverbed edge location matrix describes the precise location of the river boundary line in a two-dimensional coordinate system, with matrix element values ​​representing the coordinate information of edge points. The matrix size corresponds to the original river image resolution of 1024×768. The riverbed edge change trend matrix describes the pattern of river boundary line change over time, with matrix element values ​​representing the speed and direction of edge change at different locations. It is calculated from the riverbed edge location matrix over multiple consecutive frames. The prediction model outputs the predicted location of the riverbed edge for the next 7 days, generating an early warning signal when the predicted change exceeds the safe range of 5.0 cm.

[0139] During actual monitoring, on the 8th day, the system detected an edge change of 8.3 cm per month at a certain location in a 6-9 km river section, exceeding the normal upper limit of 7.2 cm per month for that location. The prediction model calculated that this location would continue to erode by 2.1 cm within the next 7 days, exceeding the safety threshold. The system immediately generated an early warning signal, including the coordinates of the changed location at 7.2 km of the river section, the change rate of 8.3 cm per month, and the expected occurrence time within the next 7 days.

[0140] The system has established a comprehensive feedback optimization mechanism, and the technical team regularly conducts on-site measurements and verifications of the warning areas. The accuracy of the edge detection results is verified by measuring the riverbank positions using GPS positioning and a laser rangefinder. The measured data shows that the system's detection accuracy reaches 94.3%, with errors mainly originating from variations in illumination and interference from water surface reflections. Based on the feedback from the measured data, the system dynamically updates the parameter values ​​of the granularity adjustment matrix and the deep feature mining matrix. The weight value corresponding to the illumination-sensitive area in the granularity adjustment matrix is ​​adjusted from 1.2 to 1.6, and the activation threshold of the relevant levels in the deep feature mining matrix is ​​increased from 0.3 to 0.4, continuously optimizing the performance of the entire detection system.

[0141] After 30 days of continuous monitoring, the system processed a total of 432,000 river images, identifying 17 potential erosion risk points. Of these, 12 were verified on-site to have erosion problems, achieving a warning accuracy rate of 70.6%. The system processes an average of 600 images per hour, with a single frame processing time of 83 milliseconds, meeting real-time monitoring requirements. The drone's average flight time is 45 minutes, covering a 5-kilometer river section, and it can complete comprehensive monitoring of the entire 15-kilometer river section within a day.

[0142] Compared to traditional manual inspection methods, this invention significantly reduces data transmission burden through airborne lightweight feature extraction, avoiding the network transmission of large amounts of raw image data and fundamentally solving the real-time problem. Traditional methods require transmitting the entire image before processing, while this invention completes preliminary feature extraction on the UAV, transmitting only the compressed feature vector, greatly reducing dependence on network bandwidth. The deep edge detection model uses a UNet encoding / decoding structure that effectively preserves spatial location information. Combined with a skip connection mechanism, it avoids information loss during deep propagation in traditional convolutional networks, significantly improving edge localization accuracy. The dynamic parameter adjustment mechanism, through the coordinated work of granularity adjustment matrices, gain amplification matrices, and mitigation suppression matrices, adaptively optimizes model parameters based on actual detection results, overcoming the limitations of fixed-parameter models that struggle to adapt to complex environmental changes. The time-series prediction model establishes a standard variation amplitude matrix based on historical data statistical patterns, providing a scientific judgment benchmark for anomaly detection. Compared to simple threshold judgment methods, it has stronger environmental adaptability and early warning accuracy.

[0143] It should be noted that the variables involved in this invention are explained in detail in Table 2.

[0144] Table 2 Variable Explanation Table

[0145]

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for calculating the edge of a river channel using unmanned aerial vehicles (UAVs) based on artificial intelligence, characterized in that, The drone, equipped with a high-resolution camera, acquires raw river images. A lightweight basic feature extraction model is deployed in the onboard processor to process the raw river images in real time and extract basic feature vectors. The basic feature vectors are transmitted to the ground server via a wireless transmission link. The ground server calls a deep edge detection model to process the basic feature vectors and outputs a river edge position matrix and a river edge change trend matrix. The detection confidence of the river edge position matrix is ​​calculated. When the detection confidence is lower than a preset threshold, an adjustment command is sent to the drone. Abnormal fluctuations in the river edge change trend matrix are analyzed. When the change amplitude exceeds the set range of the standard change amplitude matrix, mitigation filtering is initiated. Based on the river edge position matrix and the river edge change trend matrix, a river erosion prediction model is constructed to generate an early warning signal. The lightweight basic feature extraction model adopts an improved MobileNet architecture, which includes three depthwise separable convolutional layers, two residual connection blocks and one feature fusion layer. The model parameter size is controlled within 8MB. The input layer receives the original river channel image. The first layer uses depthwise separable convolutional kernels to extract edge contour features, the second layer extracts texture detail features, and the third layer performs feature dimensionality reduction through convolution. The adjustment instruction adjusts the feature enhancement parameters of the lightweight basic feature extraction model through the gain amplification matrix. The gain amplification matrix is ​​a coefficient matrix used to amplify the intensity of the feature signal. The adjustment process involves the scaling parameters and bias parameters of the batch normalization layer in the lightweight basic feature extraction model. When the detection confidence is lower than the threshold, the scaling coefficient of the batch normalization is increased, the bias value is adjusted, and the learning rate of the convolutional layer is increased. The dimension of the basic feature vector is dynamically adjusted according to the granularity adjustment matrix, which is a weight matrix that controls the fineness of feature extraction. The adjustment process is achieved by modifying the stride parameter of the convolution kernel and the size of the pooling window. When it is necessary to increase the granularity, the convolution stride is reduced, the pooling window is reduced, and the spatial resolution of the feature map is increased.

2. The method for calculating the edge of a river channel using unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, The deep edge detection model uses a deep feature mining matrix to control the depth of feature extraction layers. The deep feature mining matrix is ​​a parameter matrix that controls the number of feature extraction layers and the size of the receptive field in the deep learning model. The deep edge detection model adopts the UNet encoder-decoder structure, with the encoder containing four downsampling blocks and the decoder containing four upsampling blocks. The deep feature mining matrix controls the activation threshold and feature weight allocation of each layer.

3. The method for calculating the edge of a river channel using unmanned aerial vehicles based on artificial intelligence according to claim 2, characterized in that, The mitigation filtering process reduces the impact of noise interference on edge detection accuracy by using a mitigation suppression matrix. The mitigation suppression matrix is ​​a filtering matrix used to suppress image noise and abnormal fluctuations. The mitigation filtering process is achieved by adding a Gaussian filter during the feature extraction process. The Gaussian filter kernel size is of standard size, and the standard deviation parameter is adjusted according to the element values ​​of the mitigation suppression matrix, while reducing the weight coefficient of high-frequency features.

4. The method for calculating the edge of a river channel using unmanned aerial vehicles based on artificial intelligence according to claim 3, characterized in that, The standard variation range matrix is ​​a reference matrix obtained by statistical analysis of historical river channel change data. Each matrix element records the normal variation range at the corresponding spatial location. The standard variation range matrix is ​​established using collected river channel monitoring data, including edge change records under different water level conditions. The mean and standard deviation of the variation at each location are obtained through statistical analysis. The normal range is set as the interval between the mean and two standard deviations above and below the mean.

5. The method for calculating the edge of a river channel using unmanned aerial vehicles based on artificial intelligence according to claim 4, characterized in that, The river edge position matrix is ​​a numerical matrix that describes the precise position of the river boundary line in a two-dimensional coordinate system. The matrix element values ​​represent the coordinate information of the edge points, and the matrix size corresponds to the resolution of the original river image. The river edge change trend matrix is ​​a prediction matrix that describes the change law of the river boundary line over time. The matrix element values ​​represent the speed and direction of edge change at different locations and are calculated from the river edge position matrix of multiple consecutive frames.

6. The method for calculating the edge of a river channel using unmanned aerial vehicles based on artificial intelligence according to claim 5, characterized in that, The river erosion prediction model is a prediction algorithm based on time series analysis. It uses the river edge location matrix and river edge change trend matrix over a continuous time period as input and outputs the predicted location of the river edge in the future time period. When the predicted change exceeds the safe range, an early warning signal is generated. The early warning signal includes the coordinates of the changed location, the change magnitude, and the expected time of occurrence.

7. The method for calculating the edge of a river channel using unmanned aerial vehicles based on artificial intelligence according to claim 6, characterized in that, It also includes: establishing a feedback optimization mechanism to dynamically update parameter values; avoiding the transmission of raw image data and significantly reducing data transmission volume through a separate architecture for airborne lightweight feature extraction and ground depth processing; and dynamically optimizing feature extraction accuracy based on detection results through an adaptive parameter adjustment mechanism.

Images