Deep learning hydraulic engineering slope measurement method

Abstract

The invention discloses a water conservancy project slope measurement method based on deep learning, and particularly relates to the technical field of water conservancy project measurement. Acquiring a remote sensing image and digital elevation model data of a to-be-measured area, and constructing a multi-scale remote sensing and topographic feature data set; performing gradient prediction by using a deep neural network model containing a multilayer convolution structure and residual connection, and generating a pixel-level gradient value distribution diagram; performing curved surface fitting and abnormal point correction on the key section to form a continuous and consistent gradient measurement result map; and finally, performing error comparison on a measurement result and historical measured data, outputting an error analysis report, and automatically adjusting model parameters based on error feedback to complete continuous optimization. The method has high precision, high adaptability and intelligent optimization capability, is especially suitable for gradient extraction tasks of complex landforms or engineering control areas, and significantly improves the measurement efficiency and result reliability in water conservancy projects.

Description

Technical Field

[0001] This invention relates to the field of hydraulic engineering surveying technology, specifically to a deep learning-based method for measuring the slope of hydraulic engineering slopes. Background Technology

[0002] In the planning and construction of water conservancy projects, accurate measurement of topography is one of the prerequisites for ensuring the safety and feasibility of the project. In particular, slope (the rate of change of topographic elevation) is a key parameter affecting surface runoff, structural stability, and drainage system design, and has a significant impact on the functionality and durability of the entire water conservancy project. Traditional slope measurement methods mainly rely on high-precision measuring instruments such as LiDAR, total stations, or GPS. Although they have high accuracy, they are inefficient in complex terrain areas such as high mountains and valleys, densely vegetated areas, or areas with obstructed visibility. They also depend heavily on the technical skill of the surveyors, and the stability of the measurement results fluctuates.

[0003] In recent years, with the development of remote sensing technology and deep learning, automated terrain identification using multi-source terrain data such as satellite imagery and UAV aerial photography has become a research hotspot. However, existing image recognition-based slope extraction methods generally suffer from problems such as blurred edges, texture interference, and elevation misjudgment, and their accuracy varies greatly when processing different terrain types. There is a lack of a unified algorithm model suitable for accurate slope identification under multiple terrains and scales. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for measuring the slope of hydraulic engineering projects, in order to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for measuring the slope of hydraulic engineering, comprising: Acquire remote sensing images and corresponding digital elevation model data of the area to be measured; The remote sensing images are preprocessed to construct a multi-scale dataset of remote sensing images and terrain features, including edge texture, gray-level gradient and NDVI index; A deep neural network model for slope extraction is constructed based on labeled samples. The deep neural network model includes multi-layer convolutional structures and residual connections. Input the multi-scale dataset into the trained deep neural network model and output a slope value distribution map of the corresponding terrain unit; Based on the slope value distribution map, the key sections are fitted and outliers are corrected to construct the final slope measurement result map; The final slope measurement results are compared and evaluated with historical measurement data, an error analysis report is output, and the parameters of the deep neural network model are continuously optimized based on the error distribution.

[0006] Preferably, the acquisition of remote sensing images and corresponding digital elevation model data of the area to be measured includes: retrieving remote sensing image data of different time phases and resolutions from a multi-source remote sensing data platform based on the geographic boundary information of the area to be measured, and prioritizing the selection of clear images with cloud cover rates below a set threshold; performing georegistration and band fusion processing on the acquired remote sensing images, and adaptively selecting digital elevation model data with matching resolution using regional elevation difference statistical parameters, and performing spatial resolution unification and raster registration with the remote sensing images.

[0007] Preferably, the preprocessing of the remote sensing image to construct a multi-scale dataset of remote sensing image and terrain features, including edge texture, grayscale gradient, and NDVI index, includes: The acquired remote sensing images are subjected to image decomposition processing, and a multi-scale image set is constructed based on the Gaussian pyramid; Edge texture extraction, gray-level gradient calculation, and normalized vegetation index (NDVI) analysis were performed on multi-scale image sets. Edge texture was extracted using the directional gradient histogram method, gray-level gradient was calculated using the Sobel operator or the Laplacian operator, and the NDVI index was obtained based on the reflectance of red light and near-infrared bands. Image features of different scales and types are fused into a unified multi-channel image dataset through feature mapping, which serves as the input sample for a deep neural network model. The feature fusion process employs weighted fusion.

[0008] Preferably, the construction of the deep neural network model for slope extraction based on labeled samples includes: A supervised training sample set is constructed based on remote sensing images and corresponding measured slope data. The sample set includes multi-channel input image data and corresponding pixel-level slope label maps. A deep neural network model containing a multi-layer convolutional structure is constructed. The deep neural network model extracts edge and texture features in the shallow layer, extracts spatial structure features in the middle layer, learns high-level semantic and terrain relationship features in the deep layer, and enhances information flow through cross-layer connections. A residual connection structure is introduced into the convolutional network. The residual connection structure uses an identity mapping method to add the features of the previous layer to the output of the next layer element by element.

[0009] Preferably, the step of inputting the multi-scale dataset into the trained deep neural network model and outputting a slope value distribution map corresponding to the terrain unit includes: The constructed multi-scale remote sensing image and terrain feature dataset was normalized and channel-arranged. The preprocessed multi-scale data is input into the trained deep neural network model. During the inference process, the data passes through the shallow feature extraction module, the middle spatial structure analysis module, and the deep semantic feature discrimination module in sequence, and the slope prediction value of each pixel is generated in the output layer. The predicted slope values ​​are spatially reorganized and post-processed, including mapping the pixel-level prediction results to the original geographic coordinates to generate a slope value distribution map.

[0010] Preferably, the step of fitting and correcting outliers in key areas based on the slope distribution map to construct the final slope measurement result map includes: Extract areas of abrupt slope changes, slope boundaries, and the neighborhood of engineering control points as key areas, and construct corresponding topographic structure boundary index maps; Surface fitting is performed on the slope prediction results of key areas; By combining the fitting results with the statistical characteristics of the neighborhood, outliers are detected and corrected. The criteria for judging outliers is that the deviation between the predicted value and the fitted value exceeds the deviation threshold. The correction strategies include neighborhood mean substitution, fitting interpolation, or structure-guided reconstruction, and finally, a slope measurement result map is generated.

[0011] Preferably, the step of comparing and evaluating the final slope measurement results with historical measurement data includes: Based on spatial coordinate registration and resolution unification processing, the final slope measurement result map is aligned with the corresponding historical measured slope data at the pixel level to construct a one-to-one corresponding comparison dataset. Based on the comparison dataset, error statistics are calculated, including mean absolute error, root mean square error, maximum error, and error distribution frequency histogram. Sub-region error summary information is also output based on the engineering region division.

[0012] Preferably, the step of continuously optimizing the deep neural network model parameters according to the error distribution includes: Identify high-error regions and error types, and construct an error-sensitive area sample set, including corresponding remote sensing inputs, multi-scale feature maps, and labeled slopes; Based on the parameters of the deep neural network model, retraining is performed using samples from the error-sensitive region to adjust the feature weight distribution and loss function parameters of the intermediate layer of the network. The parameters of the deep neural network model are updated based on the results of the new round of training, and an error threshold is set. When the error in multiple consecutive regions exceeds the set range, retraining is automatically triggered.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a multi-scale terrain feature set and combines it with a deep neural network model featuring a hierarchical convolutional structure and residual connections to achieve high-precision automatic extraction of terrain slope. Compared with traditional methods based on DEM calculation or manual interpretation, the method of this invention exhibits stronger robustness and adaptability under complex terrain conditions, significantly improving the spatial resolution, continuity, and physical interpretability of slope measurement.

[0014] 2. This invention also introduces an error feedback-driven model optimization mechanism, which can automatically identify high-error areas and adjust model parameters based on historical measurement data, continuously improving overall measurement accuracy. This method has strong engineering applicability, especially suitable for slope monitoring and design support in key water conservancy engineering areas such as high slopes, dams, and landslides, providing a stable and reliable technical means for realizing intelligent water conservancy engineering surveying. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] 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.

[0018] For examples, please refer to Figure 1 As shown in this embodiment, a deep learning-based method for measuring slope in hydraulic engineering includes: Acquire remote sensing images and corresponding digital elevation model data of the area to be measured; The remote sensing images are preprocessed to construct a multi-scale dataset of remote sensing images and terrain features, including edge texture, gray-level gradient and NDVI index; A deep neural network model for slope extraction is constructed based on labeled samples. The deep neural network model includes multi-layer convolutional structures and residual connections. Input the multi-scale dataset into the trained deep neural network model and output a slope value distribution map of the corresponding terrain unit; Based on the slope value distribution map, the key sections are fitted and outliers are corrected to construct the final slope measurement result map; The final slope measurement results are compared and evaluated with historical measurement data, an error analysis report is output, and the parameters of the deep neural network model are continuously optimized based on the error distribution.

[0019] In this embodiment of the invention, in order to accurately extract the slope of the water conservancy project area, it is necessary to first acquire and preprocess the remote sensing image data and digital elevation model (DEM) data of the target area.

[0020] First, the geographical boundary information of the area to be measured is used to define the area's scope. Geographical boundary information can be obtained by user-input coordinate ranges (e.g., latitude and longitude selection) or by automatically extracting regional vector boundary files based on hydraulic engineering design drawings, ensuring a clear and unambiguous spatial scope during data retrieval. After defining the target boundary, remote sensing imagery data covering the area at different temporal phases and spatial resolutions is retrieved from multi-source remote sensing data platforms (such as the Sentinel satellite data platform, Landsat remote sensing imagery platform, or high-resolution commercial remote sensing databases). To improve the accuracy of subsequent slope measurements, remote sensing images with cloud cover below a set threshold (e.g., 20%) are prioritized from the multi-temporal images. Cloud cover information is generally automatically determined through image metadata or NDVI index analysis, thus avoiding elevation misjudgments caused by cloud obstruction.

[0021] Secondly, to ensure the spatial and spectral consistency of remote sensing image data, georegistration and band fusion processing are performed on the acquired remote sensing images. The georegistration process can employ control point-based image transformation algorithms, such as first-order affine transformation or polynomial transformation. By selecting invariant ground feature points (such as road intersections, mountain peaks, or buildings) as registration benchmarks, geometric deviations caused by changes in satellite orbit and attitude between remote sensing images are corrected. For band fusion, for multi-source, multi-band remote sensing images (such as visible light and near-infrared), principal component analysis (PCA) or weighted average fusion algorithms are used to fuse multi-band information, enhancing the overall contrast and boundary feature clarity of the image, thereby improving the response capability of subsequent slope extraction models to local terrain undulations.

[0022] Furthermore, to enhance the accuracy coupling between remote sensing imagery and DEM data, this invention proposes an adaptive matching resolution DEM data approach based on regional elevation difference statistical characteristics. Specifically, firstly, preliminary topographic relief estimation is performed on the acquired remote sensing imagery. For example, spatial frequency analysis is conducted using the image grayscale gradient distribution to obtain the elevation change complexity index of the region; or, with the assistance of existing coarse-resolution DEM data, statistical quantities such as regional elevation difference variance, maximum slope value, and contour line density are extracted. Based on these characteristics, the area to be measured is divided into three categories: high-relief areas, medium-relief areas, and gently undulating areas.

[0023] Based on this, DEM data sources with different spatial resolutions are automatically matched. For example, DEM data with a resolution of 5 meters or higher is preferentially used in areas with high undulations, 10-meter resolution DEMs are used in areas with moderate undulations, and standard 30-meter resolution DEMs are used in flat areas. This process avoids data redundancy or information loss problems caused by uniform resolution, and realizes a data selection strategy of on-demand acquisition and accuracy adaptation.

[0024] Subsequently, the selected DEM data and remote sensing imagery are processed to unify their spatial resolution. If there is a difference in resolution between the DEM and the remote sensing imagery, the remote sensing imagery resolution is prioritized, and the DEM is spatially resampled using bilinear interpolation or spline interpolation methods to achieve the same pixel size as the remote sensing imagery. Simultaneously, the spatial reference coordinate systems of both are unified (e.g., both use WGS-84 or UTM projection), and raster registration methods are used to ensure spatial alignment of the imagery and elevation data at the pixel level, meaning that each pixel has a unique correspondence between the remote sensing imagery and the DEM.

[0025] To improve the adaptability and accuracy of deep learning-based slope measurement methods for water conservancy projects under different terrain conditions, this invention constructs a multi-channel input dataset that integrates features such as edge texture, grayscale gradient, and normalized vegetation index (NDVI) to provide high-quality training input for subsequent neural network models.

[0026] First, the acquired remote sensing image data undergoes image decomposition processing, and a multi-scale image set is generated using the Gaussian pyramid construction method. The Gaussian pyramid is a typical multi-resolution image representation method that can effectively extract spatial detail features at different scales while preserving the overall image structure.

[0027] Specifically, the original remote sensing image is first processed by Gaussian filtering to remove high-frequency noise, and then downsampled to generate several layers of image copies with different resolutions, each layer having half the resolution of the layer above. For example, if the original image resolution is 1024×1024 pixels, then the first layer is 512×512 pixels, the second layer is 256×256 pixels, and so on, until a set minimum size threshold is reached. This forms an image hierarchy structure from full scale to low scale.

[0028] This multi-scale image set can capture both large-scale undulations (such as mountain distribution) and local detailed changes (such as gullies and fissures) on the land surface, providing basic input conditions for the extraction of features at different scales.

[0029] In the constructed Gaussian pyramid image set, the following three key terrain-related feature extraction operations were performed respectively: Edge texture feature extraction: For each multi-scale image layer, the histogram of oriented gradients (HARQ) method is used to extract edge texture features. This method generates a set of directional edge response features by calculating the gradient direction distribution of pixel gray values ​​within a local region of the image, which can effectively reflect geometric features such as terrain contours, riverbank boundaries, and ridgelines. To improve noise resistance and directional sensitivity, the feature extraction window size and the number of directional channels are adaptively adjusted according to the image scale.

[0030] Gray-level gradient feature calculation: At each image scale, the Sobel operator or Laplacian operator is used to perform edge enhancement and gray-level gradient calculation. The Sobel operator is a first-order derivative filter that can extract the edge intensity in the horizontal and vertical directions of the image, while the Laplacian operator, as a second-order derivative operator, is suitable for capturing high-frequency changing areas, such as abrupt changes in terrain or points of rapid slope change. The calculation results will generate a gradient magnitude map to reflect the intensity of local elevation changes.

[0031] NDVI Calculation: NDVI, or Normalized Difference Vegetation Index, is an important indicator used in remote sensing images to measure vegetation cover. Its calculation formula is: (Near-infrared reflectance minus red reflectance) divided by (Near-infrared reflectance plus red reflectance). This formula yields the NDVI value for each pixel, typically ranging from -1 to 1. A higher NDVI value indicates denser vegetation cover, while a lower value indicates bare land, water bodies, or artificial surfaces. Since vegetation cover often obscures or affects topographic relief information, introducing NDVI features can help models identify and filter out anomalous areas caused by vegetation disturbance when extracting slope.

[0032] After completing the multi-scale image feature extraction described above, it is necessary to integrate image features of different scales and types into input samples of a unified format for use by subsequent deep neural networks. To this end, this invention adopts the following fusion strategy: Within each image scale level, the corresponding edge texture feature map, grayscale gradient feature map, and NDVI index map are stacked according to the channel dimension to form a three-channel or multi-channel image representation. Subsequently, the feature channel maps of each scale are unified to the original image resolution through upsampling or interpolation reconstruction methods, and channel alignment is performed.

[0033] A weighted fusion mechanism is used to integrate features at multiple scales. This mechanism adaptively assigns fusion weights based on the spatial distribution and gradient intensity of features extracted from images at different scales. Specifically, if the edge response at a certain scale is strong, then that scale is given a higher weight during the fusion process; while for regions with uniform NDVI distribution, their weight can be appropriately reduced to avoid interfering with slope discrimination. The fusion result generates a unified multi-channel remote sensing feature dataset, which is used as the training input for a deep learning model.

[0034] Furthermore, the fusion process can optionally incorporate an attention mechanism, guiding the model to autonomously learn the correlation and importance between different feature channels through training, thereby further enhancing the overall feature representation capability. The final input sample not only contains rich geometric and spectral information but also possesses multi-scale terrain perception capabilities, providing a solid data foundation for subsequent slope prediction.

[0035] To achieve high-precision and automated extraction of slope in complex terrain areas from remote sensing images, this invention constructs a deep neural network model based on labeled samples, on the basis of feature engineering processing. This model integrates multi-scale image information, multi-layer convolutional feature expression capabilities, and residual connection optimization structure, which significantly improves the slope recognition performance of the model under various landform types.

[0036] First, a supervised sample set needs to be constructed for training. This sample set takes remote sensing imagery as input and measured slope data of the corresponding area as supervision labels, forming an input-output pairing structure.

[0037] Specifically, the input data is a multi-channel remote sensing image dataset that has undergone preprocessing and feature fusion. Each sample unit (such as an image patch) contains multiple image channels, including but not limited to red light, near-infrared, NDVI index map, grayscale gradient map, and edge texture map. Each channel is consistent with the other channels in terms of spatial resolution and registration accuracy, ensuring pixel-level synchronization during network input.

[0038] The output label is the slope label map corresponding to the region, which can be obtained by combining an existing digital elevation model (DEM) with a slope calculation method. Specifically, the elevation value change is extracted within the neighborhood of each pixel, and the local slope is calculated by dividing the elevation difference by the horizontal distance. The slope is then expressed as an angle or percentage, thereby generating a pixel-level slope image as a supervision signal.

[0039] To enhance the model's generalization ability, the sample data also needs to undergo data augmentation processing, including but not limited to: Image rotation: Rotate the input image patch by 90 degrees, 180 degrees, and 270 degrees to enhance the model's robustness to changes in orientation; Scale: Zoom in or out of the image to simulate different aerial photography altitude conditions; Noise disturbance: Gaussian noise simulating sensor error is added to improve the model's anti-interference ability; Occlusion simulation: Introduce random occlusion regions to train the model's recognition ability under conditions of local information loss.

[0040] After the above processing, a training dataset containing thousands to tens of thousands of sample pairs is constructed for training deep neural network models.

[0041] To address the typical continuous spatial structural feature of slope, this invention designs a deep neural network model with a multi-layer convolutional structure. This model can extract terrain features at different levels, improving the learning efficiency and expressive power of the model. The overall network structure is divided into three feature extraction stages from bottom to top: shallow, middle, and deep layers, supplemented by a cross-layer connection mechanism to enhance the feature integration effect.

[0042] Shallow convolutional structures are used to extract basic image features from the input image, such as edges, textures, and gray-level gradients. This part typically contains 2-3 convolutional layers, each with a small kernel size (e.g., 3×3) and a stride of 1. The ReLU activation function is used to enhance non-linear expressiveness. Shallow feature extraction helps the model quickly locate key structural information such as terrain boundaries, slope toes, and slope crests.

[0043] Mid-level convolutional structure: The mid-level network mainly extracts spatial structural features and regional distribution patterns, including terrain undulations and slope trends. This part uses deeper convolutional stacking and downsampling (pooling) operations to expand the receptive field while maintaining spatial feature representation capabilities. Max pooling is generally used, combined with batch normalization to improve training stability.

[0044] Deep convolutional structures: Deep layers are used to learn higher-order semantic information, such as local elevation relationships, slope orientation, and landform types. This part uses a larger number of convolutional channels (such as 128 or 256 channels) to increase network capacity, while combining global average pooling and Dropout structures to prevent overfitting.

[0045] Through a multi-level structure design of shallow-medium-deep layers, the model can realize slope expression modeling from local to overall, and has a good ability to perceive hierarchy.

[0046] In deep network structures, as the number of convolutional layers increases, problems such as gradient vanishing and feature degradation easily occur, especially when processing high-resolution remote sensing images. To address this, this invention introduces residual connection structures into convolutional networks to optimize information transmission paths and improve network stability and training efficiency.

[0047] The residual connection structure employs an identity mapping method, which involves directly adding the input feature map of a certain layer to the output of the subsequent layer element-wise. Mathematically, this is expressed as follows: if the input of a convolutional module is X, and the output after the convolution operation is F(X), then the final output is Y = F(X) + X.

[0048] This structure allows the network to directly learn the "difference between input and output," learning additional feature changes while retaining the original information, effectively mitigating the performance degradation problem that occurs as the network deepens. Especially in complex slope areas, such as multi-level steps, cliffs, or artificial slopes, the residual connection structure can improve the model's ability to perceive and predict local slope abrupt changes.

[0049] In actual network implementations, a set of residual connection units is introduced every two to three convolutional layers, and combined with activation functions and normalization layers to construct complete residual modules, so as to ensure that information flows efficiently in the network.

[0050] Before inputting data into the model, the constructed multi-scale remote sensing imagery and terrain feature dataset needs to be preprocessed to meet the requirements of deep neural networks for input data format and distribution, ensuring the stability of the inference process and the accuracy of the prediction results.

[0051] The first step is normalization. Because the numerical ranges of different channels in remote sensing images vary significantly (for example, NDVI values ​​are typically between -1 and 1, while the original red band may range from 0 to 255), directly inputting them into the model could easily lead to some channels having their weights ignored or amplified during training and inference. Therefore, it is necessary to perform normalization on each image channel separately, scaling its values ​​to the same range. Common normalization methods include min-max scaling (subtracting the minimum value of each channel from the pixel value and then dividing by the difference between the maximum and minimum values) or Z-score normalization (subtracting the mean and then dividing by the standard deviation). This invention preferably uses min-max normalization to preserve the relative distribution structure between pixels.

[0052] The next step is channel arrangement. Multi-scale image features may originate from different bands, scales, and processing methods, therefore the input channels must be arranged in a uniform order. For example, the data can be organized as follows: channel 1 is the red light band, channel 2 is the near-infrared band, channel 3 is NDVI, channel 4 is the edge feature map, channel 5 is the grayscale gradient map, channels 6-8 are multi-scale texture feature maps, and so on. By unifying the channel order, the model's recognition and parsing of input features remains consistent during the training and inference phases.

[0053] After the above normalization and channel arrangement are completed, the input data forms a tensor structure with uniform dimensions and format, ready to enter the deep neural network model for forward inference.

[0054] The preprocessed data is used as input to the trained deep neural network model for forward inference. The model structure remains consistent with the training phase, including a shallow feature extraction module, a mid-level spatial structure analysis module, and a deep semantic feature discrimination module. The overall inference process is as follows: Shallow Feature Extraction Module: This module first extracts low-level image features from the input image, such as edges, textures, colors, and grayscale variations, through multiple shallow convolutional layers. The shallow convolutional kernel size is typically 3×3 with a stride of 1, and the output feature map retains the original spatial dimensions. This module can quickly capture basic information such as terrain boundaries, slope toe lines, and vegetation distribution edges, providing support for subsequent spatial structure analysis.

[0055] The intermediate structure analysis module expands the receptive field through convolution and pooling operations, extracting the spatial distribution patterns of slope changes, such as slope continuity, slope direction, and topographic structural features like the ratio of slope length to slope height. This module is responsive to both smooth and abrupt slope regions and is a key component of the model's slope estimation.

[0056] Deep semantic discrimination module: This module further learns high-level terrain semantic information, such as slope type (natural slope / artificial slope) and typical slope combination patterns (such as composite slope, stepped slope, etc.), through deep convolutional structures and residual connection mechanisms, and generates high-dimensional semantic feature maps accordingly. The model output layer uses a fully convolutional structure to predict each pixel of the input image, and the output is a single-channel slope image with the same size as the input image.

[0057] Ultimately, the tensor structure output by the model is a two-dimensional slope prediction map, where the value of each pixel represents the estimated slope at that location, and the unit can be angle (°) or percentage (%), depending on the type of training labels.

[0058] After completing the pixel-level prediction of slope values, the output needs to be reconstructed into a slope value distribution map that conforms to the actual geographic coordinate system, so as to achieve docking with the original topography, engineering layout map, and surveying benchmark. This process includes the following two steps: Spatial Reassembly and Geographic Mapping: The slope image output by the model is geographically reassembled according to the spatial reference system of the original remote sensing image. This process, based on the georegistration information of the input image (including projected coordinate system, resolution, starting coordinates, etc.), accurately maps the predicted slope value of each pixel to its corresponding location in geospace. After mapping, it can be output in raster format (such as GeoTIFF) for further analysis or overlay display on a GIS platform.

[0059] Post-processing optimization: Due to potential local prediction jitter or outliers in edge regions and feature transition zones, deep learning models require post-processing of the predicted slope map. Two main methods are employed: Local smoothing filtering: Median filtering or Gaussian filtering is used to locally smooth the slope map, suppressing isolated outliers while preserving the boundaries of slope abrupt changes; Outlier detection and correction: Statistical analysis of slope values ​​is performed. If the difference between a pixel and the average value of its neighborhood exceeds a set threshold, it is considered an outlier and replaced or corrected to improve the continuity and stability of the overall slope field.

[0060] After completing the above processing, the final slope distribution map can be used in applications such as water conservancy engineering design and analysis, slope stability assessment, and drainage path layout.

[0061] After generating the slope distribution map based on the deep neural network model, in order to further improve the continuity, reliability and accuracy of the predicted slope data in engineering applications, it is necessary to perform targeted fitting and outlier correction on key sections of the distribution map, so as to generate the final slope measurement result map with spatial consistency and structural integrity.

[0062] Slope distribution maps typically contain several terrain transition zones or engineering-sensitive areas. Due to abrupt changes in terrain, complex landforms, or mapping obstructions, these areas are highly susceptible to local discontinuities or misjudgments in model predictions. To ensure the engineering usability of the survey results maps, these areas need to be prioritized.

[0063] This invention automatically extracts key areas and constructs an index map by performing spatial variation rate analysis and engineering location identification on slope value distribution maps. The determination of key areas includes the following three categories: Slope abrupt change zone: This refers to an area where the slope value changes drastically within a local region. A slope gradient map is formed by calculating the slope difference between each pixel and its neighboring pixels. When the gradient value of a pixel exceeds a set abrupt change threshold (e.g., 5 degrees / pixel), the area is marked as an abrupt change zone.

[0064] Slope boundary zone: In natural or artificial landforms, the junction of the slope toe, slope top, and slope surface is often a structural boundary area with prominent changes. Edge detection algorithms (such as the Canny operator) are used to identify the edges of the slope map, which further enhances the spatial positioning ability of these areas.

[0065] The neighborhood of engineering control points: This refers to the area surrounding dams, slopes, and retaining structures, and is a critical region upon which structural safety analysis is based. The neighborhood boundary is constructed by setting a buffer radius using engineering drawings or control point coordinate information.

[0066] The above identification results were integrated into a key area boundary index map, providing a spatial reference for subsequent fitting and correction processing.

[0067] In order to improve the continuity and engineering physical interpretation of slope data within the identified key areas, this invention uses an adaptive weighted least squares surface fitting method to locally reconstruct the slope data.

[0068] The specific method is as follows: Using the center pixel of each key region as a reference, a sliding window (such as 5×5, 7×7, or 9×9 pixels) is constructed. The window size can be automatically adjusted according to the terrain complexity. The fitting process establishes the following functional relationship: A weighted least squares error objective function is established for the spatial location (represented by row and column coordinates) of all pixels within the window and their predicted slope values. A set of fitting coefficients is calculated so that the fitted surface is closest to the original slope value within the window and the boundary changes are smooth.

[0069] The weighting factors are set according to the following principles: The closer the pixel is to the center, the higher its weight. If a pixel is located in a known high-confidence area (such as a bare area with a slow slope and low NDVI value), its weight should be appropriately increased.

[0070] Through this fitting operation, the reconstructed slope surface is more continuous and has a smoother trend in space, which can more realistically reflect the essential characteristics of the terrain and avoid steep changes caused by local interference in the model prediction.

[0071] Although the aforementioned deep neural networks have good generalization capabilities, they may still produce individual anomalous pixels when dealing with situations such as shadow occlusion, building interference, and edge overlap, meaning that their predicted slope values ​​deviate significantly from the surrounding trend. To further improve the data reliability of the resulting maps, this invention designs a multi-strategy outlier identification and correction mechanism.

[0072] Outlier detection: The fitted results are compared with the original predicted slope values. If the absolute difference between the predicted value and the corresponding fitted value of a pixel exceeds the set threshold (e.g., the slope angle is greater than ±10 degrees), or the deviation from the average value in the neighborhood exceeds 3 times the standard deviation, it is marked as an outlier.

[0073] Anomaly correction strategy: For identified outliers, one of the following three correction methods will be used, automatically selected based on the region type: Neighborhood mean substitution method: suitable for locally flat areas, directly using the average value of non-outlier points in a 3×3 or 5×5 neighborhood; Fitting interpolation method: Applicable to areas with continuous slope changes, it uses fitted surface values ​​to replace outliers and maintain the trend of change; Structural-guided reconstruction method: Applicable to slope boundary zones, it combines edge direction information to perform directional weighted reconstruction of outlier points while preserving boundary continuity.

[0074] After the correction is completed, the resulting graph is subjected to local smoothing filtering (such as median filtering) again to eliminate the numerical jump phenomenon at the correction boundary.

[0075] Finally, the image with outlier correction and slope surface reconstruction is output as a standardized raster format (such as GeoTIFF) and includes spatial reference information (such as projection method, resolution, starting coordinates, etc.) to form the final slope measurement result map.

[0076] To ensure the accuracy and robustness of the deep learning-based slope measurement method proposed in this invention in engineering applications, after generating the slope value distribution map and constructing the result map, it is necessary to systematically compare and evaluate the final results with existing historical measured slope data to identify potential error areas and generate an error analysis report. Furthermore, based on the error distribution results, feedback drives the deep neural network model to adjust parameters and retrain, achieving the model's adaptive and continuous optimization capabilities.

[0077] First, based on a spatial registration algorithm, the final slope measurement map is aligned with historical measured slope data in geographic coordinates. Historical measured data can come from GPS measurement points, UAV measurements, or traditional surveying maps; their coordinate information may have projection differences or spatial offsets. Therefore, coordinate transformation and image resampling methods are used to transform the historical data to match the final map. Figure 1 Under consistent projection systems and resolutions, a pixel-level contrast dataset with consistent structure is formed.

[0078] Specifically, for each historical data point, the corresponding pixel is retrieved from the slope result map based on its latitude, longitude, or projected coordinates, and a "predicted value - measured value" pairing list is constructed. If a historical measurement point covers multiple pixels, the neighborhood average value is used instead of the single-point value for comparison.

[0079] After constructing the comparison dataset, the prediction error was calculated for the metadata of each pair of images, and the following key evaluation metrics were statistically analyzed: Mean Absolute Error (MAE): Represents the average of the absolute values ​​of errors of all pixels, and measures the overall level of deviation; Root mean square error (RMSE): The square root of the average of the squared errors. It is sensitive to large error points and measures the stability of errors. Maximum error (MaxE): Identifies the pixel error that deviates most from the measured value, reflecting extreme prediction bias; Error distribution frequency histogram: shows the distribution of the number of pixels in different error intervals, which helps to understand the stability of model prediction; Regional Error Summary Table: The comparison areas are classified and statistically analyzed according to engineering divisions (such as dam slope area, natural mountain area, platform area, etc.), and the mean, variance and maximum deviation of each regional error are output to identify areas with local failure or weak prediction.

[0080] All results are integrated into the error analysis report, serving as an important basis for model adjustment.

[0081] To overcome the problem of decreased prediction accuracy of deep learning models in complex terrain areas or under special spectral interference conditions, this invention further proposes a continuous model optimization strategy based on error feedback. This strategy takes the error distribution as its starting point, constructs a local augmentation sample set, and uses a retraining mechanism to fine-tune the model, thereby improving its learning ability in key areas.

[0082] According to the error analysis report, the first step is to identify areas where the error exceeds the threshold. These typically refer to continuous pixel areas or locally concentrated anomalous areas where the error is greater than a certain set value (e.g., 10 degrees). These areas are defined as high-error areas, and the reasons for this may include complex surface textures, shading, slope orientation deviations, or insufficient coverage of training samples.

[0083] The original remote sensing input, multi-scale feature maps (such as NDVI, texture gradient, edge map, etc.) and measured slope labels of these areas are extracted to construct an error-sensitive area sample set, which serves as the key data source for model retraining.

[0084] Based on the originally trained neural network model, a transfer learning mechanism is used for retraining. The parameters of the basic convolutional layers in the early stages of the model remain unchanged, and only the weights of the intermediate and deep feature extraction layers are adjusted. Dynamic weighted optimization is performed on the loss function to make the model more sensitive to feature changes in the error region.

[0085] For example, in densely vegetated areas where errors are frequent, the feature weights of the NDVI channel can be increased in the loss function; for areas with abrupt slope changes, the response strength of the gradient features can be increased. During training, a mini-batch gradient update method is used to prevent new samples from excessively perturbing the existing model structure.

[0086] During the optimization process, this invention sets up a dynamic error threshold monitoring mechanism. After each prediction, the system automatically evaluates the error index of each region. If the error index exceeds the preset threshold (e.g., RMSE > 8 degrees) multiple times (e.g., three times), a retraining task is automatically triggered, and the model is updated using the latest high-error sample set to ensure that the model has long-term adaptability and learning ability.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application 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 this application should be included within the scope of protection of this application.

Claims

1. A deep learning method for measuring the slope of a water conservancy project, characterized in that: The application relates to a slope measurement method based on deep neural network, comprising the following steps: acquiring remote sensing images and corresponding digital elevation model data of a to-be-measured region; preprocessing the remote sensing images, and constructing a multiscale data set of the remote sensing images and terrain features, including edge texture, gray gradient and NDVI index; constructing a deep neural network model for slope extraction based on labeled samples, wherein the deep neural network model comprises a multilayer convolution structure and a residual connection; inputting the multiscale data set into the trained deep neural network model, and outputting a slope value distribution map of a corresponding landform unit; fitting and correcting abnormal points of a key section according to the slope value distribution map, and constructing a final slope measurement result map; comparing and evaluating the final slope measurement result with historical measurement data, outputting an error analysis report, and continuously optimizing the deep neural network model parameters according to the error distribution.

2. The deep learning method for hydraulic engineering slope measurement according to claim 1, characterized in that: The remote sensing images and corresponding digital elevation model data of the to-be-measured region are acquired by the following steps: based on the geographical boundary information of the to-be-measured region, remote sensing image data of different time phases and different resolutions are called from a multi-source remote sensing data platform, and clear images with a cloud coverage rate lower than a set threshold are preferentially selected; the acquired remote sensing images are subjected to geographical registration and band fusion processing, and adaptive selection of digital elevation model data of a matching resolution is carried out by using regional height difference statistical parameters, and spatial resolution unification and grid registration of the remote sensing images are carried out.

3. The deep learning method for hydraulic engineering slope measurement according to claim 2, characterized in that: The remote sensing images are preprocessed, and a multiscale data set of the remote sensing images and terrain features is constructed, including edge texture, gray gradient and NDVI index, which comprises the following steps: image decomposition processing is carried out on the acquired remote sensing images, and a multiscale image set is constructed based on a Gaussian pyramid; edge texture extraction, gray gradient calculation and normalized difference vegetation index (NDVI) analysis are respectively carried out on the multiscale image set, wherein the edge texture is extracted by using a direction gradient histogram method, the gray gradient is calculated by using a Sobel operator or a Laplace operator, and the NDVI index is obtained based on red light and near-infrared band reflectivity calculation; different scales and different types of image features are fused into a unified multiscale image data set through feature mapping, and the multiscale image data set is used as input samples of the deep neural network model, wherein the feature fusion process adopts weighted fusion.

4. The deep learning method for hydraulic engineering slope measurement according to claim 3, characterized in that: The deep neural network model for slope extraction is constructed based on labeled samples, which comprises the following steps: supervised training sample sets are constructed based on remote sensing images and corresponding slope measurement data, and the sample sets comprise multiscale input image data and corresponding pixel-level slope label maps; a deep neural network model comprising a multilayer convolution structure is constructed, wherein the deep neural network model extracts edge and texture features in a shallow network, extracts spatial structure features in a middle layer, learns high-level semantic and terrain relationship features in a deep layer, and enhances information flow through cross-layer connection; a residual connection structure is introduced in the convolution network, and the residual connection structure adopts an identity mapping method to add the features of a previous layer and output of a next layer element by element.

5. The deep learning method for hydraulic engineering slope measurement according to claim 4, characterized in that: The multiscale data set is input into the trained deep neural network model, and a slope value distribution map of a corresponding landform unit is output, which comprises the following steps: the constructed multiscale remote sensing image and terrain feature data set is subjected to normalization and channel arrangement processing; The pretreated multi-scale data is input into the trained deep neural network model, and in the inference process, it sequentially passes through the shallow feature extraction module, the middle spatial structure analysis module and the deep semantic feature discrimination module, and generates the slope prediction value of each pixel at the output layer; The predicted slope value is spatially reorganized and post-processed, including mapping the pixel-level prediction results to the original geographic coordinates to generate a slope value distribution map.

6. The deep learning method for hydraulic engineering slope measurement according to claim 5, characterized in that: The key sections are fitted and the abnormal points are corrected according to the slope value distribution map to construct the final slope measurement result map, including: Extracting the slope mutation area, the slope junction zone and the engineering control point neighborhood as the key sections, and constructing the corresponding topographic structure boundary index map; Curved surface fitting is performed on the slope prediction results of the key sections; Combined with the fitting results and the neighborhood statistical characteristics, the abnormal points are detected and corrected, wherein the abnormal point judgment is based on the deviation of the predicted value and the fitted value exceeding the deviation threshold, and the correction strategy includes neighborhood mean replacement, fitting interpolation or structure guided reconstruction, and finally the slope measurement result map is generated.

7. The deep learning method for hydraulic engineering slope measurement according to claim 6, characterized in that: The final slope measurement result is compared and evaluated with the historical measurement data, including: Based on spatial coordinate registration and resolution uniform processing, the final slope measurement result map and the corresponding historical measured slope data are aligned at the pixel level to construct a one-to-one comparison data set; On the basis of the comparison data set, error statistical indicators are calculated, including mean absolute error, root mean square error, maximum error and error distribution frequency histogram, and based on the engineering area division, the sub-region error summary information is output.

8. The deep learning method for hydraulic engineering slope measurement according to claim 7, characterized in that: The deep neural network model parameters are continuously optimized according to the error distribution, including: Determine the high error area and the error type, construct the error sensitive area sample set, including the corresponding remote sensing input, multi-scale feature map and labeled slope label; Based on the deep neural network model parameters, the error sensitive area samples are retrained to adjust the intermediate layer feature weight distribution and the loss function parameters in the network; According to the new round of training results, the deep neural network model parameters are updated, and the error threshold is set, and when the error of continuous multiple regions exceeds the set range, the retraining is automatically triggered.

Images