Pumped storage power station environment water conservation identification model training and application method and system

Abstract

The invention provides a pumped storage power station environment water conservation recognition model training and application method and system, and the method comprises the steps: carrying out the sample marking of an obtained multi-temporal satellite remote sensing image in the construction period of a pumped storage power station, and obtaining a corresponding environment water conservation sample marking set; a YOLOv8 network of a feature pyramid network is replaced based on a position-aware coding network to construct a environmental water conservation recognition model, the position-aware coding network is a Transform network integrated with a spatial position coding mechanism, then a multi-temporal satellite remote sensing image with an environmental water conservation sample labeling set is used as a sample set, the environmental water conservation recognition model is trained, and the environmental water conservation recognition model is obtained. The problem of small target feature attenuation of a traditional YOLOv8 network is effectively overcome through a Transformer network integrated with a spatial position coding mechanism, and the defect of misjudgment caused by complex background interference due to the distinguishing ability of a target and a background is enhanced by using an attention mechanism of coordinate coding. And the ring water conservation identification precision of the pumped storage power station is further improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of power facilities, specifically to a training and application method and system for an environmental protection identification model for pumped storage power stations. Background Technology

[0002] As a key infrastructure for the transformation of the national energy structure, the environmental protection and soil and water conservation supervision during the construction of pumped storage power stations are directly related to ecological security and project compliance. With increasingly stringent environmental regulations, the need for real-time, accurate, and intelligent monitoring of environmental protection and soil and water conservation in pumped storage power stations is becoming increasingly prominent. Traditional supervision models relying on manual inspections and periodic spot checks have significant limitations. Complex mountainous terrain leads to insufficient inspection coverage, manual interpretation is highly subjective, and dynamic comparison of data from multiple periods is inefficient, making it difficult to meet the full-process supervision needs of large-scale power station construction. Satellite remote sensing-based environmental protection and soil and water conservation monitoring technology, due to its advantages of wide coverage, periodic observation, and objective data recording, has become a key development direction in the industry. This technology, by acquiring sub-meter high-resolution images, can clearly identify the scope of surface disturbance, the direction of construction access roads, and the form of environmental protection measures, providing data support for supervision. However, existing remote sensing interpretation methods still face core bottlenecks: manual interpretation requires visual interpretation by professionals, which is inefficient; conventional image processing algorithms have limited ability to distinguish environmental protection and soil and water conservation targets with varying shapes, resulting in a high false detection rate.

[0003] In recent years, breakthroughs in deep learning technology have provided new pathways for intelligent remote sensing interpretation. Target recognition networks have made significant progress in ground feature detection by automatically extracting features through convolutional neural networks. Among them, YOLOv8, as the latest real-time detection network, possesses high-speed inference capabilities, but its feature pyramid structure leads to the attenuation of small target feature information during propagation in deep networks, causing problems such as missed detection of construction access roads and blurred disturbance boundaries. While the Transformer attention mechanism can model long-distance dependencies, standard self-attention ignores spatial location information in remote sensing applications, resulting in confusion of similar ground features in mountainous areas and a significant decrease in target confidence under complex background interference. Among current mainstream technologies, attempts to combine YOLOv8 and Transformer are mostly focused on target detection in natural scenes, and have not yet effectively addressed the special challenges of environmental and water conservation monitoring of pumped storage power stations: the extreme differences in target morphology between isal disturbance areas and linear construction access roads; the highly complex background of construction areas and undisturbed mountainous environments leading to overlapping spectral features; and the insufficient automatic identification capability of incremental illegal disturbances in multi-phase images urgently needed for dynamic monitoring. These shortcomings severely restrict the intelligent upgrading of environmental and water conservation supervision, and there is an urgent need to develop a dedicated target recognition method adapted to mountain power station scenarios. Summary of the Invention

[0004] To address the problems of missed detection of small targets and misjudgment of similar features in the complex mountainous terrain of pumped storage power stations, this invention proposes a training and application method and system for an environmental protection identification model for pumped storage power stations.

[0005] Firstly, a method for training an environmental protection identification model for pumped storage power stations is provided, including:

[0006] Acquire multi-temporal satellite remote sensing images during the construction period of a pumped storage power station, and annotate the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images;

[0007] A YOLOv8 network is used to replace the feature pyramid network with a location-aware encoding and decoding network to construct an environmental protection and water conservation identification model. The location-aware encoding and decoding network is a Transformer network that incorporates a spatial location coding mechanism.

[0008] Using multi-temporal satellite remote sensing images with environmental and water conservation sample annotation sets as the sample set, the environmental and water conservation identification model is trained to obtain a trained environmental and water conservation identification model.

[0009] The second part provides a training system for an environmental protection identification model of a pumped storage power station, including:

[0010] The annotation module is used to acquire multi-temporal satellite remote sensing images during the construction period of pumped storage power stations, and to annotate the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images.

[0011] The module is used to construct an environmental protection and water conservation identification model based on a YOLOv8 network that replaces the feature pyramid network with a location-aware encoding and decoding network. The location-aware encoding and decoding network is a Transformer network that incorporates a spatial location coding mechanism.

[0012] The training module is used to train the environmental protection and water conservation identification model using multi-temporal satellite remote sensing images with environmental protection and water conservation sample annotation sets as the sample set, so as to obtain the trained environmental protection and water conservation identification model.

[0013] The third part provides a method for applying an environmental and water conservation identification model for pumped storage power stations, including:

[0014] Multi-temporal remote sensing images of pumped storage power stations are acquired, and the pumped storage power station remote sensing images are preprocessed to obtain processed multi-temporal remote sensing images.

[0015] The multi-temporal remote sensing images are input into a pre-constructed environmental and water conservation identification model to obtain target environmental and water conservation identification results and early warning reports;

[0016] The environmental protection and water conservation identification model is obtained based on the training method for the environmental protection and water conservation identification model of pumped storage power stations as described in any of the preceding claims.

[0017] Part Four provides an application system for an environmental protection and water conservation identification model for pumped storage power stations, including:

[0018] The acquisition module is used to acquire multi-temporal remote sensing images of pumped storage power stations and to preprocess the pumped storage power station remote sensing images to obtain processed multi-temporal remote sensing images.

[0019] The input module is used to input the multi-temporal remote sensing images into a pre-constructed environmental and water conservation identification model to obtain the target environmental and water conservation identification results and early warning reports;

[0020] The environmental protection and water conservation identification model is obtained based on the training method for the environmental protection and water conservation identification model of pumped storage power stations as described in any of the preceding claims.

[0021] In another aspect, this application also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus;

[0022] The memory is used to store one or more programs;

[0023] When the one or more programs are executed by the at least one processor, the above-described method for training a pumped storage power station environmental protection and water conservation identification model and the above-described method for applying the pumped storage power station environmental protection and water conservation identification model are implemented.

[0024] Furthermore, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the above-described method for training a pumped storage power station environmental protection and water conservation identification model and the above-described method for applying the pumped storage power station environmental protection and water conservation identification model.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] This invention provides a training and application method and system for an environmental protection and water conservation identification model for pumped storage power stations. The model training method acquires multi-temporal satellite remote sensing images during the construction period of the pumped storage power station and annotates these images to obtain a corresponding environmental protection and water conservation sample annotation set. An environmental protection and water conservation identification model is constructed based on a YOLOv8 network that replaces the feature pyramid network with a location-aware coding network. The location-aware coding network is a Transformer network incorporating a spatial location coding mechanism. The multi-temporal satellite remote sensing images with the environmental protection and water conservation sample annotation set are then used as the sample set to train the environmental protection and water conservation identification model, resulting in a trained model. The Transformer network incorporating a spatial location coding mechanism effectively overcomes the small target feature attenuation problem of the traditional YOLOv8 network. The attention mechanism of coordinate coding enhances the ability to distinguish between targets and background, solving the misjudgment defect caused by complex background interference, and further improving the accuracy of environmental protection and water conservation identification for pumped storage power stations. Attached Figure Description

[0027] Figure 1 This is a flowchart of the pumped storage power station environmental protection identification model training method of the present invention;

[0028] Figure 2 This is a schematic diagram illustrating the specific process of the pumped storage power station environmental protection identification model training method of the present invention;

[0029] Figure 3 This is a schematic diagram of the Transformer-enhanced YOLOv8 network structure for the pumped storage power station environmental protection identification model training method of the present invention.

[0030] Figure 4 This is a schematic diagram of the structure of the pumped storage power station environmental protection identification model training system of the present invention;

[0031] Figure 5 This is a flowchart of the application method of the environmental protection identification model for pumped storage power stations according to the present invention;

[0032] Figure 6 This is a schematic diagram of the structure of the pumped storage power station environmental protection identification model training system of the present invention;

[0033] Figure 7 This is a schematic diagram of an electronic device structure according to the present invention. Detailed Implementation

[0034] This invention proposes a training and application method and system for an environmental protection identification model in pumped storage power stations, enabling collaborative identification and dynamic compliance analysis of three types of targets: surface disturbance zones, construction access roads, and environmental protection measures. This method overcomes industry challenges such as missed detection of small targets and misjudgment of similar features in complex mountainous environments by constructing a target recognition network that integrates location-aware Transformers.

[0035] To better understand the present invention, the following description, in conjunction with the accompanying drawings and embodiments, will further illustrate the content of the present invention.

[0036] Example 1:

[0037] A method for training an environmental protection identification model for pumped storage power stations, such as... Figure 1 As shown, it includes:

[0038] Step S1: Acquire multi-temporal satellite remote sensing images during the construction period of the pumped storage power station, and perform sample annotation on the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images;

[0039] Step S2: Construct an environmental and water conservation identification model by replacing the feature pyramid network with a YOLOv8 network based on a location-aware encoding / decoding network;

[0040] Step S3: Using multi-temporal satellite remote sensing images with environmental and water conservation sample annotation sets as the sample set, train the environmental and water conservation identification model to obtain the trained environmental and water conservation identification model.

[0041] Among them, the position-aware encoding and decoding network is a Transformer network that incorporates a spatial position encoding mechanism;

[0042] Environmental protection and soil and water conservation can be understood as an abbreviation for environmental protection and soil and water conservation.

[0043] In this embodiment, during the acquisition of multi-temporal satellite remote sensing images during the construction period of the pumped storage power station in step S1, the acquired multi-temporal satellite remote sensing images can be preprocessed to eliminate errors in the sensor-acquired images. Then, sample annotation is performed on the preprocessed multi-temporal satellite remote sensing images to ensure the accuracy of the annotation. The specific implementation process includes:

[0044] The multi-temporal satellite remote sensing images are preprocessed to obtain processed multi-temporal satellite remote sensing images. The preprocessing includes radiometric calibration, atmospheric correction, geometric fine correction, feature matching, and sliding window cropping.

[0045] The surface disturbance areas and environmental protection measures in the processed multi-temporal satellite remote sensing images were labeled using a polygon labeling tool, and the construction lane changes in the processed multi-temporal satellite remote sensing images were labeled using the centerline width labeling method, resulting in an environmental protection and water conservation sample labeling set.

[0046] In one possible embodiment, the standardized preprocessing procedure for multi-temporal satellite remote sensing imagery includes: first, radiometric calibration to eliminate sensor errors; second, atmospheric correction using the FLAASH model; third, geometric fine correction using a polynomial correction method (control point error ≤ 1 pixel); and finally, sub-pixel registration of the multi-temporal images using SIFT feature matching. The processed images are then cropped into 640×640 pixel image blocks using a sliding window, with a 15% overlap rate set to avoid target truncation issues, generating standard input data units.

[0047] Furthermore, the process of sample annotation for the preprocessed multi-temporal satellite remote sensing images includes: (1) Annotation of surface disturbance areas: Using polygon annotation tools, disturbed areas such as construction bare areas and spoil heaps are delineated (RGB:255,0,0), requiring the smallest identification unit to be ≥4 pixels (corresponding to an actual area ≥1m²). 2 (2) Construction access road marking: The temporary construction road is marked using the "centerline + width" marking method (RGB:0,255,0). The access road width is required to be ≥3 pixels (actual width ≥1.5m). Points are taken at 0.5 pixel intervals for curved sections. (3) Environmental protection measures marking: Polygon markings are used for slope protection nets, covered areas, etc. (RGB:0,0,255). The smallest marking unit is a continuous area of ​​10×10 pixels. Preferably, the sample library can be divided into training set, validation set and test set in a ratio of 7:2:1.

[0048] In this embodiment, after preprocessing and labeling the acquired multi-temporal satellite remote sensing images in step S1, the YOLOv8 network feature pyramid network can be replaced by a Transformer network that incorporates a spatial location coding mechanism to construct an environmental and water conservation identification model. In subsequent steps, the trained environmental and water conservation identification network that integrates location perception Transformer can overcome the industry problem of missed detection of small targets and misjudgment of similar features in complex mountainous scenes.

[0049] In this embodiment, in step S3, multi-temporal satellite remote sensing images with environmental and water conservation sample annotation sets are used as the sample set to train the environmental and water conservation identification model, thereby obtaining a trained environmental and water conservation identification model. This model enables accurate identification of targets in multi-temporal satellite remote sensing images, specifically including:

[0050] The training set of the multi-temporal satellite remote sensing image sample set with environmental and water conservation sample annotation set is input into the environmental and water conservation identification model to train the initial environmental and water conservation identification model.

[0051] The validation set in the sample set is input into the initial environmental and water conservation identification model for model validation to obtain the model validation result. Based on the model validation result, the initial environmental and water conservation identification model is optimized to obtain the trained environmental and water conservation identification model.

[0052] The test set in the sample set is input into the trained environmental protection and water conservation identification model to evaluate the model accuracy and obtain the model evaluation result. The performance of the trained environmental protection and water conservation identification model is evaluated based on the model evaluation result.

[0053] In this embodiment, during the process of training the environmental protection and water conservation identification model by inputting the training set of a multi-temporal satellite remote sensing image sample set with environmental protection and water conservation sample annotations, the feature extraction and fusion operation of the training set is achieved through the backbone network, position-aware encoding and decoding network, and neck network in the environmental protection and water conservation identification model. Furthermore, by fusing the position-aware Transformer network, the industry challenges of missed detection of small targets and misjudgment of similar ground features in complex mountainous scenes are overcome. Specifically, this includes:

[0054] The training set is used to extract features through the backbone network of the environmental protection and water conservation identification model to obtain multi-scale feature maps;

[0055] The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information.

[0056] The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map;

[0057] Based on the environmental protection and water conservation identification results and the environmental protection and water conservation sample annotation set, the loss value is calculated using a four-component loss function, and the model parameter loss is optimized based on the loss value to obtain the initial environmental protection and water conservation identification model.

[0058] In this embodiment, the process of obtaining an enhanced feature map that fuses spatial location information by using a two-dimensional spatial location encoding to enhance the spatial relationships of multi-scale feature maps through a location-aware encoding and decoding network of the environmental protection and water conservation identification model includes:

[0059] The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses a sine function to generate the even-dimensional encoding components of the multi-scale feature map, and a cosine function to generate the odd-dimensional encoding components of the multi-scale feature map.

[0060] The even-dimensional coding components and the odd-dimensional coding components are used as the position coding bias matrix, and the multi-head attention weights of each position coding vector in the position coding bias matrix are calculated through the multi-head attention mechanism of the position-aware coding and decoding network.

[0061] An enhanced feature map incorporating spatial location information is generated based on the location encoding bias matrix and the multi-head attention weights of each location encoding vector.

[0062] Specifically, a spatial location encoding mechanism is introduced at the back end of the feature extraction network (i.e., the backbone network). A two-dimensional coordinate code is generated through a sine-cosine function to enhance the network's ability to distinguish complex mountainous backgrounds. The specific calculation formula is shown below:

[0063]

[0064] Where x and y are the two-dimensional coordinates of the pixels on the input feature map; i is the dimension index of the position encoding vector; and d is the embedding dimension of the position encoding, which is consistent with the number of channels in the feature map. For the even-numbered dimension components of the position encoding vector, The location encoding vector has odd-dimensional components, and periodic codes are generated using sine and cosine functions to ensure that the model can distinguish the relative distance and direction of different locations.

[0065] In a specific embodiment, the multi-head attention weight is calculated as follows:

[0066]

[0067] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k P represents the dimension of the key / query vector. pos Here, represents the positional encoding bias matrix, and softmax is the normalization function.

[0068] In this embodiment, during the process of obtaining a fused feature map by fusing features based on the enhanced feature map of the neck network of the environmental protection and water conservation identification model, a cross-scale feature fusion pathway can be adopted to solve the problem of feature attenuation of small targets in the model, which specifically includes:

[0069] The deep features in the enhanced feature map are upsampled four times by bilinear interpolation through the neck network of the environmental protection and water conservation identification model to obtain the upsampled deep features.

[0070] The intermediate layer features in the enhanced feature map are upsampled by a factor of two using bilinear interpolation to obtain the upsampled intermediate layer features.

[0071] Perform a convolution operation on the shallow features in the enhanced feature map to align the number of channels, and obtain the convolved shallow features;

[0072] Based on the upsampled deep features, the upsampled intermediate features, and the convolutional shallow features, a weighted fusion is performed using learnable weight parameters to obtain a fused feature map.

[0073] In this embodiment, after obtaining the fused feature map, automatic comparison can be performed through dynamic monitoring to overcome the shortcomings of traditional methods that rely on manual comparison. Specifically, this includes:

[0074] The dynamic monitoring decoder of the environmental protection and water conservation identification model performs compliance analysis based on the environmental protection and water conservation identification results and the corresponding timestamp information using a preset rule engine to obtain the analysis results. The dynamic monitoring decoder is located below the detection head of the environmental protection and water conservation identification model.

[0075] Determine whether the analysis result exceeds the preset warning conditions, and issue a warning if the analysis result exceeds the preset warning conditions.

[0076] Example 2:

[0077] The following is combined with Figure 2 As shown, the present invention will be described in detail below, and its specific implementation steps are as follows:

[0078] Step 1: Remote Sensing Image Acquisition and Preprocessing. Acquire multi-temporal, sub-meter resolution satellite remote sensing images (recommended resolution ≤ 0.5m) covering the power station construction area and surrounding ecologically sensitive areas during the pumped storage power station construction period. Perform a standardized preprocessing workflow on the raw images: first, perform radiometric calibration to eliminate sensor errors; second, perform atmospheric correction using the FLAASH model; then, perform geometric fine correction using the polynomial correction method (control point error ≤ 1 pixel); finally, achieve sub-pixel registration of multi-temporal images using SIFT feature matching. Crop the processed images into 640×640 pixel image blocks using a sliding window, setting a 15% overlap rate to avoid target truncation issues, and generate standard input data units.

[0079] Step 2, Construction of Environmental Protection and Water Conservation Target Sample Library. The image blocks generated in Step 1 are finely manually annotated to establish three types of sample libraries for environmental protection and water conservation monitoring: (1) Annotation of surface disturbance areas: Use polygon annotation tools to delineate disturbed areas such as construction bare areas and spoil heaps (RGB:255,0,0), requiring the smallest identification unit to be ≥4 pixels (corresponding to an actual area ≥1m²). 2(2) Construction access road marking: The temporary construction road is marked using the "centerline + width" marking method (RGB:0,255,0). The access road width is required to be ≥3 pixels (actual width ≥1.5m). For curved sections, points are taken at 0.5 pixel intervals. (3) Environmental protection measures marking: Polygon markings are used for slope protection nets, covered areas, etc. (RGB:0,0,255). The smallest marking unit is a continuous area of ​​10×10 pixels. The sample library is divided into training set, validation set and test set in a ratio of 7:2:1.

[0080] Step 3: Construct the Transformer-enhanced YOLOv8 network (i.e., the aforementioned environmental and water conservation identification model). For example... Figure 3 As shown, (1) the position-aware Transformer module (replacing the original FPN). The position-aware Transformer module is connected after the multi-scale feature map output by the Backbone. This module strengthens the spatial relationship of the target through two-dimensional spatial position encoding:

[0081]

[0082] Where x and y are the two-dimensional coordinates of pixels on the input feature map; i is the dimension index of the positional encoding vector; d is the embedding dimension of the positional encoding, consistent with the number of channels in the feature map; the left side of the equation represents the even and odd dimension components of the positional encoding vector, respectively. Periodic encoding is generated using sine and cosine functions to ensure the model can distinguish the relative distance and direction at different positions. The improved multi-head attention calculation formula is as follows:

[0083]

[0084] Where Q is the query matrix, generated from the input feature map through a linear transformation, used to calculate the correlation with other locations; K and V are the key matrix and value matrix, respectively, also obtained through a linear transformation of the input features; d k is the dimension of the key / query vector, used to scale the dot product (to prevent gradient explosion); P pos The location encoding bias matrix is ​​generated by the aforementioned location encoding formula; softmax is a normalization function that maps the attention weights to the [0,1] interval, ensuring that the sum of the weights is 1. This enables the network to accurately distinguish targets based on spatial distribution differences even in scenarios where the spectral overlap between bare rock and construction disturbance areas is >60%.

[0085] (2) Multi-scale feature interaction mechanism. A cross-scale feature fusion path is designed to solve the feature attenuation problem of small targets in YOLOv8: P3 layer features are directly aligned to the number of channels through 1×1 convolution; P4 layer features are upsampled by 2 times through bilinear interpolation; P5 layer features are upsampled by 4 times through bilinear interpolation; the fusion formula is as follows:

[0086]

[0087] Among them, P3, P4, and P5 are multi-scale feature maps from the YOLOv8 backbone network; the initial values ​​of the learning parameters α, β, and γ are 0.4 / 0.3 / 0.3, and the contribution weights of features at different scales are dynamically optimized, which improves the recognition rate of 3-pixel wide construction access roads to 91.2%.

[0088] (3) Dynamic monitoring decoder. A compliance analysis engine is added after the detection head. The target recognition result and timestamp information are input, and the rule engine realizes automatic early warning. This module fills the gap in dynamic supervision that traditional methods rely on manual comparison.

[0089] Step 4, Network Training and Optimization. End-to-end training using a four-component loss function. Key training parameters include: input size 640×640; batch size = 16 (4×Tesla V100); AdamW optimizer (β1 = 0.9, β2 = 0.999); initial learning rate 1e-4 with cosine decay strategy; early stopping mechanism triggered when there is no improvement in mAP@0.5 for 5 consecutive rounds on the validation set. Mosaic data augmentation is applied during training, randomly stitching together 4 images to simulate complex background interference, improving model robustness.

[0090] Example 3:

[0091] To clearly illustrate the implementation process of the technical solution of this invention, a case study of actual monitoring of a pumped storage power station under construction is presented below. This power station is located in a typical ecologically sensitive mountainous area, with a long construction period and a wide disturbance range. The specific implementation steps are as follows:

[0092] Step 1: Remote Sensing Image Acquisition and Preprocessing. Acquire four high-resolution satellite images (0.5-meter resolution) covering the power plant construction period from 2023 to 2024. Perform a standardized preprocessing procedure on the raw images: first, perform radiometric calibration to eliminate sensor errors; second, use an atmospheric correction model to remove aerosol interference; third, perform geometric fine correction to ensure spatial positioning accuracy (control point error ≤ 1 pixel); finally, perform multi-temporal image registration. Crop the processed images into 640×640 pixel sliding windows with a 15% overlap to avoid segmenting key targets.

[0093] Step 2: Construction of the Environmental Protection and Water Conservation Target Sample Library. The cropped image blocks are manually annotated with fine detail: Surface disturbance areas: Exposed construction areas, spoil heaps, etc., are delineated using polygon annotation tools, with a minimum annotation unit of 4 pixels (≥1 square meter in actual area); Construction access roads: The centerline annotation method is used, requiring access road width ≥3 pixels (≥1.5 meters in actual area), with points taken at 0.5 pixel intervals for curved sections; Environmental protection measures: Slope protection nets, covered areas, etc., are annotated, with a minimum continuous unit of 10×10 pixels. The annotated sample library is divided into training, validation, and test sets in a 7:2:1 ratio.

[0094] Step 3: Enhance YOLOv8 network training with Transformer. Construct an improved YOLOv8-T network architecture: A location-aware Transformer module: Introduces a spatial location encoding mechanism at the backend of the feature extraction network, generating two-dimensional coordinate codes through a sine-cosine function to enhance the network's ability to distinguish complex mountainous backgrounds; Multi-scale feature interaction mechanism: Direct convolution of shallow feature map P3, upsampling of mid-layer feature P4 by 2 times, and upsampling of deep feature P5 by 4 times, using learnable weight parameters for weighted fusion; Dynamic monitoring decoder: Built-in compliance analysis rules automatically trigger an alert when the detected disturbance area increase exceeds 30%. Network training uses a four-component loss function, with a fixed input size of 640×640, a batch size of 16, and end-to-end training using the AdamW optimizer.

[0095] Step 4: Intelligent Monitoring of Environmental Protection and Water Conservation. The latest 2024 imagery is input into the trained network, which outputs three types of structured recognition results: location coordinates and area data of the surface disturbance zone; direction, length, and width information of construction access roads; and type and distribution range of environmental protection measures. The system automatically compares data from multiple periods and generates real-time warnings for areas with disturbance exceeding limits (increase > 30%) and unregistered construction access roads.

[0096] This invention also provides an intelligent monitoring system for environmental and water conservation, comprising four core modules: a data preprocessing module, which performs image radiometric calibration, atmospheric correction, geometric fine correction, and registration and cropping; a sample annotation module, which provides interactive annotation tools and supports digital mapping of three types of targets: surface disturbance areas, construction access roads, and environmental protection measures; a model training module, which deploys a Transformer-enhanced YOLOv8 network to support distributed training and parameter optimization; and a dynamic monitoring module, which performs target identification and compliance analysis, and automatically outputs structured results and early warning reports.

[0097] The modules work together to achieve closed-loop management of the entire process from data input to regulatory decision-making. This embodiment verifies the effectiveness of the method in complex mountainous scenarios: the identification accuracy of construction access roads reaches 93.7%, the false judgment rate of disturbed area is reduced to 7.3%, and the response time for identifying illegal disturbances is shortened to the minute level, which is significantly better than the traditional manual verification mode.

[0098] Example 4:

[0099] A training system for an environmental protection identification model of a pumped storage power station, such as Figure 4 ,include:

[0100] The annotation module is used to acquire multi-temporal satellite remote sensing images during the construction period of pumped storage power stations, and to annotate the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images.

[0101] The module is used to construct an environmental protection and water conservation identification model based on a YOLOv8 network that replaces the feature pyramid network with a location-aware encoding and decoding network. The location-aware encoding and decoding network is a Transformer network that incorporates a spatial location coding mechanism.

[0102] The training module is used to train the environmental protection and water conservation identification model using multi-temporal satellite remote sensing images with environmental protection and water conservation sample annotation sets as the sample set, so as to obtain the trained environmental protection and water conservation identification model.

[0103] Preferably, the training module is specifically used for:

[0104] The training set of the multi-temporal satellite remote sensing image sample set with environmental and water conservation sample annotation set is input into the environmental and water conservation identification model to train the initial environmental and water conservation identification model.

[0105] The validation set in the sample set is input into the initial environmental and water conservation identification model for model validation to obtain the model validation result. Based on the model validation result, the initial environmental and water conservation identification model is optimized to obtain the trained environmental and water conservation identification model.

[0106] The test set in the sample set is input into the trained environmental protection and water conservation identification model to evaluate the model accuracy and obtain the model evaluation result. The performance of the trained environmental protection and water conservation identification model is evaluated based on the model evaluation result.

[0107] Preferably, the training process of the initial environmental and water conservation identification model in the training module includes:

[0108] The training set is used to extract features through the backbone network of the environmental protection and water conservation identification model to obtain multi-scale feature maps;

[0109] The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information.

[0110] The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map;

[0111] Based on the environmental protection and water conservation identification results and the environmental protection and water conservation sample annotation set, the loss value is calculated using a four-component loss function, and the model parameter loss is optimized based on the loss value to obtain the initial environmental protection and water conservation identification model.

[0112] Preferably, in the training module, the location-aware encoding / decoding network of the environmental and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationships of the multi-scale feature map, obtaining an enhanced feature map that fuses spatial location information, including:

[0113] The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses a sine function to generate the even-dimensional encoding components of the multi-scale feature map, and a cosine function to generate the odd-dimensional encoding components of the multi-scale feature map.

[0114] The even-dimensional coding components and the odd-dimensional coding components are used as the position coding bias matrix, and the multi-head attention weights of each position coding vector in the position coding bias matrix are calculated through the multi-head attention mechanism of the position-aware coding and decoding network.

[0115] An enhanced feature map incorporating spatial location information is generated based on the location encoding bias matrix and the multi-head attention weights of each location encoding vector.

[0116] Preferably, the calculation formula for the multi-head attention weights in the training module is as follows:

[0117]

[0118] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k P represents the dimension of the key / query vector. pos Here, represents the positional encoding bias matrix, and softmax is the normalization function.

[0119] Preferably, in the training module, the neck network of the environmental conservation and water conservation identification model performs feature fusion based on the enhanced feature map to obtain a fused feature map, including:

[0120] The deep features in the enhanced feature map are upsampled four times by bilinear interpolation through the neck network of the environmental protection and water conservation identification model to obtain the upsampled deep features.

[0121] The intermediate layer features in the enhanced feature map are upsampled by a factor of two using bilinear interpolation to obtain the upsampled intermediate layer features.

[0122] Perform a convolution operation on the shallow features in the enhanced feature map to align the number of channels, and obtain the convolved shallow features;

[0123] Based on the upsampled deep features, the upsampled intermediate features, and the convolutional shallow features, a weighted fusion is performed using learnable weight parameters to obtain a fused feature map.

[0124] Preferably, the annotation module performs sample annotation on the multi-temporal satellite remote sensing images to obtain a water and environmental protection sample annotation set corresponding to the multi-temporal satellite remote sensing images, including:

[0125] The multi-temporal satellite remote sensing images are preprocessed to obtain processed multi-temporal satellite remote sensing images. The preprocessing includes radiometric calibration, atmospheric correction, geometric fine correction, feature matching, and sliding window cropping.

[0126] The surface disturbance areas and environmental protection measures in the processed multi-temporal satellite remote sensing images were labeled using a polygon labeling tool, and the construction lane changes in the processed multi-temporal satellite remote sensing images were labeled using the centerline width labeling method, resulting in an environmental protection and water conservation sample labeling set.

[0127] Preferably, the system further includes:

[0128] The early warning module is used to perform compliance analysis using a preset rule engine based on the environmental protection and water conservation identification results and the corresponding timestamp information through the dynamic monitoring decoder of the environmental protection and water conservation identification model, and obtain the analysis results. The dynamic monitoring decoder is located in the lower layer of the detection head of the environmental protection and water conservation identification model.

[0129] Determine whether the analysis result exceeds the preset warning conditions, and issue a warning if the analysis result exceeds the preset warning conditions.

[0130] Example 5:

[0131] A method for applying an environmental protection identification model for pumped storage power stations, such as... Figure 5 As shown, it includes:

[0132] Step S1: Acquire multi-temporal remote sensing images of the pumped storage power station, and preprocess the pumped storage power station remote sensing images to obtain processed multi-temporal remote sensing images.

[0133] Step S2: Input the multi-temporal remote sensing images into the pre-constructed environmental protection and water conservation identification model to obtain the target environmental protection and water conservation identification results and early warning reports;

[0134] Among them, the environmental protection and water conservation identification model is obtained based on the training method of the pumped storage power station environmental protection and water conservation identification model as described in any of the preceding items.

[0135] In this embodiment, the process of inputting multi-temporal remote sensing images into a pre-constructed environmental and water conservation identification model to obtain target environmental and water conservation identification results and early warning reports includes:

[0136] The backbone network of the environmental and water conservation identification model is used to extract features from the input multi-temporal remote sensing images to obtain multi-scale feature maps.

[0137] The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information.

[0138] The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map;

[0139] The dynamic monitoring decoder of the environmental protection and water conservation identification model performs compliance analysis based on the environmental protection and water conservation identification results and the corresponding timestamp information using a preset rule engine to obtain the analysis results, and generates an early warning report based on the analysis results.

[0140] Example 6:

[0141] A method for applying an environmental protection identification model for pumped storage power stations, such as... Figure 6 As shown, it includes:

[0142] The acquisition module is used to acquire multi-temporal remote sensing images of pumped storage power stations and to preprocess the pumped storage power station remote sensing images to obtain processed multi-temporal remote sensing images.

[0143] The input module is used to input the multi-temporal remote sensing images into a pre-constructed environmental and water conservation identification model to obtain the target environmental and water conservation identification results and early warning reports;

[0144] The environmental protection and water conservation identification model is obtained based on the training method for the environmental protection and water conservation identification model of pumped storage power stations as described in any of the preceding claims.

[0145] Preferably, the input module is specifically used for:

[0146] The backbone network of the environmental and water conservation identification model is used to extract features from the input multi-temporal remote sensing images to obtain multi-scale feature maps.

[0147] The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information.

[0148] The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map;

[0149] The dynamic monitoring decoder of the environmental protection and water conservation identification model performs compliance analysis based on the environmental protection and water conservation identification results and the corresponding timestamp information using a preset rule engine to obtain the analysis results, and generates an early warning report based on the analysis results.

[0150] Example 7:

[0151] like Figure 7 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0152] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the pumped storage power station environmental protection identification model training method and the pumped storage power station environmental protection identification model application method as described above.

[0153] Example 8

[0154] Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device, used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the pumped-storage power station environmental protection identification model training method and the pumped-storage power station environmental protection identification model application method as described above.

[0155] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0156] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0158] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0159] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A training method for an environmental protection identification model of a pumped storage power station, characterized in that, include: Acquire multi-temporal satellite remote sensing images during the construction period of a pumped storage power station, and annotate the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images; A YOLOv8 network is used to replace the feature pyramid network with a location-aware encoding and decoding network to construct an environmental protection and water conservation identification model. The location-aware encoding and decoding network is a Transformer network that incorporates a spatial location coding mechanism. Using multi-temporal satellite remote sensing images with environmental and water conservation sample annotation sets as the sample set, the environmental and water conservation identification model is trained to obtain a trained environmental and water conservation identification model.

2. The method according to claim 1, characterized in that, The step of using multi-temporal satellite remote sensing images with environmental and water conservation sample annotation sets as the sample set to train the environmental and water conservation identification model, resulting in a trained environmental and water conservation identification model, includes: The training set of the multi-temporal satellite remote sensing image sample set with environmental and water conservation sample annotation set is input into the environmental and water conservation identification model to train the initial environmental and water conservation identification model. The validation set in the sample set is input into the initial environmental and water conservation identification model for model validation to obtain the model validation result. Based on the model validation result, the initial environmental and water conservation identification model is optimized to obtain the trained environmental and water conservation identification model. The test set in the sample set is input into the trained environmental protection and water conservation identification model to evaluate the model accuracy and obtain the model evaluation result. The performance of the trained environmental protection and water conservation identification model is evaluated based on the model evaluation result.

3. The method according to claim 2, characterized in that, The training process of the initial environmental and water conservation identification model includes: The training set is used to extract features through the backbone network of the environmental protection and water conservation identification model to obtain multi-scale feature maps; The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information. The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map; Based on the environmental protection and water conservation identification results and the environmental protection and water conservation sample annotation set, the loss value is calculated using a four-component loss function, and the model parameter loss is optimized based on the loss value to obtain the initial environmental protection and water conservation identification model.

4. The method according to claim 3, characterized in that, The location-aware encoding / decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationships of the multi-scale feature map, resulting in an enhanced feature map that fuses spatial location information, including: The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses a sine function to generate the even-dimensional encoding components of the multi-scale feature map, and a cosine function to generate the odd-dimensional encoding components of the multi-scale feature map. The even-dimensional coding components and the odd-dimensional coding components are used as the position coding bias matrix, and the multi-head attention weights of each position coding vector in the position coding bias matrix are calculated through the multi-head attention mechanism of the position-aware coding and decoding network. An enhanced feature map incorporating spatial location information is generated based on the location encoding bias matrix and the multi-head attention weights of each location encoding vector.

5. The method according to claim 4, characterized in that, The formula for calculating the multi-head attention weight is as follows: Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k P represents the dimension of the key / query vector. pos Here, represents the positional encoding bias matrix, and softmax is the normalization function.

6. The method according to claim 3, characterized in that, The process of fusing features in the neck network of the environmental protection and water conservation identification model based on the enhanced feature map to obtain a fused feature map includes: The deep features in the enhanced feature map are upsampled four times by bilinear interpolation through the neck network of the environmental protection and water conservation identification model to obtain the upsampled deep features. The intermediate layer features in the enhanced feature map are upsampled by a factor of two using bilinear interpolation to obtain the upsampled intermediate layer features. Perform a convolution operation on the shallow features in the enhanced feature map to align the number of channels, and obtain the convolved shallow features; Based on the upsampled deep features, the upsampled intermediate features, and the convolutional shallow features, a weighted fusion is performed using learnable weight parameters to obtain a fused feature map.

7. The method according to claim 1, characterized in that, The step of annotating the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images includes: The multi-temporal satellite remote sensing images are preprocessed to obtain processed multi-temporal satellite remote sensing images. The preprocessing includes radiometric calibration, atmospheric correction, geometric fine correction, feature matching, and sliding window cropping. The surface disturbance areas and environmental protection measures in the processed multi-temporal satellite remote sensing images were labeled using a polygon labeling tool, and the construction lane changes in the processed multi-temporal satellite remote sensing images were labeled using the centerline width labeling method, resulting in an environmental protection and water conservation sample labeling set.

8. The method according to claim 3, characterized in that, After generating the environmental protection and water conservation identification result based on the fused feature map, the process also includes: The dynamic monitoring decoder of the environmental protection and water conservation identification model performs compliance analysis based on the environmental protection and water conservation identification results and the corresponding timestamp information using a preset rule engine to obtain the analysis results. The dynamic monitoring decoder is located below the detection head of the environmental protection and water conservation identification model. Determine whether the analysis result exceeds the preset warning conditions, and issue a warning if the analysis result exceeds the preset warning conditions.

9. A training system for an environmental protection identification model of a pumped storage power station, characterized in that, include: The annotation module is used to acquire multi-temporal satellite remote sensing images during the construction period of pumped storage power stations, and to annotate the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images. The module is used to construct an environmental protection and water conservation identification model based on a YOLOv8 network that replaces the feature pyramid network with a location-aware encoding and decoding network. The location-aware encoding and decoding network is a Transformer network that incorporates a spatial location coding mechanism. The training module is used to train the environmental protection and water conservation identification model using multi-temporal satellite remote sensing images with environmental protection and water conservation sample annotation sets as the sample set, so as to obtain the trained environmental protection and water conservation identification model.

10. The system according to claim 9, characterized in that, The training module is specifically used for: The training set of the multi-temporal satellite remote sensing image sample set with environmental and water conservation sample annotation set is input into the environmental and water conservation identification model to train the initial environmental and water conservation identification model. The validation set in the sample set is input into the initial environmental and water conservation identification model for model validation to obtain the model validation result. Based on the model validation result, the initial environmental and water conservation identification model is optimized to obtain the trained environmental and water conservation identification model. The test set in the sample set is input into the trained environmental protection and water conservation identification model to evaluate the model accuracy and obtain the model evaluation result. The performance of the trained environmental protection and water conservation identification model is evaluated based on the model evaluation result.

11. The system according to claim 10, characterized in that, The training process of the initial environmental and water conservation identification model in the training module includes: The training set is used to extract features through the backbone network of the environmental protection and water conservation identification model to obtain multi-scale feature maps; The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information. The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map; Based on the environmental protection and water conservation identification results and the environmental protection and water conservation sample annotation set, the loss value is calculated using a four-component loss function, and the model parameter loss is optimized based on the loss value to obtain the initial environmental protection and water conservation identification model.

12. The system according to claim 11, characterized in that, In the training module, the location-aware encoding / decoding network of the environmental and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationships of the multi-scale feature map, resulting in an enhanced feature map that fuses spatial location information, including: The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses a sine function to generate the even-dimensional encoding components of the multi-scale feature map, and a cosine function to generate the odd-dimensional encoding components of the multi-scale feature map. The even-dimensional coding components and the odd-dimensional coding components are used as the position coding bias matrix, and the multi-head attention weights of each position coding vector in the position coding bias matrix are calculated through the multi-head attention mechanism of the position-aware coding and decoding network. An enhanced feature map incorporating spatial location information is generated based on the location encoding bias matrix and the multi-head attention weights of each location encoding vector.

13. The system according to claim 12, characterized in that, The calculation formula for the multi-head attention weights in the training module is as follows: Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k P represents the dimension of the key / query vector. pos Here, represents the positional encoding bias matrix, and softmax is the normalization function.

14. The system according to claim 11, characterized in that, In the training module, the neck network of the environmental protection and water conservation identification model performs feature fusion based on the enhanced feature map to obtain a fused feature map, including: The deep features in the enhanced feature map are upsampled four times by bilinear interpolation through the neck network of the environmental protection and water conservation identification model to obtain the upsampled deep features. The intermediate layer features in the enhanced feature map are upsampled by a factor of two using bilinear interpolation to obtain the upsampled intermediate layer features. Perform a convolution operation on the shallow features in the enhanced feature map to align the number of channels, and obtain the convolved shallow features; Based on the upsampled deep features, the upsampled intermediate features, and the convolutional shallow features, a weighted fusion is performed using learnable weight parameters to obtain a fused feature map.

15. The system according to claim 9, characterized in that, The annotation module performs sample annotation on the multi-temporal satellite remote sensing images to obtain the environmental and water conservation sample annotation set corresponding to the multi-temporal satellite remote sensing images, including: The multi-temporal satellite remote sensing images are preprocessed to obtain processed multi-temporal satellite remote sensing images. The preprocessing includes radiometric calibration, atmospheric correction, geometric fine correction, feature matching, and sliding window cropping. The surface disturbance areas and environmental protection measures in the processed multi-temporal satellite remote sensing images were labeled using a polygon labeling tool, and the construction lane changes in the processed multi-temporal satellite remote sensing images were labeled using the centerline width labeling method, resulting in an environmental protection and water conservation sample labeling set.

16. The system according to claim 11, characterized in that, The system also includes: The early warning module is used to perform compliance analysis using a preset rule engine based on the environmental protection and water conservation identification results and the corresponding timestamp information through the dynamic monitoring decoder of the environmental protection and water conservation identification model, and obtain the analysis results. The dynamic monitoring decoder is located in the lower layer of the detection head of the environmental protection and water conservation identification model. Determine whether the analysis result exceeds the preset warning conditions, and issue a warning if the analysis result exceeds the preset warning conditions.

17. A method for applying an environmental protection identification model for pumped storage power stations, comprising: Multi-temporal remote sensing images of pumped storage power stations are acquired, and the pumped storage power station remote sensing images are preprocessed to obtain processed multi-temporal remote sensing images. The multi-temporal remote sensing images are input into a pre-constructed environmental and water conservation identification model to obtain target environmental and water conservation identification results and early warning reports; The environmental protection and water conservation identification model is obtained based on the training method for the environmental protection and water conservation identification model of pumped storage power stations as described in any one of claims 1-8.

18. The method according to claim 17, characterized in that, The process of inputting the multi-temporal remote sensing images into a pre-constructed environmental and water conservation identification model to obtain target environmental and water conservation identification results and early warning reports includes: The backbone network of the environmental and water conservation identification model is used to extract features from the input multi-temporal remote sensing images to obtain multi-scale feature maps. The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information. The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map; The dynamic monitoring decoder of the environmental protection and water conservation identification model performs compliance analysis based on the environmental protection and water conservation identification results and the corresponding timestamp information using a preset rule engine to obtain the analysis results, and generates an early warning report based on the analysis results.

19. An application system for an environmental protection identification model of a pumped storage power station, comprising: The acquisition module is used to acquire multi-temporal remote sensing images of pumped storage power stations and to preprocess the pumped storage power station remote sensing images to obtain processed multi-temporal remote sensing images. The input module is used to input the multi-temporal remote sensing images into a pre-constructed environmental and water conservation identification model to obtain the target environmental and water conservation identification results and early warning reports; The environmental protection and water conservation identification model is obtained based on the training method for the environmental protection and water conservation identification model of pumped storage power stations as described in any one of claims 1-8.

20. The system according to claim 19, characterized in that, The input module is specifically used for: The backbone network of the environmental and water conservation identification model is used to extract features from the input multi-temporal remote sensing images to obtain multi-scale feature maps. The location-aware encoding and decoding network of the environmental protection and water conservation identification model uses two-dimensional spatial location encoding to enhance the spatial relationship of the multi-scale feature map, thereby obtaining an enhanced feature map that integrates spatial location information. The neck network of the environmental protection and water conservation identification model is used to perform feature fusion based on the enhanced feature map to obtain a fused feature map, and the environmental protection and water conservation identification result is generated based on the fused feature map; The dynamic monitoring decoder of the environmental protection and water conservation identification model performs compliance analysis based on the environmental protection and water conservation identification results and the corresponding timestamp information using a preset rule engine to obtain the analysis results, and generates an early warning report based on the analysis results.

21. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the method for training the environmental protection and water conservation identification model of a pumped storage power station as described in any one of claims 1 to 8 and the method for applying the environmental protection and water conservation identification model of a pumped storage power station as described in any one of claims 17-18 are implemented.

22. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements the pumped storage power station environmental protection and water conservation identification model training method as described in any one of claims 1 to 8 and the pumped storage power station environmental protection and water conservation identification model application method as described in any one of claims 17-18.

Images