A Calculation Method for Red Mud Volume Monitoring Based on Optical Satellite Imagery Data and Multi-Task Deep Learning Model

Abstract

This invention discloses a method for monitoring and calculating red mud volume based on optical imagery and a multi-task model. The method includes: S1, constructing a multi-task deep learning model containing a main network and a sub-network; acquiring optical satellite imagery data of the study area and inputting it into the multi-task deep learning model; the main network first extracts deep feature maps, and then extracts deep semantic feature maps through a deep semantic extraction network; S2, the sub-network first extracts shallow feature maps, and a multi-head convolutional module performs concatenation and fusion of deep and shallow features to obtain a fused feature map; S3, merging the semantic segmentation results and height reconstruction results of red mud identification using the fused feature map and the deep semantic feature map to output three-dimensional data of red mud features; S4, calculating the red mud accumulation volume of the study area using the three-dimensional data of red mud features. This invention employs parallel interactive fusion processing of the main network and the sub-network to achieve coarse localization, precise localization, and height reconstruction, thereby improving the overall recognition accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of alumina solid waste identification and volume monitoring, and in particular to a method for calculating the volume of red mud based on optical satellite imagery data and a multi-task deep learning model. Background Technology

[0002] Red mud is a highly alkaline solid waste discharged during alumina production, characterized by large output, continuously increasing stockpiles, high alkalinity, and complex heavy metal content. Long-term open-air stockpiling not only occupies land resources but also easily leads to alkaline leachate spills, dust diffusion, and heavy metal migration, posing potential risks to the surrounding soil and water environment. Therefore, accurate and rapid target identification and volume monitoring of red mud stockpiles has become an important requirement for environmental supervision and safety assessment. Current red mud stockpile monitoring mainly relies on the following methods: 1) Manual inspection and ground measurement: Relying on manual sampling, total station or GPS measurements of the stockpile boundaries and elevations, this method is inefficient, time-consuming, costly, and difficult to achieve large-scale, continuous dynamic monitoring. 2) Airborne photogrammetry and LiDAR measurement: This can acquire high-precision three-dimensional point clouds or digital surface models (DSMs), but the data acquisition and processing process is complex, requiring high levels of expertise from equipment and operators, and is prone to data loss and noise issues. In summary, current technologies cannot utilize satellite imagery to achieve low-cost and efficient volume monitoring and accurate calculation of red mud, alumina solid waste, and cannot achieve continuous red mud volume monitoring and quantitative calculation over large geographical areas. Summary of the Invention

[0003] The purpose of this invention is to solve the technical problems pointed out in the background art and provide a method for red mud volume monitoring and calculation based on optical satellite image data and a multi-task deep learning model. The main network of the multi-task deep learning model is responsible for obtaining various information about red mud by using high-level features containing more semantic information and extracting and processing them to obtain three-layer deep feature maps P1 to P3. The sub-network performs local feature extraction through shallower convolution and focuses on extracting shallow feature map P4 from the details. The sub-network can effectively improve the recognition accuracy. The semantic segmentation results and height reconstruction results of red mud recognition by fusing feature maps R1 to R3 and deep semantic feature maps Q1 to Q3 are merged and processed to output three-dimensional data of red mud features.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for calculating red mud volume monitoring based on optical satellite imagery data and a multi-task deep learning model, the method comprising:

[0006] S1. Construct a multi-task deep learning model containing a main network and a sub-network. Obtain optical satellite image data of the study area and input it into the multi-task deep learning model. The main network first extracts three levels of deep feature maps P1 to P3. Then, the deep feature maps P1 to P3 are extracted by the deep semantic extraction network to obtain deep semantic feature maps Q1 to Q3.

[0007] S2. The sub-network first extracts the shallow feature map P4. The multi-head convolution module of the sub-network concatenates and fuses the deep feature maps P1 to P3 with the shallow feature map P4 to obtain the fused feature maps R1 to R3.

[0008] S3. Merge the semantic segmentation results and height reconstruction results of the fused feature maps R1~R3 and the deep semantic feature maps Q1~Q3 for red mud identification, and output the three-dimensional data of red mud features.

[0009] S4. Construct a sample set of optical satellite imagery data containing labeled data and input it into a multi-task deep learning model for training on red mud identification and height reconstruction using methods S1 to S3. Calculate the red mud accumulation volume in the study area using the following formula based on the three-dimensional data of red mud features:

[0010] Where n is the pixel index of the red mud feature in the three-dimensional data, and N is the total number of red mud pixels in the three-dimensional data. For the length data of pixel n, For the width data of cell n, Here is the height data for pixel n.

[0011] To better implement this invention, in method S1, the main network extracts deep feature maps P1 to P3 as follows:

[0012] S11. The main network includes a deep feature extraction module D1 to D3. The deep feature extraction module D1 extracts and processes the optical satellite image data through two shifted convolutional layers SConv and a random depth feature extraction module RDFE, and compresses the spatial scale to obtain a deep feature map P1.

[0013] S12, the deep feature extraction module D2 extracts and processes the deep feature map P1 through a shifted convolutional layer SConv and a random depth feature extraction module RDFE, and compresses the spatial scale to obtain the deep feature map P2;

[0014] S13, the deep feature extraction module D3 processes the deep feature map P2 sequentially through a shifted convolutional layer SConv, a random depth feature extraction module RDFE, and a spatial pyramid pooling fusion module SPPF, and compresses the spatial scale to obtain the deep feature map P3.

[0015] Preferably, the method for the main network to extract deep semantic feature maps Q1 to Q3 is as follows: the deep feature maps P1 to P3 are processed by the feature pyramid network FPN to obtain deep semantic feature maps Q1 to Q3 of three target feature levels. Each target feature level is processed by the result header to obtain the scene classification, object detection, semantic segmentation and height reconstruction results.

[0016] Preferably, the shifted convolutional layer SConv consists of a separable convolutional module, a deep convolutional module, a batch normalization layer, and a LeakReLU layer; the random deep feature extraction module RDFE includes two configurable residual connection network modules DDM, one of which has its residual connection shortcut set to False, and the other has its residual connection shortcut set to True; the spatial pyramid pooling fusion module SPPF uses multi-scale pooling operations to allow the deep feature map P3 to acquire global information at different resolutions while preserving detailed features in the image.

[0017] Preferably, in method S2, the method for the sub-network to extract the shallow feature map P4 is as follows: the shallow feature map P4 is obtained by extracting and processing the feature map through two shifted convolutional layers SConv and a random depth feature extraction module RDFE and compressing the spatial scale.

[0018] Preferably, in method S2, the deep feature maps P1 to P3 and the shallow feature map P4 are subjected to multi-scale global information fusion, channel splicing and fusion processing by the multi-head convolution module MSG-Attention to obtain the fused feature map R0.

[0019] Preferably, the fused feature map R0 is processed by a random depth feature extraction module RDFE to obtain the fused feature map R1; the fused feature map R1 is processed by a shifted convolutional layer SConv, a skip connection deep feature map P2, and a random depth feature extraction module RDFE to obtain the fused feature map R2; the fused feature map R2 is processed by a shifted convolutional layer SConv, a skip connection deep feature map P3, and a random depth feature extraction module RDFE to obtain the fused feature map R3; the fused feature maps R1 to R3 are obtained through three target feature levels, and each target feature level is processed by a result header to obtain the scene classification, object detection, semantic segmentation, and height reconstruction results.

[0020] Preferably, the fused feature maps R1-R3 and the deep semantic feature maps Q1-Q3 comprise a total of six feature maps. Each feature map is processed to obtain a scene classification result, an object detection result, a semantic segmentation result, and a height reconstruction result. The six scene classification results and the six semantic segmentation results are merged using the DS evidence fusion method. The six height reconstruction results are merged by averaging. The six object detection results are merged using a game theory approach. After merging the fused feature maps R1-R3 and the deep semantic feature maps Q1-Q3, a red mud feature map is obtained, which is three-dimensional data of red mud features.

[0021] Preferably, the optical satellite image data and optical satellite image data samples input to the multi-task deep learning model have undergone preprocessing including image denoising and size normalization cropping. The optical satellite image data samples in the optical satellite image data sample set have undergone sample data enhancement processing including rotation, mirror flipping, random cropping, and scaling. The optical satellite image data samples also include data samples with adjustments to brightness, contrast, and saturation and the addition of data samples under different lighting conditions.

[0022] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0023] (1) The main network of the multi-task deep learning model of the present invention is responsible for obtaining various information of red mud by using high-level features containing more semantic information and extracting and processing them to obtain three-layer deep feature maps P1 to P3. The sub-network performs local feature extraction through shallower convolution and focuses on extracting shallow feature map P4 in detail. The sub-network can effectively improve the recognition accuracy. The semantic segmentation result and height reconstruction result of red mud recognition by merging the fused feature map R1 to R3 and the deep semantic feature map Q1 to Q3 are processed and output as three-dimensional data of red mud features.

[0024] (2) The present invention integrates feature maps R1 to R3 and deep semantic feature maps Q1 to Q3, and uses result head processing to obtain scene classification, target detection, semantic segmentation and height reconstruction results. Scene classification and target detection adopt the detection head structure commonly used in YOLOv5, and semantic segmentation and height reconstruction results are obtained by 1×1 convolution. Scene classification and target detection are used to quickly locate the distribution of red mud in a large study area. Spatial range recognition and height reconstruction obtain the spatial boundary and height information of red mud at each location based on the location. Finally, volume quantization estimation is completed based on spatial range and height data. The multi-task deep learning model processes high-level features and low-level features respectively through the parallel structure of the main network and the sub-network. It can still maintain high task accuracy and robustness under complex and changeable environmental conditions, and provides a feasible solution for solving the problem of red mud volume calculation.

[0025] (3) The end-to-end integrated parallel processing method of the present invention eliminates the accumulation of process fragmentation error, outputs scene classification, detection box, segmentation mask and height reconstruction results simultaneously in the same network, and directly completes the accurate volume quantization estimation. Scene classification and target detection provide coarse localization, semantic segmentation finely delineates and accurately locates, and height reconstruction provides red mud height information. The four together complete the end-to-end closed-loop process from "finding - delineating - measuring height - calculating volume", mutually constraining and complementing each other, improving the overall recognition and measurement accuracy and efficiency.

[0026] (4) The present invention achieves the fusion of high and low layer feature results, taking into account both semantic global and detailed boundary. The main network is based on high and deep semantic features, focusing on global structure and category determination. The sub-network is based on low and shallow detailed features, strengthening boundary, texture and small target recognition. The two networks mutually restrict and complement each other through result fusion and game mechanism at the corresponding level, improving the overall recognition accuracy and robustness.

[0027] (5) The random depth feature extraction module RDFE of this invention can help the model maintain efficient feature extraction capability when dealing with different types of red mud by dynamically adjusting the extraction depth, while reducing redundant calculations. This enables the model to adapt to red mud of different scales and shapes, significantly reducing the number of parameters and floating-point operations, and improving the efficiency of model training and inference. In the main network, the feature pyramid network FPN realizes the information transmission of features from bottom to top and from top to bottom, integrates information of different scales, highlights the salience of multi-scale targets and global context association, and improves the model recognition accuracy and robustness. Attached Figure Description

[0028] Figure 1 This is a flowchart of the red mud volume monitoring and calculation method of the present invention;

[0029] Figure 2 This is a schematic diagram illustrating the processing principle of the multi-task deep learning model in the embodiment;

[0030] Figure 3 This is a schematic diagram illustrating the processing principle of the shifted convolutional layer SConv in the embodiment;

[0031] Figure 4 This is a schematic diagram illustrating the processing principle of the Random Depth Feature Extraction (RDFE) module in the embodiment.

[0032] Figure 5 for Figure 4 A schematic diagram of the structure of the two network modules DDM in the diagram;

[0033] Figure 6 This is a schematic diagram illustrating the processing principle of the SPPF module in the embodiment of the spatial pyramid pool fusion module;

[0034] Figure 7This is a schematic diagram illustrating the processing principle of the multi-head convolution module MSG-Attention in the embodiment. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to embodiments:

[0036] Example

[0037] like Figure 1 As shown, a method for monitoring and calculating red mud volume based on optical satellite imagery data and a multi-task deep learning model is presented. The method includes:

[0038] S1. Construct a multi-task deep learning model that includes a main network and sub-networks, such as... Figure 2 As shown, the main network is primarily used to extract three levels of deep feature maps P1 to P3 (i.e., three layers of high-level deep feature maps), while the sub-network is primarily used to extract a shallow feature map P4 (i.e., one layer of low-level features). The main network is responsible for obtaining various information about red mud using the high-level deep feature maps containing more semantic information, while the sub-network extracts local features through shallower convolutions, focusing on details, especially in the recognition of complex backgrounds and small objects. The sub-network can effectively improve recognition accuracy. Optical satellite imagery data of the study area is input into the multi-task deep learning model. In this embodiment, both the optical satellite imagery data and the optical satellite imagery data samples use image patches of 512×512×3 scale, and this is used as an example for technical introduction. The main network first extracts three levels of deep feature maps P1 to P3. In some embodiments, such as... Figure 2 As shown, the method for the main network to extract deep feature maps P1 to P3 is as follows:

[0039] S11. The main network sequentially includes a deep feature extraction module D1 to D3. Deep feature extraction modules D1 to D3 process the optical satellite image data through two shifted convolutional layers SConv and a random depth feature extraction module RDFE, compressing the spatial scale to obtain a deep feature map P1. Taking an input image scale of 512×512×3 as an example, the first shifted convolutional layer SConv is used to remove a large amount of invalid background information to obtain a feature map (256×256×32). The shifted convolutional layer SConv consists of a separable convolution module, a deep convolution module, a batch normalization layer, and a LeakReLU layer (an activation function layer). See [link to documentation]. Figure 3The separate convolution module is used to separate the number of convolutions, which is consistent with the number of channels in the input feature map (in_channel). The deep convolution module has the same number of deep convolutions as the number of channels in the target feature map (out_channel). The batch normalization layer normalizes the feature map to accelerate model training. The shift convolution layer SConv of this invention can significantly reduce the number of model parameters and improve model training efficiency while maintaining model performance.

[0040] S12, the deep feature extraction module D2 extracts and processes the deep feature map P1 through a shifted convolutional layer SConv and a random depth feature extraction module RDFE, and compresses the spatial scale to obtain the deep feature map P2.

[0041] S13, the deep feature extraction module D3 processes the deep feature map P2 sequentially through a shifted convolutional layer SConv, a random depth feature extraction module RDFE, and a spatial pyramid pooling fusion module SPPF, and compresses the spatial scale to obtain the deep feature map P3. Taking an input image scale of 512×512×3 as an example, the scales of deep feature maps P1 to P3 are (128×128×64), (64×64×128), and (32×32×256), respectively.

[0042] like Figure 4 As shown, in some embodiments, Figure 4 A backbone network for a random deep feature extraction module (RDFE) is provided. Before each operation, channels of number β×c_out are selected by random sampling and directly skipped and concatenated with the features of the backbone network. Here, β is a random decimal between 0 and 1, but only a multiple of 0.1, and c_out is the number of channels of the output features of the random deep feature extraction module (RDFE). Figure 4 In this context, h represents the feature height, and w represents the feature width. For example... Figure 5 As shown, the Random Deep Feature Extraction (RDFE) module includes two configurable residual connection network modules (DDM). One DDM has its residual connection shortcut set to False, while the other DDM has its residual connection shortcut set to True. This approach enriches deep information through splicing and fusion without increasing shallow layer computation. Figure 6 As shown, the Spatial Pyramid Pooling Fusion (SPPF) module uses multi-scale pooling operations to allow the deep feature map P3 to acquire global information at different resolutions while preserving detailed features in the image. Figure 4As shown, this embodiment designs a random depth feature extraction module (RDFE). By dynamically adjusting the extraction depth, the RDFE module can help the model maintain efficient feature extraction capabilities when dealing with different types of red mud, thereby enabling the model to adapt to red mud of different scales and morphologies.

[0043] Next, the deep feature maps P1 to P3 are extracted using a deep semantic extraction network to obtain deep semantic feature maps Q1 to Q3. In some embodiments, the method for the main network to extract deep semantic feature maps Q1 to Q3 is as follows: the deep feature maps P1 to P3 are processed by the Feature Pyramid Network (FPN) to obtain three target feature levels of deep semantic feature maps Q1 to Q3. Taking an input image scale of 512×512×3 as an example, the scales of deep feature maps P1 to P3 are (128×128×64), (64×64×128), and (32×32×256), respectively, and the scales of deep semantic feature maps Q1 to Q3 are: Q1 (128×128×832), Q2 (64×64×1600), and Q3 (32×32×1856), respectively. Each target feature level is processed by the result head to obtain the scene classification, object detection, semantic segmentation, and height reconstruction results. The scene classification and object detection adopt the detection head structure commonly used in YOLOv5, and the semantic segmentation and height reconstruction results are obtained by 1×1 convolution. In this embodiment, the Feature Pyramid Network (FPN) in the main network enables the information transmission of features from bottom to top and from top to bottom. Finally, the constructed result head is processed to obtain scene classification results, object detection results, semantic segmentation results, and height reconstruction results.

[0044] S2. The sub-network first extracts a shallow feature map P4. Taking an input image scale of 512×512×3 as an example, the multi-head convolutional module of the sub-network concatenates and fuses the deep feature maps P1 to P3 with the shallow feature map P4 to obtain fused feature maps R1 to R3. In some embodiments, the sub-network extracts the shallow feature map P4 as follows: it extracts, processes, and compresses the spatial scale through two shifted convolutional layers SConv and a random depth feature extraction module RDFE to obtain the shallow feature map P4.

[0045] The deep feature maps P1-P3 and the shallow feature map P4 are processed by a multi-head convolutional module MSG-Attention to perform multi-scale global information fusion, channel concatenation, and fusion processing to obtain a fused feature map R0 (forming a more information-rich feature map). In some embodiments, such as Figure 7As shown, the multi-head convolutional module MSG-Attention includes convolutional kernels of size=16×16 and stride=16, which divide the feature map into a sequence of 16×16 graphs. Position Embedding then encodes the positions of these graph sequences, helping the model understand the relationships between different positions and enhancing the network's sensitivity to specific locations. Multi-HeadAttention calculates the relationship between each position and other positions, and obtains a new feature representation for each position through weighted summation. ① By performing a linear transformation of the input feature map into query, key, and value, the relationship between each position and other positions (i.e., attention score) is calculated. ② Multiple attention heads are computed in parallel, each capturing information in a different subspace, enabling the model to focus on information from different regions and scales in the image. ③ The attention score is used to weighted aggregate the input features, ultimately obtaining a new feature representation. The Normalize layer refers to normalization, ensuring a uniform distribution of features. The Feed Forward neural network... The network consists of two fully connected layers. The first layer maps the input features to a larger dimensional space, increasing the model's expressive power. The second layer maps the features back to the original dimensional space. A ReLU activation function is used between the two layers. The Feed Forward Network helps capture the complex relationships between input features and further enhances the expressiveness of the features through non-linear transformations. The Splicing layer restores the graph sequence to a 128×128×256 feature map, and then obtains the target layer feature maps R1 (128×128×256), R2 (64×64×786), and R3 (32×32×1024) through random depth feature extraction, depthwise separable convolution, and concatenation operations. In this embodiment, the multi-head convolution module MSG-Attention enhances the model's ability to recognize multi-scale red mud targets in complex environments by fusing feature information from different scales. The multi-head convolution module MSG-Attention can effectively improve the recognition accuracy of occlusion, small objects, and details.

[0046] The fused feature map R0 is processed by a random deep feature extraction module RDFE to obtain fused feature map R1. Fusion feature map R1 is processed by a shifted convolutional layer SConv, a skip connection deep feature map P2, and another random deep feature extraction module RDFE to obtain fused feature map R2. Fusion feature map R2 is processed by a shifted convolutional layer SConv, a skip connection deep feature map P3, and another random deep feature extraction module RDFE to obtain fused feature map R3. Fusion feature maps R1 to R3 are obtained through three target feature levels. Each target feature level is processed using a result header to obtain scene classification, object detection, semantic segmentation, and height reconstruction results. Scene classification and object detection use the detection header structure commonly used in YOLOv5, while semantic segmentation and height reconstruction results are obtained using a 1×1 convolution.

[0047] S3. Merge the semantic segmentation results and height reconstruction results of the fused feature maps R1-R3 and the deep semantic feature maps Q1-Q3 for red mud identification, and output the three-dimensional data of red mud features. There are a total of six feature maps, including the fused feature maps R1-R3 (three features) and the deep semantic feature maps Q1-Q3 (three features). Each feature map is processed to obtain a scene classification result, an object detection result, a semantic segmentation result, and a height reconstruction result. The scene classification and object detection of the feature maps adopt the detection head structure commonly used in YOLOv5, and the semantic segmentation and height reconstruction results of the feature maps are obtained by 1×1 convolution. The six scene classification results and six semantic segmentation results were merged using the DS evidence fusion method; the six height reconstruction results were merged by averaging; and the six object detection results were merged using a game-theoretic approach. Specifically, after the feature maps are processed by the main and sub-networks, a series of candidate boxes (object detection results) are generated, and the confidence and classification probability of each box are calculated. The position of the candidate boxes is predicted by a regression model, and the confidence is calculated by a classification model. These candidate boxes usually have a large overlapping area, especially red mud, which may have multiple detection boxes at different scales. During the non-maximum suppression (NMS) algorithm, the model's candidate boxes will engage in a game-theoretic process, with the goal of selecting the most representative box from multiple candidate boxes. Non-maximum suppression (NMS) performs feature game theory in the following ways: ① Sort by confidence: First, NMS sorts each candidate box by its confidence value, retaining the box with the highest confidence. This box is considered the optimal candidate box, indicating that the model's prediction of the target within this box is most reliable. ② Calculate IoU (Intersection over Union): NMS calculates the IoU value between other candidate boxes and the currently selected box. The IoU value measures the degree of overlap between two boxes. Boxes with high IoU values ​​indicate that they may be detecting the same target, so redundant boxes should be removed. ③ Filtering and suppression: If the IoU between two candidate boxes exceeds a set threshold (usually 0.5), NMS removes the box with a higher IoU value but lower confidence. This ensures that each target object is detected only once and avoids redundancy and false detections. The above process is repeated until all candidate boxes have been processed.In this invention, the scene classification and semantic segmentation results of the main and auxiliary networks of the multi-task deep learning model are fused through DS evidence fusion, the height reconstruction results are obtained through mean fusion, and the target detection results are fused through a game mechanism to ensure that the optimal target box is retained. The game process relies on the non-maximum suppression algorithm (NMS). The NMS algorithm filters candidate boxes based on their confidence and intersection-over-union (IoU) values, removes redundant detection boxes, and retains only the most representative red mud target boxes. After multiple rounds of training and optimization, the model can accurately identify and locate red mud in complex environments, greatly improving the automatic detection and classification capabilities of red mud.

[0048] The red mud feature map is obtained by merging the feature maps R1 to R3 with the deep semantic feature maps Q1 to Q3. The red mud feature map is three-dimensional data of red mud features.

[0049] S4. Construct an optical satellite imagery data sample set containing labeled data and input it into the multi-task deep learning model for training on red mud identification and height reconstruction using methods S1 to S3. In some embodiments, the optical satellite imagery data and optical satellite imagery data samples input into the multi-task deep learning model have undergone preprocessing including image denoising and size normalization cropping. Both the optical satellite imagery data and the optical satellite imagery data samples use Gaofen-2 satellite imagery (Gaofen-2 satellite has a high-resolution image of 1 meter, which can provide rich red mud image data, suitable for red mud identification and analysis). Taking the optical satellite imagery data samples as an example, in order to ensure the diversity and representativeness of the optical satellite imagery data samples, the range of the optical satellite imagery data samples covers different red mud scenes. Each region may have different types of red mud, and each type of red mud is different in morphology, size and distribution. Acquiring optical satellite imagery data samples typically involves several quality issues, such as noise, uneven illumination, and background interference. Optical satellite imagery data samples require preprocessing, including image denoising and size normalization cropping. Denoising methods include: satellite images are often affected by various noises, including atmospheric interference and sensor noise, which affect image quality and consequently feature extraction. To eliminate this noise, filtering techniques are used; common denoising methods include median filtering, mean filtering, and Gaussian filtering. These methods effectively remove random noise from images while preserving important image features. Size normalization cropping involves: satellite images usually have high resolution. To accommodate multi-task deep learning models, the acquired satellite images need to be size normalized, adjusting all images to the same size (e.g., 512×512 pixels) to ensure input data consistency. The labeled data of the optical satellite image data samples include scene classification labels (i.e. whether the image patch contains red mud), target detection labels (the precise location or four-way coordinates of red mud in the image patch), semantic segmentation labels (the spatial range of red mud in the image patch), and height reconstruction labels (the corresponding height of red mud at each location).

[0050] When training the multi-task deep learning model using methods S1 to S3 for red mud identification and height reconstruction on an optical satellite imagery data sample set, the specific steps are as follows: The multi-task deep learning model is trained using the red mud training set (i.e., the training set in the optical satellite imagery data sample set); the model accuracy after each iteration is tested using the validation set (i.e., the validation set in the optical satellite imagery data sample set); red mud identification and height reconstruction are performed using the test set (i.e., the test set in the optical satellite imagery data sample set), and the final accuracy is calculated. In this embodiment, the multi-task deep learning model employs multiple loss functions during model training. Scene classification and semantic segmentation use the cross-entropy loss function, height reconstruction uses the root mean square loss function, and object detection uses the IOU loss function. In evaluating the accuracy of the multi-task deep learning model results, scene classification and semantic segmentation results use precision, recall, and F1 score; object detection results use the IOU index; and height reconstruction results use root mean square error and absolute error.

[0051] C2. Set the number of training iterations (epochs) and perform multiple iterations. In each iteration, use gradient descent to reduce the model loss value, while simultaneously optimizing and updating the model parameters.

[0052] The optical satellite imagery data samples in the optical satellite imagery data sample set have undergone sample data augmentation processing including rotation, mirroring, random cropping, and scaling. The samples also include adjustments to brightness, contrast, and saturation, and the addition of data samples under different lighting conditions. New training samples are generated through a series of random transformations (including sample data augmentation processing) to increase the diversity and richness of the training data, thereby improving the model's adaptability to changes in different scenes, lighting, and angles. For example, rotating satellite imagery samples by randomly selecting angles (e.g., -30° to +30°) allows the model to learn object features from different perspectives; horizontally or vertically flipping satellite imagery samples enhances the model's ability to recognize red mud in different directions; random cropping, by randomly cropping a region from the original image, simulates the behavior of objects at different positions and sizes, helping the model better adapt to changes in target scale; scaling processing, by enlarging or reducing the image at different scales, enhances the model's ability to recognize red mud of different sizes; and color adjustment processing, by adjusting the brightness, contrast, and saturation of the image, simulates recognition scenarios under different lighting conditions.

[0053] The volume of red mud deposits in the study area was calculated using the following formula based on the three-dimensional data of red mud characteristics:

[0054] Where n is the pixel index of the red mud feature in the three-dimensional data, and N is the total number of red mud pixels in the three-dimensional data. For the length data of pixel n, For the width data of cell n, Here is the height data for pixel n.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the volume of red mud based on optical satellite image data and a multi-task deep learning model, characterized in that: The method comprises the following steps: S1, constructing a multi-task deep learning model comprising a main network and a secondary network, inputting optical satellite image data of a study area into the multi-task deep learning model, the main network extracting three levels of deep feature maps P1-P3, and then extracting deep semantic feature maps Q1-Q3 from the deep feature maps P1-P3 through a deep semantic extraction network; S2, the secondary network first extracts a shallow feature map P4, and a multi-head convolution module of the secondary network splices and fuses the deep feature maps P1-P3 and the shallow feature map P4 to obtain fusion feature maps R1-R3; S3, the fusion feature maps R1-R3 and the deep semantic feature maps Q1-Q3 are respectively subjected to semantic segmentation and height reconstruction processing for red mud identification, and then all the semantic segmentation results and the height reconstruction results are combined to obtain red mud feature three-dimensional data; S4, constructing an optical satellite image data sample set comprising labeled data, inputting the sample set into the multi-task deep learning model to learn and train the red mud identification and height reconstruction through the method S1 to the method S3, and calculating the volume of the red mud deposit in the study area according to the following formula by using the red mud feature three-dimensional data: wherein n is the sequence number of the red mud pixel in the red mud characteristic three-dimensional data, N is the total number of red mud pixels in the red mud characteristic three-dimensional data, is the length data of the pixel n, is the width data of the pixel n, is the height data of the pixel n. 2.The method for monitoring the volume of red mud based on optical satellite image data and multi-task deep learning model according to claim 1, characterized in that: In the method S1, the main network extracts the deep feature maps P1-P3 in the following manner: S11, the main network comprises deep feature extraction modules D1-D3 in sequence, the deep feature extraction module D1 extracts and compresses the spatial scale of the optical satellite image data through two shift convolution layers SConv and a random deep feature extraction module RDFE in sequence to obtain the deep feature map P1; S12, the deep feature extraction module D2 extracts and compresses the spatial scale of the deep feature map P1 through a shift convolution layer SConv and a random deep feature extraction module RDFE in sequence to obtain the deep feature map P2; S13, the deep feature extraction module D3 extracts and compresses the spatial scale of the deep feature map P2 through a shift convolution layer SConv, a random deep feature extraction module RDFE and a spatial pyramid pooling fusion module SPPF in sequence to obtain the deep feature map P3. 3.The method for monitoring the volume of red mud based on optical satellite image data and multi-task deep learning model according to claim 1, characterized in that: The main network extracts the deep semantic feature maps Q1-Q3 in the following manner: the deep feature maps P1-P3 are processed through a feature pyramid network FPN to obtain three target feature levels of deep semantic feature maps Q1-Q3, and each target feature level is processed by a result head to obtain scene classification, target detection, semantic segmentation and height reconstruction results. 4.The method of monitoring red mud volume based on optical satellite image data and multi-task deep learning model according to claim 2, characterized in that: The shift convolution layer SConv is composed of a separation convolution module, a deep convolution module, a batch normalization layer and a LeakRelu layer; the random deep feature extraction module RDFE internally comprises two network modules DDM with configurable residual connections, one of which has a residual connection shortcut set to False, and the other has a residual connection shortcut set to True; the spatial pyramid pooling fusion module SPPF allows the deep feature map P3 to obtain global information at different resolutions and retain the detailed features in the image through a multi-scale pooling operation. 5.The method for monitoring the volume of red mud based on optical satellite image data and multi-task deep learning model according to claim 1, characterized in that: In the method S2, the method for extracting the shallow feature map P4 by the sub-network is as follows: the shallow feature map P4 is obtained by processing and compressing the spatial scale through two shift convolution layers SConv and a random depth feature extraction module RDFE.

6. The method for red mud volume monitoring based on optical satellite image data and multi-task deep learning model according to claim 1 or 5, characterized in that: In the method S2, the deep feature maps P1-P3 and the shallow feature map P4 are subjected to multi-scale global information fusion and channel splicing fusion processing through a multi-head convolution module MSG-Attention to obtain a fusion feature map R0.

7. The method for red mud volume monitoring based on optical satellite image data and multi-task deep learning model according to claim 6, characterized in that: The fusion feature map R0 is subjected to extraction processing through a random depth feature extraction module RDFE to obtain a fusion feature map R1; the fusion feature map R1 is subjected to processing through a shift convolution layer SConv, fusion processing through a jump connection deep feature map P2 and extraction processing through a random depth feature extraction module RDFE to obtain a fusion feature map R2; the fusion feature map R2 is subjected to processing through a shift convolution layer SConv, fusion processing through a jump connection deep feature map P3 and extraction processing through a random depth feature extraction module RDFE to obtain a fusion feature map R3; the fusion feature maps R1-R3 are obtained through three target feature levels respectively, and each target feature level obtains a scene classification result, a target detection result, a semantic segmentation result and a height reconstruction result through result head processing respectively. 8.The method for red mud volume monitoring based on optical satellite image data and multi-task deep learning model according to claim 1, characterized in that: The fusion feature maps R1-R3 and the deep semantic feature maps Q1-Q3 are six feature maps in total, and each feature map is subjected to processing to obtain a scene classification result, a target detection result, a semantic segmentation result and a height reconstruction result; The six scene classification results and the six semantic segmentation results are combined through a DS evidence fusion method; the six height reconstruction results are combined through an average method; the six target detection results are combined through a game method; the fusion feature maps R1-R3 and the deep semantic feature maps Q1-Q3 are combined to obtain a red mud feature map, and the red mud feature map is a red mud feature three-dimensional data. 9.The method for monitoring red mud volume based on optical satellite image data and multi-task deep learning model according to claim 1, characterized in that: The optical satellite image data and the optical satellite image data samples input into the multi-task deep learning model are preprocessed, including image denoising and size standardization and cropping; the optical satellite image data samples in the optical satellite image data sample set are subjected to sample data enhancement processing, including rotation, mirror flipping, random cropping and scale reduction; the optical satellite image data samples further include brightness, contrast and saturation adjustment and increase of data samples under different lighting conditions.

Images