A mining area change monitoring method, device and medium based on remote sensing data explicit reconstruction

Abstract

The present application relates to the field of remote sensing data processing, and discloses a mine area change monitoring method and device based on explicit reconstruction of remote sensing data and a medium, comprising: obtaining mine area pre- and post-time phase visible light remote sensing data, then performing explicit reconstruction based on mine area feature space distribution physical constraints at the original data level, jointly extracting directional geometric features, first-order / two-order sparsity features, spectral mean variance and mine area specific spectral index to generate 82-dimensional high-dimensional explicit physical features; then inputting into a change monitoring network containing 5 asymmetric channel attention modules to output change monitoring results. The present application replaces implicit extraction with physically driven explicit feature reconstruction, improves robustness under illumination and radiation interference, and has the advantages of strong interpretability, good mine area pertinence and high detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing, and in particular to a method, equipment, and medium for monitoring changes in mining areas based on explicit reconstruction of remote sensing data. Background Technology

[0002] Mining area change monitoring utilizes technical methods to qualitatively and quantitatively analyze changes in land features within mining areas, reflecting the changes in the surface of the mining area across different time dimensions. The changes in the mining area effectively reflect the progress of mining activities during that period. Accurate and effective assessment of these changes is a crucial technical means of understanding the progress of mining activities and provides important data support and reference for relevant management departments in their mining monitoring work. Currently, mainstream research methods include traditional and deep learning-based change monitoring methods, both of which can achieve good monitoring results. Traditional change monitoring methods typically use traditional techniques such as difference analysis, image ratio analysis, and principal component analysis to analyze remote sensing data from different time phases. While these methods do not require dataset construction, they often lack good data compatibility and adaptability. Deep learning-based change monitoring methods, on the other hand, use data-driven logic for network design to extract spectral, spatial, and temporal features from remote sensing data. The high-dimensional semantic features obtained through this extraction can better handle mining area change monitoring tasks and exhibit better adaptability. However, most current deep learning-based methods directly use raw remote sensing data to implicitly extract high-dimensional features through complex network structures, and then use these implicitly extracted high-dimensional features to complete the task of monitoring changes in mining areas. This leads to two problems: firstly, the implicit feature extraction mode, as a black box, is difficult to effectively explain and resolve the problem of some data failures; secondly, the raw remote sensing data is susceptible to the influence of light intensity and radiation, making the network's extraction of high-dimensional semantic features less robust and unable to effectively extract changed areas. Therefore, to solve these problems, this method proposes a mining area change monitoring method based on explicit reconstruction of remote sensing data. It explicitly reconstructs remote sensing data into high-dimensional physical feature data according to geometric and spectral characteristics. Then, a mining area change monitoring network based on explicit reconstruction of physical features is designed to process the explicitly reconstructed high-dimensional physical feature data. This allows for the mining of physical-driven data representations from different time periods, rather than the simple representations of the raw remote sensing data, to robustly obtain mining area change monitoring results at different times, thus realizing automated monitoring based on remote sensing data. Summary of the Invention

[0003] The purpose of this invention is to propose a method, equipment, and medium for monitoring changes in mining areas based on explicit reconstruction of remote sensing data. This addresses the problems of implicit feature extraction modes being difficult to interpret effectively as black boxes and failing to resolve some data, as well as the issues that the original remote sensing data is susceptible to the influence of light intensity and radiation, making the network extraction of high-dimensional semantic features insufficiently robust and unable to effectively extract changed areas.

[0004] Specifically, the present invention provides a method, equipment, and medium for monitoring changes in mining areas based on explicit reconstruction of remote sensing data. The method includes the following steps: S1. Obtain visible light remote sensing data of the mining area in the previous time phase. Visible light remote sensing data from the next time phase ; S2, the visible light remote sensing data of the previous time phase. Visible light remote sensing data from the next time phase At the raw data level, explicit physical feature reconstruction based on the spatial distribution of land features in the mining area is performed, transforming the raw remote sensing data into high-dimensional explicit physical features. The explicit physical feature reconstruction includes feature reconstruction based on the geometric characteristics of the mining area and feature reconstruction based on the spectral characteristics of the mining area. S3, the high-dimensional explicit physical features Input to the mining area change monitoring network based on explicit reconstruction of physical features The data is processed and reconstructed based on physical drives at different time phases to obtain the monitoring results of changes in the mining area. ; S4. Output the monitoring results of changes in the mining area. The results are used to indicate the areas of change and areas of no change in the geological features of the mining area.

[0005] A storage medium storing instructions and data for implementing a method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data.

[0006] A mining area change monitoring device based on explicit reconstruction of remote sensing data includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a mining area change monitoring method based on explicit reconstruction of remote sensing data.

[0007] The beneficial effects provided by this invention are: 1. Strong physical interpretability: Remote sensing data is explicitly reconstructed into high-dimensional physical features according to geometric and spectral physical properties, avoiding the problem of the "black box" implicit features of traditional deep learning methods being difficult to interpret. Each feature channel has a clear physical meaning (such as directional texture statistics, sparse edge curvature, mining area-specific spectral index, etc.), which facilitates the analysis of failure cases and the optimization of physical constraints.

[0008] 2. Significantly improved robustness: The physical features after explicit reconstruction are highly insensitive to changes in light intensity, atmospheric radiation differences, and sensor noise in the original remote sensing data. They can effectively suppress spurious changes and can still stably extract real change areas under complex lighting conditions, overcoming the shortcomings of directly using the original remote sensing data which is easily interfered with.

[0009] 3. Targeted design for mining areas: Customized texture statistics at a 7×7 neighborhood scale with four main directions (0°, 45°, 90°, 135°) were developed for the characteristics of the terrain in mining areas. In addition, proprietary spectral features such as bare soil index, vegetation degradation index and water pollution index were added to make the method more consistent with typical change patterns caused by mining activities, such as pit expansion, spoil heap changes, vegetation destruction and tailings dam pollution.

[0010] 4. Lightweight and efficient network structure: The asymmetric channel attention mechanism (compression ratio of 4 and 2) has fewer parameters than the standard SENet. The depth of the 5 PFPB modules matches the number of channels of the 82-dimensional physical features, which avoids overfitting while ensuring sufficient interaction of features and has high computational efficiency.

[0011] 5. Post-processing further reduces false alarms: By combining prior physical models such as mining boundaries and tailings dam shorelines with area filtering and shape constraints, isolated noise patches can be effectively filtered out, improving the engineering practicality of change detection.

[0012] 6. Adapting to Imbalanced Samples: The loss function introduces dynamic imbalance weights based on historical changes in the mining area, which solves the problem of the network being biased towards negative samples due to the extremely small number of pixels in the changed areas, and improves the recall rate of the changed areas. Attached Figure Description

[0013] Figure 1 This is a simplified flowchart of the method of the present invention; Figure 2 This is a schematic diagram of a mining area change monitoring network based on explicit reconstruction of physical characteristics; Figure 3 This is a schematic diagram of the physical feature processing module structure; Figure 4 This is a schematic diagram of the change monitoring module structure; Figure 5 This is a map of the monitoring results of changes in the mining area extracted by the method; Figure 6 This is a schematic diagram of the hardware device operation according to an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0015] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0016] Please refer to Figure 1 The present invention provides a method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data, comprising: S1. Obtain visible light remote sensing data of the mining area in the previous time phase. Visible light remote sensing data from the next time phase ; As one embodiment, step S1 of the present invention specifically involves: acquiring visible light remote sensing data of the mining area in the previous time phase (e.g., the early stage of mining activities or the previous monitoring period). Visible light remote sensing data from the next time phase (e.g., the current monitoring period) The visible light remote sensing data consists of standard remote sensing images containing red, green, and blue bands, with each temporal phase having a data size of [missing data]. ,in Image height (in pixels). The image width (in pixels) is represented by the superscripts 1 and 2, which indicate the previous and next time phases, respectively. In this embodiment, the remote sensing data can be sourced from satellite remote sensing platforms (such as Landsat, Sentinel-2, and Gaofen series satellites) or UAV aerial photography platforms. The data has undergone preprocessing such as radiometric calibration, atmospheric correction, and geometric registration to ensure pixel-level spatial correspondence between the images of different time phases.

[0017] S2, the visible light remote sensing data of the previous time phase. Visible light remote sensing data from the next time phase At the raw data level, explicit physical feature reconstruction based on the spatial distribution of land features in the mining area is performed, transforming the raw remote sensing data into high-dimensional explicit physical features. The explicit physical feature reconstruction includes feature reconstruction based on the geometric characteristics of the mining area and feature reconstruction based on the spectral characteristics of the mining area. As one embodiment, step S2 of the present invention specifically involves: processing the input remote sensing data from the previous time phase. and subsequent time-phase remote sensing data Instead of directly feeding the data into a deep learning network for implicit feature extraction, the data is first explicitly reconstructed based on the physical characteristics of the mining area's features at the raw data level.

[0018] The "physical constraints based on the spatial distribution of features in the mining area" refers to the fact that typical features in the mining area (such as pits, spoil heaps, stockpiles, transport roads, tailings ponds, etc.) have specific spatial scale ranges (usually from several meters to tens of meters), main directions of slope texture (commonly 0°, 45°, 90°, 135°, etc.) and spectral reflectance characteristics (exposed rocks, vegetation degradation, water pollution, etc. have unique spectral features).

[0019] This invention directly encodes these physical constraints into the feature reconstruction process, resulting in reconstructed high-dimensional explicit physical features. It possesses explicit physical meaning, rather than being a black-box implicit feature. Specifically, explicit physical feature reconstruction includes the following sub-steps: feature reconstruction based on the geometric characteristics of mining area features, feature reconstruction based on the spectral characteristics of mining area features, and generation of high-dimensional explicit physical features; through the above steps, this invention converts the input pre- and post-temporal remote sensing data... Convert to size High-dimensional explicit physical characteristics The first three channels of this feature represent the original spectral characteristics, while channels four through eighty-two represent the mined geometric and spectral physical characteristics. Each channel represents a specific physical attribute of the mining area (such as the texture mean in a certain direction, the spectral variance in a certain band, and a proprietary spectral index). This explicit reconstruction method allows subsequent networks to monitor changes based on the physically driven data representation, rather than on shallow, raw remote sensing data, thereby significantly improving robustness to interference factors such as light intensity and radiation changes.

[0020] Specifically, the feature reconstruction based on the geometric characteristics of the mining area in step S2 further includes: Visible light remote sensing data of the previous time phase Visible light remote sensing data from the next time phase Based on the spatial distribution scale of land features in the mining area, the main direction of slope texture and topographic curvature characteristics, we jointly extract directional physical geometric features, first-order sparse geometric features and second-order sparse geometric features. The extraction of the directional physical geometric features employs a 7×7 neighborhood that matches the spatial distribution scale of typical landforms in the mining area, and four directional angles that match the main directions of the mining area slopes and transportation roads. And calculate six statistical measures in each direction: mean, variance, homogeneity, contrast, dissimilarity and entropy, and then obtain the arithmetic mean of the directional dimensions; The first-order sparse geometric features are used to characterize the degree of abrupt changes in the edges of features in the mining area, and the second-order sparse geometric features are used to characterize the range of changes in the curvature of the terrain in the mining area. The three together constitute a multi-level physical representation of the geometric characteristics of features in the mining area.

[0021] As one embodiment, step S21 of the present invention specifically involves: the geometric and physical characteristics of the mining area exhibit a unique spatial distribution pattern—artificial features such as pits and spoil heaps typically present regular geometric shapes, slope textures have obvious directionality (common directions are 0°, 45°, 90°, and 135°), and the curvature of the terrain varies considerably. The present invention, targeting these characteristics, designs a combined extraction scheme: (I) Extraction of directional physical geometric features First, the remote sensing data from before and after the time phases are combined. and Convert to grayscale images respectively and The conversion formula uses the standard luminance formula:

[0022] Then, at each pixel position Take a 7×7 neighborhood centered at the center This reduces the grayscale levels of pixels in the neighborhood from the original 256 levels to... Levels are used to reduce computational complexity and enhance texture stability:

[0023] Next, at four directional angles Construct local gray-level direction feature representation matrices respectively:

[0024] in The reduced grayscale levels This is a conditional function (output 1 if the condition is met, otherwise output 0).

[0025] Normalize the directional dimension of the above matrix:

[0026] Then, calculate the mean, variance, homogeneity, contrast, dissimilarity, and entropy for each direction: Mean: ; variance: ; Homogeneity: ; Contrast: ; Dissimilarity: ; entropy: ; Finally, the arithmetic mean of the similar statistics in all four directions is taken to obtain the comprehensive statistics in the directional dimension (e.g., , (etc.), as a physical geometric feature of direction.

[0027] Note that the superscript 1 in the above formula indicates one direction, for example... , indicating direction The average value. This invention only uses the calculation formula for the direction of 0° as an example; the calculation methods for the other three directions are the same. The appearance of subscript 1 in the following text is also due to the same reason; the calculation for other directions should also be the same, and this invention will not provide further explanation.

[0028] (ii) Joint extraction of first-order and second-order sparse geometric features First-order sparse geometric features are used to characterize the degree of abrupt changes in the edges of features in mining areas, such as the boundaries of mining pits and the edges of transportation roads. This invention employs the following operator:

[0029] in For the first-order sparse operator in the horizontal direction, It is a first-order sparse operator in the vertical direction.

[0030] Second-order sparse geometric features are used to characterize the curvature variation of mining area terrain, such as the convex and concave transitions of slopes and the edges of spoil heaps. This invention employs the following operator:

[0031] in It is a second-order sparsity operator in the horizontal direction. It is a second-order sparse operator in the vertical direction.

[0032] The three geometric features mentioned above (directional physical geometry, first-order sparsity geometry, and second-order sparsity geometry) are used in parallel in subsequent steps to form a geometric integrity description of the mining area's features, from texture distribution and edge abrupt changes to curvature variations. Experiments show that using all three together can improve the F1 score by approximately 15% to 20% in mining area change monitoring compared to using any one or two features alone.

[0033] It should be noted that the joint extraction of first-order sparse geometric features and second-order sparse geometric features specifically includes: S211, Utilizing the first-order sparsity operator in the horizontal direction and the first-order sparse operator in the vertical direction For grayscale images respectively and Perform convolution and calculate the first-order sparse geometric features of the previous time phase based on the convolution result. and the first-order sparsity geometric features of the subsequent time phase ; As one embodiment, step S211 of the present invention specifically involves: for grayscale images (size is) ),use and Perform convolution operations separately to obtain the horizontal gradient map. and vertical gradient plot Then calculate the gradient magnitude. This amplitude map reflects the intensity of abrupt changes in terrain features in space. In mining scenarios, areas of change such as pit boundaries, road edges, and building outlines typically exhibit large gradient amplitudes. Similarly, for... Calculated .

[0034] S212, Utilizing the second-order sparsity operator in the horizontal direction and second-order sparse operators in the vertical direction For grayscale images respectively and Perform convolution and calculate the second-order sparse geometric features of the previous time phase based on the convolution result. Second-order sparsity geometric features of the subsequent time phase ; The first-order sparse geometric features and the second-order sparse geometric features are used in parallel with the directional physical geometric features in the subsequent high-dimensional explicit physical features to form a description of the geometric integrity of the mining area features from the edge to the curvature.

[0035] As one embodiment, step S212 of the present invention specifically involves: for grayscale images ,use and Perform convolution operations separately to obtain the second derivative plot in the horizontal direction. Graph of the second derivative in the vertical direction Then calculate the second gradient magnitude. This amplitude map reflects the variation in the curvature of terrain features. In mining scenarios, locations such as the convex and concave transitions of slopes, the edges of spoil heap steps, and the toe of tailings dams exhibit significant higher second-order derivative responses. Similarly, for... Calculated .

[0036] In this invention, first-order and second-order sparse geometric features are used in parallel with directional physical geometric features in subsequent high-dimensional explicit physical features to form a geometric integrity description of mining area features from edge to curvature. This combined use is not a simple superposition, but a physically driven design targeting the multi-level geometric attributes of mining area features, including edge-texture-curvature, with the three complementing and synergistically enhancing each other.

[0037] It should be noted that the feature reconstruction based on the spectral characteristics of land cover in the mining area in step S2 further includes: S221, at each pixel position Within a 7×7 neighborhood centered on the pixel, calculate the mean and variance of the pixel values ​​in the red, green, and blue channels to obtain the anisotropic spectral mean values ​​for the previous and subsequent time phases. and and anisotropic spectral variance and ; Specifically, in this step, spectral physical features are mined from the remote sensing data of the preceding and following time periods based on their spectral physical characteristics. These features include the mean of anisotropic spectra, the mean of isotropic spectra, the variance of anisotropic spectra, and the variance of isotropic spectra. The calculation process will be described in detail below. Specifically, for the input visible light remote sensing data of the preceding and following time periods... , We first calculate the mean values ​​of each anisotropic spectrum locally, as shown below:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] in, These are the anisotropic spectral mean values ​​in the red, green, and blue channels, respectively, obtained from the remote sensing data of the previous time phase. These are the anisotropic spectral mean values ​​in the red, green, and blue channels obtained from remote sensing data of the next time phase, respectively. Based on pixel position Take a 7x7 pixel neighborhood centered at the center. This represents the total number of pixels within the pixel set. These are remote sensing data from the previous time phase. Red, green, and blue channel data, These are remote sensing data from the next time phase. The red, green, and blue channel data.

[0044] Then, the visible light remote sensing data from the previous and subsequent time phases are analyzed. , The mean values ​​of the isotropic spectra are calculated as follows:

[0045]

[0046] in, These are the isotropic spectral mean values ​​for the previous and subsequent time phases, respectively. Based on pixel position Take a 7x7 pixel neighborhood centered at the center. This represents the total number of pixels within the pixel set. These are remote sensing data from the previous time phase. Red, green, and blue channel data, These are remote sensing data from the next time phase. The red, green, and blue channel data.

[0047] For the input visible light remote sensing data of the previous and next time phases , Let's calculate the variance of each anisotropic spectrum, as shown below:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] in, These are the anisotropic spectral variances in the red, green, and blue channels obtained from the remote sensing data of the previous time phase. These are the anisotropic spectral variances in the red, green, and blue channels obtained from the remote sensing data of the next time phase, respectively. Based on pixel position Take a 7x7 pixel neighborhood centered at the center. This represents the total number of pixels within the pixel set. These are remote sensing data from the previous time phase. Red, green, and blue channel data, These are remote sensing data from the next time phase. Red, green, and blue channel data; These are the anisotropic spectral mean values ​​in the red, green, and blue channels, respectively, obtained from the remote sensing data of the previous time phase. These are the anisotropic spectral mean values ​​obtained from the remote sensing data of the next time phase in the red, green, and blue channels, respectively.

[0054] S222. Within the 7×7 neighborhood, calculate the mean and variance of the vector magnitudes of all channel pixel values ​​to obtain the isotropic spectral mean values ​​for the previous and subsequent time phases. and and the variance of each isotropic spectrum and ; Specifically, for the input visible light remote sensing data of the previous and subsequent time phases... , Let's calculate the variance of the isotropic spectra, as shown below:

[0055] in, , These are the variances of the isotropic spectra of the preceding and following time phases, respectively. Based on pixel position Take a 7x7 pixel neighborhood centered at the center. This represents the total number of pixels within the pixel set. These are remote sensing data from the previous time phase. Red, green, and blue channel data, These are remote sensing data from the next time phase. Red, green, and blue channel data; These are the isotropic spectral mean values ​​for the previous and subsequent time phases, respectively.

[0056] S223. Within the 7×7 neighborhood, additionally calculate the mining area-specific spectral index features, which include at least: a bare soil index for characterizing exposed surface / mining face, a vegetation degradation index for characterizing the degree of vegetation destruction in the mining area, and a water pollution index for characterizing the impact of tailings pond or ore dressing plant drainage. As one embodiment, step S223 of the present invention specifically involves: Mining areas contain typical features not found in ordinary natural scenes, therefore requiring the design of specialized spectral indices. Specifically, as follows: (1) Bare Soil Index (BSI): Used to characterize exposed ground surfaces or mining faces. The calculation formula is:

[0057] In this invention, since visible light remote sensing data (excluding near-infrared (NIR) bands) is used, an approximation is employed: the blue band is used as a substitute for near-infrared (because exposed rocks have low reflectivity in the blue band and high reflectivity in the red-green bands). The actual calculations use the following approach:

[0058] The higher the index, the greater the degree of surface exposure, corresponding to mining areas such as pits and spoil heaps in the mining area.

[0059] (2) Vegetation Degradation Index (VDI): Used to characterize the degree of vegetation damage in mining areas. The degradation form of the Normalized Difference Vegetation Index (NDVI) is adopted:

[0060] The index has a lower value in areas with healthy vegetation and a significantly higher value in areas where vegetation is damaged (becomes bare soil or collapses).

[0061] (3) Water Pollution Index (WPI): Used to characterize the impact of tailings ponds or ore dressing plant wastewater discharge. The calculation formula is:

[0062] Tailings dam drainage typically contains suspended solids and heavy metal ions, causing the water body to have increased reflectivity in the green band and decreased reflectivity in the blue band. Therefore, an abnormally high WPI value can be used as an indicator of water pollution.

[0063] The three types of spectral indices specific to the mining area were calculated separately in the preceding and following time phases to obtain the corresponding characteristic maps.

[0064] S224. The general spectral physical features obtained in steps S221 and S222 are fused with the mine-specific spectral index features obtained in step S223 to form composite spectral physical features.

[0065] As one embodiment, step S224 of the present invention specifically involves: splicing the general spectral physical features (3 channels of anisotropic mean + 3 channels of anisotropic variance + 1 channel of isotropic mean + 1 channel of isotropic variance, totaling 8 channels) with the mining area-specific spectral index features (bare soil index, vegetation degradation index, and water pollution index, each with 2 time phases, totaling 6 channels) along the channel dimension to form a composite spectral physical feature with 14 channels. This composite feature retains both general spectral statistical information and introduces specific indicators for typical features in the mining area, enabling a more effective characterization of changes in the mining area.

[0066] It should be noted that the generation of high-dimensional explicit physical features in step S2 further includes: S231, Original spectral characteristics All composite spectral physical features and all calculated geometric features obtained in steps S221 to S224 are spliced ​​together along the channel dimension to obtain preliminary high-dimensional explicit physical features. The number of channels for this preliminary feature is dynamically determined to be 82 channels based on the total number of physical features in the mining area; The above processes have extracted geometric and spectral physical features, including directional physical geometric features, first-order sparsity geometric features, second-order sparsity geometric features, anisotropic / isotropic spectral mean / variance, and mining area-specific spectral indices. In this step, we summarize and standardize the extracted geometric and spectral physical features along with the original spectral characteristics to generate high-dimensional explicit physical features.

[0067] Specifically, for the geometric and spectral physical features mined above, we concatenate them along the channel dimension and add the original spectral properties to generate preliminary high-dimensional explicit physical features, as shown below:

[0068] in, This is a preliminary high-dimensional explicit physical characteristic, its size is ; This indicates an operation of feature stitching along the feature dimension. The specific composition of the 82 channels is as follows: 6 channels of original spectrum + 6 channels of anisotropic spectral mean + 2 channels of isotropic spectral mean + 6 channels of anisotropic spectral variance + 2 channels of isotropic spectral variance + 6 channels of mining area-specific spectral index + directional physical geometric features (6 statistical measures in each direction × 4 directions × 2 phases = 48 channels; in actual stitching, both the features in each direction and the features after directional averaging are included, for a total of 82 channels).

[0069] S232, Regarding the preliminary high-dimensional explicit physical features For each channel, a standardization process is performed by subtracting its mean and dividing by its variance to obtain the final high-dimensional explicit physical feature. ; The 82 channels have a direct physical correspondence with the number of stacked physical feature processing modules in the subsequent network, with each module responsible for processing a specific category of physical feature subset.

[0070] For the initially generated high-dimensional explicit physical features Channel-by-channel standardization is performed to generate the final high-dimensional explicit physical features, as shown below:

[0071] in, The final high-dimensional explicit physical feature has a size of ; , The c-th channel at pixel position The final high-dimensional explicit physical eigenvalues, and the initially generated high-dimensional explicit physical eigenvalues. , These are the mean and variance of the c-th channel image of the initially generated high-dimensional explicit physical feature values, respectively.

[0072] Specifically, this invention establishes a correspondence between the 82-channel features and five physical feature processing modules in the subsequent network: the first module mainly processes geometric feature channels (directional texture, first-order / second-order sparsity), the second module mainly processes spectral feature channels (anisotropic / isotropic mean and variance), the third module processes mining-specific spectral index channels, and the fourth and fifth modules handle cross-feature interaction fusion. This correspondence ensures the physical interpretability of the network design, rather than being a black-box stacking.

[0073] S3, the high-dimensional explicit physical features Input to the mining area change monitoring network based on explicit reconstruction of physical features The data is processed and reconstructed based on physical drives at different time phases to obtain the monitoring results of changes in the mining area. ; It should be noted that the mining area change monitoring network in step S3 Includes 5 cascaded physical feature processing modules And one change monitoring module, the stacking number of the five physical feature processing modules is related to the high-dimensional explicit physical features. The channel number 82 matches, and the specific processing steps are as follows: S31, the high-dimensional explicit physical features The data is sequentially input into five cascaded physical feature processing modules. Each module employs an asymmetric channel attention structure to process the physical features between different channels, outputting change monitoring features. ; As one embodiment, step S3 of the present invention specifically involves: processing the high-dimensional explicit physical features generated in step S2. Input to the mining area change monitoring network based on explicit reconstruction of physical features In this study, a carefully designed network structure is used to mine high-dimensional explicit physical features, thereby robustly obtaining monitoring results of changes in the mining area over different time periods. Specifically:

[0074] in, This is the mining area change monitoring result extracted by this method, and its size is... ; The structural diagram of the mining area change monitoring network based on explicit reconstruction of physical features proposed in this paper is shown below. Figure 2As shown below, the network structure is described in detail.

[0075] For the high-dimensional explicit physical features of the input The data is then input into five consecutive physical feature processing modules to fully process and mine the physical features between different channels, as shown below:

[0076] in, These are the change monitoring features processed by the continuous physical feature processing module. The proposed physical feature processing module is shown in the schematic diagram below. Figure 3 As shown, the five physical feature processing modules here have the same structure, and their stacking number has a direct physical correspondence with the 82 channels of the high-dimensional explicit physical features: each module is responsible for processing a specific category of physical feature subset (such as the first module focusing on geometric features, the second module focusing on spectral features, etc.).

[0077] Each of the physical feature processing modules Specific implementation: S311. After batch normalizing the input features, the preliminary physical features are obtained through a 3×3 convolutional layer and a ReLU activation function, as follows: F 1 = ReLU(conv1(BatchNorm( H ))) in, The physical characteristics are for preliminary processing, and their size is ; This refers to the BatchNorm2d function in PyTorch. This refers to the ReLU function in PyTorch. It is a 3×3 convolution operation with 82 input and 82 output channels.

[0078] S312. The preliminary physical features are sequentially passed through global average pooling, a first fully connected layer with a compression ratio of 4, a ReLU activation function, a second fully connected layer with a compression ratio of 2, and a Sigmoid activation function to establish asymmetric channel attention weights. These weights are then multiplied by the preliminary physical features to obtain attention-enhanced features. This invention employs asymmetric channel attention, unlike the conventional SENet. Specifically, we utilize a global average pooling layer, a fully connected layer, and an activation function to establish channel attention, as shown below:

[0079] in, This is the feature vector after channel attention processing, and its size is... ; This is the adaptive_avg_pool2d operation (global average pooling) in PyTorch, which outputs a vector of length 82. The first fully connected layer has an input of 82 and an output of 20 (compression ratio of approximately 4). activation; This is the second fully connected layer, with an input of 20 and an output of 41 (compression ratio of approximately 2). The output is mapped to the (0,1) interval. Since the output has 41 channels and the input features have 82 channels, this invention employs grouped attention: the 82 channels are divided into two groups of 41 channels each, sharing the same set of attention weights. Then, the attention weights are... Multiply by each channel to get .

[0080] S313. Repeat steps S311 and S312 once to perform convolution and asymmetric channel attention processing on the attention-enhanced features again. Feature vectors after channel attention processing After performing batch normalization again, input a convolutional layer and activation function, as shown below:

[0081] in, For further processing of physical features, its size is ; It is a 3×3 convolution with 82 input and output channels.

[0082] Next, we utilize asymmetric channel attention again:

[0083] S314. Perform batch normalization on the reprocessed features, and output the features of this module through the last 3×3 convolutional layer and ReLU activation function; Finally, the feature vector after full channel attention processing After performing batch normalization, a convolutional layer and activation function are input to obtain the feature vector output by this module, as shown below:

[0084] in, Physical feature processing module The output feature vector has a size of ; It is a 3×3 convolution with 82 input and output channels.

[0085] S32, The change monitoring features The data is input to the change monitoring module, and after multiple convolution operations, it finally outputs a mining area change monitoring result map with 1 channel. .

[0086] It should be noted that the change monitoring module in step S32 specifically performs the following: S321. Monitoring the changes... After batch normalization, intermediate features are obtained by passing the data through a 3×3 convolutional layer and the ReLU activation function. ; S322, Regarding the intermediate feature Batch normalization is performed again, and deep features are obtained by passing the data through another 3×3 convolutional layer and the ReLU activation function. ; S323, The deep features The system sequentially passes through a 3×3 convolutional layer with 16 output channels, a ReLU activation function, and a 3×3 convolutional layer with 1 output channel to generate preliminary change monitoring results. S324. Apply constraints based on the mining area topographic physical model to the preliminary change monitoring results, the constraints including: filtering out isolated noise areas with areas smaller than the minimum change threshold of the mining area based on the prior geometry of the mining boundary or tailings dam shoreline, to obtain the final mining area change monitoring results. .

[0087] Through five consecutive physical feature processing modules, we can process high-dimensional explicit physical features. We have fully explored the data representation between channels. Next, we will focus on change monitoring features. The data is input into the change monitoring module to obtain the mining area change monitoring results extracted by this method. For example... Figure 4 As shown in the diagram, it illustrates the structure of the change monitoring module, which we will now describe in detail.

[0088] Specifically, the change monitoring features obtained from processing five consecutive physical feature processing modules... After batch normalization, the data is input into a convolutional layer and activation function, as shown below:

[0089] in, The change monitoring feature vector, which is to be further processed, has a size of [missing information]. ; It is a 3×3 convolution with 82 input and output channels.

[0090] Next, the change monitoring feature vector will be further processed. After performing batch normalization, a convolutional layer and activation function are input to fully extract the change detection feature vectors, as shown below:

[0091] in, To fully process the change monitoring feature vector, its size is ; It is a 3×3 convolution with 82 input and output channels.

[0092] Finally, the fully processed change monitoring feature vectors The input is fed into a convolutional layer, an activation function, and another convolutional layer to obtain the mining area change monitoring results extracted by this method, as shown below:

[0093] in, This is the mining area change monitoring result extracted by this method, and its size is... ; It is a 3×3 convolution with 82 input channels and 16 output channels; It is a 3×3 convolution with 16 input channels and 1 output channel.

[0094] Furthermore, as a preferred embodiment of the present invention, the preliminary change monitoring results are further subjected to post-processing based on a physical model of the mining area topography: Area filtering: Binarization with a threshold of 0.5 is used to extract all connected components, and the pixel area of ​​each patch is calculated. Based on the minimum varying physical size of features in the mining area (e.g., 50 square meters, corresponding to approximately 20 pixels on the image), isolated noise patches with areas smaller than this threshold are filtered out.

[0095] Geometric constraints: Utilizing the prior geometry of the mining area boundary or tailings dam shoreline (which can be obtained from the mining area GIS database), only variation patches intersecting with these prior areas are retained. For example, if the mining pit boundary is known to be a polygon, only the variation detection results within that polygon are retained.

[0096] After the above post-processing, the final monitoring results of changes in the mining area are obtained. .

[0097] S4. Output the monitoring results of changes in the mining area. The results are used to indicate the areas of change and areas of no change in the geological features of the mining area.

[0098] As one embodiment, step S4 of the present invention specifically involves: processing the change monitoring results obtained in step S3. The probabilistic map is thresholded (typically with a threshold of 0.5) to generate a binary change map, where white pixels (value 1) indicate a changed area and black pixels (value 0) indicate an unchanged area. This result can be overlaid on a subsequent time-phase remote sensing image to visually demonstrate land use / cover changes caused by mining activities. Simultaneously, the resulting map can be exported to raster formats such as GeoTIFF for subsequent area statistics, change vector analysis, and mine supervision report generation.

[0099] It should be noted that the method also includes: before step S3, using a binary cross-entropy loss function to analyze the mining area change monitoring network. During training, the loss function incorporates imbalance weights between samples from changed and unchanged areas of the mining area, and these weights are dynamically determined based on historical change statistics of the mining area.

[0100] Specifically, for the proposed mining area change monitoring network based on explicit reconstruction of physical features, we use an improved binary cross-entropy function as the loss function for training, incorporating the imbalance weights of samples from changed and unchanged mining areas into the loss function.

[0101]

[0102] in, The true value is the calibrated value (1 indicates change, 0 indicates no change). Weight and Dynamically determined based on historical change statistics of the mining area:

[0103] in and These represent the total number of changed and unchanged pixels in the training batch, respectively. Alternatively, they can be calculated based on the historical average percentage of changed area in the mining area. Set fixed weights .

[0104] The Adam optimizer was used during training, with an initial learning rate of 0.001, which was decayed to 0.9 times the original rate every 20 epochs, for a total of 100 epochs.

[0105] like Figure 5 As shown, this method can still accurately and robustly obtain monitoring results of changes in mining areas for remote sensing images of different time phases where light intensity and radiation have a significant impact. Figure 5In the image, (a) is the remote sensing image of the later time phase, (b) is the remote sensing image of the earlier time phase, and (c) is the mining area change monitoring result map extracted by this method (white represents the changed area, and black represents the unchanged area). The results show that this method can effectively suppress spurious changes caused by differences in illumination and radiation, and accurately extract the true areas of change in the mining area.

[0106] Example 2: Please see Figure 6 , Figure 6 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a mining area change monitoring device 401 based on explicit reconstruction of remote sensing data, a processor 402, and a storage medium 403.

[0107] A mining area change monitoring device 401 based on explicit reconstruction of remote sensing data: The mining area change monitoring device 401 based on explicit reconstruction of remote sensing data implements the mining area change monitoring method based on explicit reconstruction of remote sensing data.

[0108] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the mining area change monitoring method based on explicit reconstruction of remote sensing data.

[0109] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the mining area change monitoring method based on explicit reconstruction of remote sensing data.

[0110] 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, improvements, etc., 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 changes in mining areas based on explicit reconstruction from remote sensing data, characterized in that: Includes the following steps: S1. Obtain visible light remote sensing data of the mining area in the previous time phase. Visible light remote sensing data from the next time phase ; S2, the visible light remote sensing data of the previous time phase. Visible light remote sensing data from the next time phase At the raw data level, explicit physical feature reconstruction based on the spatial distribution physical constraints of land features in the mining area is performed, transforming the raw remote sensing data into high-dimensional explicit physical features. The explicit physical feature reconstruction includes feature reconstruction based on the geometric characteristics of the mining area and feature reconstruction based on the spectral characteristics of the mining area. S3, the high-dimensional explicit physical features Input to the mining area change monitoring network based on explicit reconstruction of physical features The data is processed and reconstructed based on physical drives at different time phases to obtain the monitoring results of changes in the mining area. ; S4. Output the monitoring results of changes in the mining area. The results are used to indicate the areas of change and areas of no change in the geological features of the mining area.

2. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 1, characterized in that, The feature reconstruction based on the geometric characteristics of the mining area in step S2 further includes: Visible light remote sensing data of the previous time phase Visible light remote sensing data from the next time phase Based on the spatial distribution scale of landforms in the mining area, the main direction of slope texture and topographic curvature characteristics, we jointly extract directional physical geometric features, first-order sparse geometric features and second-order sparse geometric features. The extraction of the directional physical geometric features employs a 7×7 neighborhood that matches the spatial distribution scale of typical landforms in the mining area, and four directional angles that match the main directions of the mining area slopes and transportation roads. The mean, variance, homogeneity, contrast, dissimilarity, and entropy of each direction are calculated and then the result is obtained by arithmetic mean of the directional dimensions. The first-order sparse geometric features are used to characterize the degree of abrupt changes in the edges of features in the mining area, and the second-order sparse geometric features are used to characterize the range of changes in the curvature of the terrain in the mining area. The three together constitute a multi-level physical representation of the geometric characteristics of features in the mining area.

3. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 2, characterized in that: The joint extraction of first-order and second-order sparse geometric features specifically includes: S211, Utilizing the first-order sparsity operator in the horizontal direction and the first-order sparse operator in the vertical direction For grayscale images respectively and Perform convolution and calculate the first-order sparse geometric features of the previous time phase based on the convolution result. First-order sparsity geometric features of the next time phase ; S212, Utilizing the second-order sparsity operator in the horizontal direction and second-order sparse operators in the vertical direction For grayscale images respectively and Perform convolution and calculate the second-order sparse geometric features of the previous time phase based on the convolution result. Second-order sparsity geometric features of the subsequent time phase ; The first-order sparse geometric features and the second-order sparse geometric features are used in parallel with the directional physical geometric features in the subsequent high-dimensional explicit physical features to form a description of the geometric integrity of the mining area features from the edge to the curvature.

4. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 2, characterized in that, The feature reconstruction based on the spectral characteristics of land cover in the mining area in step S2 further includes: S221, at each pixel position Within a 7×7 neighborhood centered on the pixel, calculate the mean and variance of the pixel values ​​in the red, green, and blue channels to obtain the anisotropic spectral mean values ​​for the previous and subsequent time phases. and and anisotropic spectral variance and ; S222. Within the 7×7 neighborhood, calculate the mean and variance of the vector magnitudes of all channel pixel values ​​to obtain the isotropic spectral mean values ​​for the previous and subsequent time phases. and and the variance of each isotropic spectrum and ; S223. Within the 7×7 neighborhood, additionally calculate the mining area-specific spectral index features, which include at least: a bare soil index for characterizing exposed surface / mining face, a vegetation degradation index for characterizing the degree of vegetation destruction in the mining area, and a water pollution index for characterizing the impact of tailings pond or ore dressing plant drainage. S224. The general spectral physical features obtained in steps S221 and S222 are fused with the mine-specific spectral index features obtained in step S223 to form composite spectral physical features.

5. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 4, characterized in that, The generation of high-dimensional explicit physical features in step S2 further includes: S231, Original spectral characteristics All composite spectral physical features and all calculated geometric features obtained in steps S221 to S224 are spliced ​​together along the channel dimension to obtain preliminary high-dimensional explicit physical features. The number of channels for this preliminary feature is dynamically determined to be 82 channels based on the total number of physical features in the mining area. S232, Regarding the preliminary high-dimensional explicit physical features For each channel, a standardization process is performed by subtracting its mean and dividing by its variance to obtain the final high-dimensional explicit physical feature. ; The 82 channels have a direct physical correspondence with the number of stacked physical feature processing modules in the subsequent network, with each module responsible for processing a specific category of physical feature subset.

6. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 1, characterized in that, Mining area change monitoring network in step S3 Includes 5 cascaded physical feature processing modules And one change monitoring module, the stacking number of the five physical feature processing modules is related to the high-dimensional explicit physical features. The channel number 82 matches, and the specific processing steps are as follows: S31, The high-dimensional explicit physical features The data is sequentially input into five cascaded physical feature processing modules. Each module employs an asymmetric channel attention structure to process the physical features between different channels, outputting change monitoring features. ; Each of the physical feature processing modules Specific implementation: S311. After batch normalizing the input features, the preliminary physical features are obtained by passing them through a 3×3 convolutional layer and the ReLU activation function. S312. The preliminary physical features are sequentially passed through global average pooling, a first fully connected layer with a compression ratio of 4, a ReLU activation function, a second fully connected layer with a compression ratio of 2, and a Sigmoid activation function to establish asymmetric channel attention weights. These weights are then multiplied by the preliminary physical features to obtain attention-enhanced features. S313. Repeat steps S311 and S312 once to perform convolution and asymmetric channel attention processing on the attention-enhanced features again. S314. Perform batch normalization on the reprocessed features, and output the features of this module through the last 3×3 convolutional layer and ReLU activation function; S32, The change monitoring features The data is input to the change monitoring module, and after multiple convolution operations, it finally outputs a mining area change monitoring result map with 1 channel. .

7. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 1, characterized in that, The change monitoring module in step S32 is specifically executed as follows: S321. Monitoring the changes... After batch normalization, intermediate features are obtained by passing the data through a 3×3 convolutional layer and the ReLU activation function. ; S322, Regarding the intermediate feature Batch normalization is performed again, and deep features are obtained by passing the data through another 3×3 convolutional layer and the ReLU activation function. ; S323, The deep features The system sequentially passes through a 3×3 convolutional layer with 16 output channels, a ReLU activation function, and a 3×3 convolutional layer with 1 output channel to generate preliminary change monitoring results. S324. Apply constraints based on the mining area topographic physical model to the preliminary change monitoring results, the constraints including: filtering out isolated noise areas with areas smaller than the minimum change threshold of the mining area based on the prior geometry of the mining boundary or tailings dam shoreline, to obtain the final mining area change monitoring results. .

8. The method for monitoring changes in mining areas based on explicit reconstruction of remote sensing data as described in claim 1, characterized in that, The method also includes: before step S3, using a binary cross-entropy loss function to analyze the mining area change monitoring network. During training, the loss function incorporates imbalance weights between samples from changed and unchanged areas of the mining area, and these weights are dynamically determined based on historical change statistics of the mining area.

9. A storage medium, characterized in that: The storage medium stores instructions and data to implement the mining area change monitoring method based on explicit reconstruction of remote sensing data as described in any one of claims 1 to 8.

10. A mining area change monitoring device based on explicit reconstruction of remote sensing data, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the mining area change monitoring method based on explicit reconstruction of remote sensing data as described in any one of claims 1 to 8.

Images