Dump boundary recognition method combining multi-driving factors and fusion network model

Abstract

The application discloses a dump boundary identification method combining a multi-driving factor and a fusion network model, and the method comprises the following steps: S1, constructing a dump boundary sample data set and making a driving factor feature library; S2, constructing a fusion network model DAMAT-U, a coding unit performs layer-by-layer multi-scale feature extraction on remote sensing image data and sequentially obtains feature maps A1-An; S3, a decoding unit performs layer-by-layer decoding processing on the feature map An, and the decoding unit comprises a driving factor adaptive module DFAM, the driving factor adaptive module DFAM generates modulation parameters of the model in real time by using driving factor feature data; in the decoding process, the corresponding layer feature maps of the coding unit are correspondingly connected and fused; and S4, remote sensing image data of a research area is input into the trained fusion network model DAMAT-U to perform dump boundary prediction. The application realizes the purpose of separating the dump boundary of the research mining area by using deep fusion, multi-scale feature extraction, attention enhancement and driving factor adaptive modulation.

Description

Technical Field

[0001] This invention relates to the field of boundary detection of spoil heaps in open-pit mines, and in particular to a method for identifying spoil heap boundaries that combines multiple driving factors and a fusion network model. Background Technology

[0002] Tailings ponds, waste rock piles, etc., generated during open-pit mining not only occupy a large amount of land resources, but also, due to their unique loose depositional structure and the influence of external environmental factors (such as rainfall, weathering, earthquakes, etc.), are highly susceptible to geological disasters such as slope instability, landslides, and debris flows, seriously threatening the safety of mining operations, the surrounding ecological environment, and the safety of people's lives and property. Tailings ponds in different climate zones (such as arid and humid regions) exhibit significant differences in remote sensing imagery. For example, in arid regions with sparse vegetation cover, the spectral characteristics of tailings ponds are prominent, but they can easily be confused with bare ground and rocks. In humid regions with abundant vegetation, tailings ponds may be partially covered by vegetation, making the boundaries spectrally blurred, but texture and topographic features may be more pronounced. Therefore, achieving high-precision, automated interpretation of tailings pond boundaries is of great significance for mine safety production, ecological environmental protection, and geological disaster prevention.

[0003] Traditional monitoring of spoil heaps relies primarily on field measurements and manual interpretation, which is costly, time-consuming, and labor-intensive, making it difficult to meet the needs of large-scale, dynamic, and real-time monitoring. Modern high-resolution remote sensing technology, with its macroscopic and dynamic advantages, can effectively compensate for the shortcomings of traditional methods. Semantic segmentation models based on deep learning (such as U-Net and its improvements) have been widely applied to target extraction and feature identification in mining areas, achieving boundary detection by automatically learning remote sensing image features. For example, the U-Net model based on multi-scale sample sets has been used for identifying industrial solid waste and open-pit mines, significantly improving identification accuracy and generalization ability. Furthermore, multi-source remote sensing data fusion models, utilizing a single deep network to simultaneously process multi-channel images, have also demonstrated the effectiveness of leveraging multi-source information to enhance segmentation results.

[0004] However, existing technologies still have shortcomings in monitoring spoil heaps: (1) Existing models are mostly trained based on single-scale or single-image sources, which are insufficient for representing spoil heaps of different shapes and sizes. Boundary segmentation often becomes blurred or missing, limiting the accuracy of identification. (2) Environmental driving factors are ignored. Different climate zones (such as arid, semi-arid, and humid) have significant differences in vegetation cover, soil moisture, and lithological weathering. These climatic factors (such as precipitation and temperature) are the main drivers of changes in surface vegetation and landforms. If the model training does not fully consider these heterogeneous features, the cross-domain performance will decline. Therefore, a new method that can deeply integrate multi-source data, adapt to environmental driving factors, and optimize boundary identification tasks is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems pointed out in the background art and to provide a method for identifying spoil heap boundaries by combining multiple driving factors and a fusion network model. This method achieves deep fusion of multi-source remote sensing data and driving factors, multi-scale feature extraction, attention enhancement, and adaptive modulation of driving factors. By using the fusion network model DAMAT-U, adaptive feature weighting is achieved through the embedding of driving factors and conditional modulation within the fusion network model DAMAT-U, ultimately realizing the goal of segmenting spoil heap boundaries in the research area.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A method for identifying spoil heap boundaries that combines multiple driving factors and a fusion network model, the method comprising:

[0008] S1. Construct a waste dump boundary sample dataset, which includes remote sensing image data and associated labeled waste dump boundaries; create a driving factor feature library containing P driving factors, and calculate and obtain driving factor feature data based on the driving factor feature library for the remote sensing image data;

[0009] S2. The DAMAT-U fusion network model is constructed with the U-Net network architecture as the core and trained using the waste dump boundary sample dataset. The DAMAT-U fusion network model includes an encoding unit and a decoding unit. The encoding unit performs multi-scale feature extraction on the remote sensing image data layer by layer and obtains feature maps A1 to An in sequence.

[0010] S3. The decoding unit performs layer-by-layer decoding processing on the feature map An. The decoding unit includes the driving factor adaptive module DFAM. The driving factor adaptive module DFAM uses the driving factor feature data to encode and transform to obtain a multi-dimensional conditional vector and performs a channel-by-channel affine transformation on the feature map to generate the modulation parameters of the model in real time. During the decoding process, corresponding skip connections are made and the feature maps of the corresponding layers of the encoding unit are fused to obtain feature maps Bn to B1 in sequence. Feature map B1 is processed by 1×1 convolution and Sigmoid function to output the predicted boundary of the spoil heap.

[0011] S4. Obtain remote sensing image data of the study area. Calculate driving factor feature data based on the driving factor feature library for the remote sensing image data of the study area. Input the remote sensing image data and driving factor feature data of the study area into the trained fusion network model DAMAT-U. The fusion network model DAMAT-U outputs the predicted boundary of the spoil heap.

[0012] To better implement the present invention, in method S1, the remote sensing image data in the waste dump boundary sample dataset consists of remote sensing image data from multiple sources and different types at the same geographical location, with each waste dump in the remote sensing image data corresponding to and labeled with the waste dump boundary; the remote sensing image data also undergoes data preprocessing including atmospheric correction, radiometric calibration, topographic correction, and filtering and denoising.

[0013] Preferably, the driving factors in the driving factor feature library include spectral feature factors, texture feature factors, topographic feature factors, temporal feature factors, and radar feature factors. The spectral feature factors include normalized vegetation index, normalized water index, normalized building index, bare soil index, improved normalized water index, normalized combustion index, red-edge normalized vegetation index, and green-red vegetation index. The texture feature factors include texture homogeneity and texture difference. The topographic feature factors include slope, aspect, topographic curvature, profile curvature, planar curvature, and topographic roughness. The temporal feature factors include seasonal variation amplitude and seasonal variation rate. The radar feature factors include backscattering coefficient VV and backscattering coefficient VH.

[0014] Preferably, the decoding unit of the fusion network model DAMAT-U uses multidimensional conditional vectors to generate scaling parameters γ and bias parameters β that match the number of channels in the feature layer, and uses a feature linear modulation mechanism to perform channel-by-channel affine transformation on the feature maps of each layer of the decoding unit and generate the modulation parameters of the model in real time.

[0015] Preferably, the encoding and decoding units of the fusion network model DAMAT-U are both five-layer structures. The first layer of the encoding unit is downsampled by a 7×7 convolutional layer to obtain the feature map A1. The second to fifth layers of the encoding unit each include a 3×3 convolutional layer, a 2×2 max pooling layer, and a fusion multi-scale feature extraction module EASPPM. The second to fifth layers of the encoding unit perform multi-scale feature extraction layer by layer.

[0016] Preferably, the EASPPM multi-scale feature extraction module includes a 1×1 convolutional layer and five parallel processing branches. The 1×1 convolutional layer performs dimensionality reduction convolution processing on the input feature map and inputs it into the five parallel processing branches. The first processing branch uses 1×1 convolution to extract local detail features. The second, third, and fourth processing branches use 3×3 dilated convolution with dilation rates of 6, 12, and 18, respectively, to extract features. The fifth processing branch first obtains the image-level global context vector through global average pooling, and then restores it to its original spatial size through 1×1 convolution and upsampling operations. The EASPPM multi-scale feature extraction module concatenates the feature maps output by the five parallel processing branches in the channel dimension to form a comprehensive feature map of local to global multi-scale information, and then outputs the feature map through feature fusion processing of the 1×1 convolutional layer.

[0017] Preferably, the fifth layer output feature map A5 of the encoding unit of the DAMAT-U fusion network model is processed by convolution in the fifth layer of the decoding unit to obtain feature map B5; the fourth layer of the decoding unit performs 2×2 upsampling on feature map B5, and then performs skip connections and multi-scale attention mechanism (MAM) enhancement key feature processing with feature map A4 output by the fourth layer of the encoding unit to obtain feature map B4; the third layer of the decoding unit performs upsampling on feature map B4 by dynamically adjusting model parameters using the driving factor adaptive module (DFAM), and then performs skip connections and multi-scale attention mechanism (MAM) enhancement key feature processing with feature map A3 output by the third layer of the encoding unit to obtain feature map B3; the second, first and third layers of the decoding unit are processed in the same way, and the second and first layers of the decoding unit output feature maps B2 and B1 respectively.

[0018] Preferably, the processing method of the multi-scale attention mechanism (MAM) is as follows:

[0019] The Multi-Scale Attention (MAM) mechanism extracts channel attention and spatial attention from the input feature map. For channel attention, features are first compressed using global average pooling, then processed through two fully connected layers and a sigmoid activation function to obtain channel weights. For spatial attention, parallel max pooling and average pooling are used, followed by concatenation of the results, and then 7×7 convolution and a sigmoid activation function to obtain spatial weights. Finally, multi-scale fusion sampling and feature fusion concatenation are performed to output the feature map.

[0020] Preferably, the total loss function expression of the fusion network model DAMAT-U is as follows: ,in For the total loss, To divide the loss, Mean square error, For domain adaptation loss, , , These represent the weights.

[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0022] (1) This invention realizes the deep fusion of multi-source remote sensing data and driving factors, multi-scale feature extraction, attention enhancement and adaptive modulation of driving factors. The fusion network model DAMAT-U is used to achieve adaptive feature weighting through the embedding of driving factors and conditional modulation within the fusion network model DAMAT-U. Finally, the purpose of boundary segmentation of the spoil heap in the research mining area is achieved. It shows excellent performance in boundary segmentation tasks. Through multi-scale feature fusion and edge enhancement mechanism, it effectively solves the problems of boundary blurring, breakage and adhesion, and reduces false detection and missed detection.

[0023] (2) The adaptive driving factor module DFAM of the present invention adopts the adaptive modulation mechanism of the driving factor feature library to realize the dynamic adjustment of network parameters. It can realize the adaptive dynamic adjustment of model parameters based on the set of driving factors, which significantly improves the adaptability and generalization performance of the model, reduces the performance decay caused by environmental differences, and effectively improves the cross-domain applicability of the model. The fusion network model DAMAT-U of the present invention integrates multi-scale convolution, multi-scale feature extraction module EASPPM and dual attention mechanism, which can simultaneously capture the macroscopic morphological features and micro-texture features of the spoil disposal site, effectively solving the problem of multi-scale target recognition.

[0024] (3) This invention realizes a complete automated process from data preprocessing to result output. Through multi-scale feature fusion and edge enhancement mechanism, it effectively solves the problems of boundary blurring, breakage and adhesion, and reduces false detection and missed detection. It greatly improves processing efficiency, and the output results can be directly and seamlessly connected with the existing geographic information system platform, providing reliable technical support for large-scale engineering applications and having important practical application value. Attached Figure Description

[0025] Figure 1 This is a flowchart of the spoil heap boundary identification method of the present invention;

[0026] Figure 2 This is a schematic diagram illustrating the principle of the spoil heap boundary identification method in this embodiment;

[0027] Figure 3 This is a schematic diagram of the principle structure of the fusion network model DAMAT-U in the embodiment;

[0028] Figure 4 This is a schematic diagram illustrating the principle structure of the EASPPM multi-scale feature extraction module in the embodiment.

[0029] Figure 5 This is a schematic diagram of the principle structure of the driving factor adaptive module DFAM in the embodiment;

[0030] Figure 6 This is a schematic diagram of the structure of the multi-scale attention mechanism (MAM) in the embodiment;

[0031] Figure 7 The following are examples of comparison diagrams showing the predicted boundary results of the spoil heap after segmentation using the DAMAT-U fusion network model for three different climate zones in the research mining area. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to embodiments:

[0033] Example

[0034] like Figure 1 As shown, a method for identifying the boundary of a spoil heap combining multiple driving factors and a fusion network model is presented. The method includes:

[0035] S1. Construct a waste dump boundary sample dataset, which includes remote sensing image data and associated labeled waste dump boundaries. In some embodiments, the remote sensing image data in the waste dump boundary sample dataset consists of several different types of multi-source remote sensing image data from the same geographical location. The collected multi-source remote sensing image data undergoes registration processing in a unified coordinate system, with the registration error controlled within 0.5 pixels. Each waste dump in the remote sensing image data corresponds to an associated labeled waste dump boundary. The remote sensing image data also undergoes data preprocessing including atmospheric correction, radiometric calibration, topographic correction, and filtering and denoising. Data preprocessing also includes data resampling processing; continuous data uses bilinear interpolation, and discrete data uses nearest-neighbor interpolation.

[0036] A driving factor feature library containing P driving factors is constructed. Driving factor feature data is then calculated from remote sensing image data based on this library. In some embodiments, the driving factors in the feature library include spectral feature factors, texture feature factors, terrain feature factors, temporal feature factors, and radar feature factors; such as... Figure 2 As shown, the driving factors in the driving factor feature library can also be divided into topographic factors (including topographic feature factors and texture feature factors), climate factors (including temporal feature factors), vegetation indices (including spectral feature factors), and human activity factors. Among them, spectral feature factors include normalized vegetation index, normalized water index, normalized building index, bare soil index, improved normalized water index, normalized combustion index, red-edge normalized vegetation index, and green-red vegetation index. Texture feature factors include texture homogeneity and texture difference. Topographic feature factors include slope, aspect, topographic curvature, profile curvature, planar curvature, and topographic roughness. Temporal feature factors include seasonal variation amplitude and seasonal variation rate. Radar feature factors include backscattering coefficient VV and backscattering coefficient VH.

[0037] This embodiment selects open-pit coal mines in three typical climate zones (arid, semi-arid, and humid regions) as research areas. These three research areas are designated as the arid research area (research area under arid climate conditions), the semi-arid research area (research area under semi-arid climate conditions), and the humid research area (research area under humid climate conditions). Multi-source remote sensing data from 2018 to 2024 were collected for the three research areas. The multi-source remote sensing data includes Sentinel-2 multispectral imagery (10-meter resolution, 12 bands), Landsat 8 thermal infrared imagery (30-meter resolution, 11 bands), Sentinel-1 SAR data (10-meter resolution, VV / VH dual polarization), and Gaofen-2 imagery (0.8-meter panchromatic, 3.2-meter multispectral). Simultaneously, SRTM DEM data (30-meter resolution), WorldClim climate dataset (1-kilometer resolution monthly mean temperature and precipitation data), MODIS vegetation index product (MOD13Q1, 250-meter resolution, 16-day composite), and VIIRS nighttime light data (VNP46A1, 500-meter resolution) were collected. Multi-source remote sensing data of the study area were uniformly registered according to the WGS84 UTM coordinate system. Data preprocessing was performed on the multi-source remote sensing data of the study area, including radiometric calibration and atmospheric correction of optical images, with atmospheric correction parameters calculated using the 6S model. Aerosol optical thickness was obtained using the MODIS atmospheric product (MOD04). SAR data underwent radiometric calibration, thermal noise removal, and topographic correction, with Gamma MAP filtering (3×3 window) used to reduce speckle noise. Topographic factors such as slope, aspect, and curvature were derived from the DEM data using ArcGIS 10.8 software.

[0038] A driving factor feature library was created, which includes spectral feature factors, texture feature factors, terrain feature factors, temporal feature factors, and radar feature factors. Based on this, 8 spectral feature factors, 6 texture feature factors, 6 terrain feature factors, 2 temporal feature factors, and 2 radar feature factors were calculated, for a total of 24 feature driving factors (their sources and descriptions are shown in Table 1). All feature data were resampled to a 10-meter resolution and Z-score standardized. The standardization formula is x'=(x-μ) / σ, where μ and σ represent the mean and standard deviation of the training set, respectively. Texture features were calculated using the Gray-Level Co-occurrence Matrix (GLCM), with average values ​​calculated in four directions (0°, 45°, 90°, 135°). Terrain features were calculated based on SRTMDEM 30m data using a 3×3 moving window. Temporal features were based on MODIS NDVI products (MOD13Q1, 250m resolution) from 2018 to 2024. Radar features were based on Sentinel-1 SAR data using the GRD product, with preprocessing including radiometric calibration and terrain correction. All features were finally resampled to 10m resolution and aligned with Sentinel-2 data.

[0039] Table 1. Sources and explanations of the 24 driving factors

[0040]

[0041]

[0042] In three research mining areas—arid, semi-arid, and humid—boundary sample datasets of spoil heaps were constructed for training the DAMAT-U fusion network model (the datasets were divided into approximately 60% training set, approximately 20% validation set, and approximately 20% test set). Driving factor feature data were also calculated based on the driving factor feature library.

[0043] S2, such as Figures 1-3 As shown, a fusion network model DAMAT-U is constructed with the U-Net network architecture as the core and trained using a waste dump boundary sample dataset. The fusion network model DAMAT-U includes encoding units and decoding units (preferably, such as...). Figure 3 As shown, the encoding and decoding units of the fusion network model DAMAT-U are both five-layer structures. The encoding unit performs multi-scale feature extraction layer by layer from the remote sensing image data and sequentially obtains feature maps A1 to An. In some embodiments, such as Figure 3As shown, the encoding unit has a five-layer structure (i.e., five downsampling stages). The first layer of the encoding unit is downsampled by a 7×7 convolutional layer to obtain the feature map A1. The second to fifth layers of the encoding unit each include a 3×3 convolutional layer, a 2×2 max pooling layer, and a multi-scale feature extraction module EASPPM (this application creatively introduces a multi-scale feature extraction module EASPPM into each layer of the encoding unit. The multi-scale feature extraction module EASPPM uses multiple 3×3 dilated convolutions with different dilation rates in parallel. The multi-scale feature extraction module EASPPM can enhance the multi-scale perception capability and provide key spatial detail information for subsequent fine upsampling). The second to fifth layers of the encoding unit perform multi-scale feature extraction layer by layer. If the input size of a certain remote sensing image in a multi-source remote sensing image dataset is H×W×C (example: H=W=512, C=32), the first layer of the encoding unit undergoes a 7×7 convolutional layer (Conv7×7 represents a convolution operation with a kernel size of 7, a stride of 2, and padding of 3) and downsampling processing (including preliminary feature extraction and downsampling processing, ReLU(BN(Conv7×7(X_input)), where BN is batch normalization and ReLU is the excitation frequency). The live function is used to obtain feature map A1, which has a size of (H / 2)×(W / 2)×64 (i.e., 256×256×64). The second to fifth layers of the coding unit output feature maps A2 to A5 respectively. Feature map A2 has a size of 128×128×128, feature map A3 has a size of 64×64×256, feature map A4 has a size of 32×32×512, and feature map A5 has a size of 32×32×1024.

[0044] In some embodiments, such as Figure 4 As shown, the multi-scale feature extraction module EASPPM includes a 1×1 convolutional layer and five parallel processing branches (by arranging five feature processing paths with different receptive fields in parallel, it captures diverse features of the target from local details to global context). The 1×1 convolutional layer performs dimensionality reduction convolution processing on the input feature map (dimensionality reduction reduces computational complexity) and inputs it into the five parallel processing branches. The first processing branch uses 1×1 convolution to extract local detail features (to preserve the local detail information of the original features). The second, third, and fourth processing branches use 3×3 dilated convolution with dilation rates of 6, 12, and 18, respectively, to extract features. The fifth processing branch first obtains the image-level global context vector through global average pooling, and then restores it to its original spatial size through 1×1 convolution (for non-linear transformation) and upsampling operations. The multi-scale feature extraction module EASPPM concatenates the feature maps output by the five parallel processing branches along the channel dimension to form a comprehensive feature map F containing multi-scale information from local to global. concat=Concat(Branch1,Branch2,Branch3,Branch4,Branch5), and then perform feature fusion processing through a 1×1 convolutional layer to output the feature map.

[0045] S3. The decoding unit performs layer-by-layer decoding processing on the feature map An. The decoding unit includes a driving factor adaptive module DFAM. The driving factor adaptive module DFAM uses the driving factor feature data to encode and transform to obtain a multi-dimensional conditional vector and performs a channel-by-channel affine transformation on the feature map to generate the modulation parameters of the model in real time. During the decoding process, corresponding skip connections are made and the feature maps of the corresponding layers of the encoding unit are fused to obtain feature maps Bn to B1 in sequence. Feature map B1 is processed by 1×1 convolution and the Sigmoid function to output the predicted boundary of the spoil heap. Preferably, the decoding unit of the fusion network model DAMAT-U uses the multi-dimensional conditional vector to generate scaling parameters γ and bias parameters β that match the number of channels of the feature layer. It uses a feature linear modulation mechanism to perform a channel-by-channel affine transformation on the feature maps of each layer of the decoding unit and generates the modulation parameters of the model in real time.

[0046] In some embodiments, the decoding unit has a five-layer structure (preferably, both the encoding and decoding units of the fusion network model DAMAT-U are five-layer structures). The feature map A5 output from the fifth layer of the encoding unit of the fusion network model DAMAT-U is processed by convolution in the fifth layer of the decoding unit to obtain feature map B5. The fourth layer of the decoding unit performs 2×2 upsampling on feature map B5, and then performs skip connections and multi-scale attention mechanism (MAM) enhancement on feature map A4 output from the fourth layer of the encoding unit to obtain feature map B4. The third layer of the decoding unit performs upsampling on feature map B4 using the driving factor adaptive module (DFAM) to dynamically adjust the model parameters, and then performs skip connections and multi-scale attention mechanism (MAM) enhancement on feature map A3 output from the third layer of the encoding unit to obtain feature map B3. The second, first, and third layers of the decoding unit are processed in the same way, and the second and first layers of the decoding unit output feature maps B2 and B1, respectively.

[0047] In some embodiments, the processing method of the multi-scale attention mechanism (MAM) is as follows: Figure 6 As shown, the Multi-Scale Attention (MAM) mechanism extracts the input feature map (e.g., Figure 6 As shown, assuming the feature map F has dimensions C×H×W, the first branch for channel attention first compresses the features using global average pooling, then processes them through two fully connected layers and a sigmoid activation function to obtain the channel weights. The fully connected processing expression is as follows: s=σ(W2·δ(W1·z)), where... , r=16, where r is the compression ratio and δ represents the ReLU activation function. The second branch processes spatial attention through parallel max pooling and average pooling, then concatenates the results, followed by a 7×7 convolution and a Sigmoid activation function to obtain spatial weights. Finally, multi-scale fusion sampling and feature fusion are performed on the two branches to output a feature map, F'=s_channel⊙F⊙s_spatial, where s_channel represents the channel attention weights and s_spatial represents the spatial attention weights.

[0048] In some embodiments, such as Figure 5 As shown, the Driver Factor Adaptive Module (DFAM) deeply integrates driver factors from the driver factor feature library into the segmentation network using conditional embedding and Feature Linear Modulation (FiLM) techniques. It generates modulation parameters in real time based on specific, dynamically input factor vectors, enabling dynamic adjustment of parameters according to different climatic zone environmental conditions. Figure 5 As shown, the Driving Factor Adaptive Module (DFAM) uses a three-layer fully connected network to process the 24-dimensional environmental factor vector. The first fully connected layer is: h1 = σ(BN(W1·x + b1)), where... , Input x∈R 24 The second fully connected layer: h2 = σ(BN(W2·h1 + b2)), , The third fully connected layer: h3 = σ(BN(W3·h2+b3)), , After outputting the 64-dimensional conditional vector c, modulation parameters are generated through linear transformation: γ = Wγ·c + bγ, β = Wβ·c + bβ, where Wγ, bγ, C represents the number of feature map channels. A linear characteristic modulation (FiLM) mechanism is employed: FiLM(F) = γ⊙F + β, where... The input feature map ⊙ represents channel-wise multiplication. This operation is applicable to different climate regions (the three research mining areas listed above), dynamically adjusts the feature response within the network, and significantly improves the model's boundary recognition accuracy and generalization ability under various climate zones such as arid, semi-arid, and humid.

[0049] S4. Acquire remote sensing image data of the study area. Calculate driving factor feature data from the remote sensing image data of the study area according to the driving factor feature library. Input the remote sensing image data and driving factor feature data of the study area into the trained fusion network model DAMAT-U. The fusion network model DAMAT-U outputs the predicted boundary of the spoil heap. The fusion network model DAMAT-U performs boundary segmentation head prediction processing on the feature map obtained by the decoding unit. The boundary segmentation head consists of a 1×1 convolutional layer, which maps the 64-channel feature map into 1 channel and compresses the output value of each pixel to between 0 and 1 through the Sigmoid activation function, directly generating the predicted spoil heap boundary.

[0050] In some embodiments, the total loss function expression for the fusion network model DAMAT-U is as follows: ,in For the total loss, To divide the loss, Mean square error, For domain adaptation loss, , , They represent the weights, , , Examples of possible values ​​are as follows: =1.0, =0.5, =0.5. Segmentation loss , For Dice's loss, , where ε is the smoothing term, taking the value of , The prediction result for pixel i. This represents the actual result for pixel i; For cross-entropy loss, Mean square error , To predict edge results, Real edge labels. Domain adaptation loss. , For the maximum mean difference loss (MMD), To combat the losses. , Denotes the square norm in the RKHS space. To map the data to the eigenfunctions of the reproducing kernel Hilbert space (RKHS), Given an expected function, find its expectation. For source domain data, For target domain data.

[0051] The network parameter settings for the DAMAT-U converged network model are shown in Table 2, and the server configuration is shown in Table 3.

[0052]

[0053]

[0054] The fusion network model DAMAT-U employs a two-stage training strategy. The first stage, pre-training, uses all source domain data for 100 epochs with an initial learning rate of 0.001, employing the CosineAnnealingLR scheduler with a batch size of 16 and mixed precision training. The second stage, fine-tuning, uses target domain data for 50 epochs, reducing the learning rate to 0.0001 and introducing a gradient inversion layer (GRL). The gradient inversion coefficient λ = 2 / (1 + exp(-10•p)) - 1, where p is the training progress percentage. An early stopping strategy is used, stopping training when the validation set loss no longer decreases within 15 epochs. The fusion network model DAMAT-U of this invention uses Intersection over Union (IoU), F1 score, accuracy, and recall as evaluation metrics, and performs quantitative evaluation on the test set. Experimental results show that the predicted waste dump boundaries in the three mining areas of this invention are as follows: Figure 7 As shown, the three research mining areas in this invention achieved superior performance, with an average IoU of 85.2%, an F1 score of 91.5%, an accuracy of 94.3%, and a recall of 89.7%, significantly outperforming all existing methods. Ablation experiments were also conducted to verify the effectiveness of the multi-scale attention mechanism, the driver factor adaptive module, and the domain adaptation strategy.

[0055] In this embodiment, the DAMAT-U fusion network model is used for model inference. The target area image is sliced ​​according to the preprocessing procedure and input into the trained DAMAT-U fusion network model to obtain the predicted spoil heap boundary. Optimization processing involves thresholding the probability map (0.5) to generate a binary mask. Noise and holes are removed through morphological opening and closing operations (kernel size 3×3). Connectivity analysis is performed using 8-neighborhood connections, with an area filtering threshold of 500 square meters (corresponding to 50 pixels), filtering small regions and merging adjacent polygons. The final output includes a vectorized boundary file (Shapefile format), a risk level raster map (GeoTIFF format), and a driving factor contribution table (CSV format). A visual analysis report is also generated to support GIS integration and decision-making applications.

[0056] 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 identifying the boundary of a spoil heap combining multiple driving factors and a fusion network model, characterized in that: The methods include: S1. Construct a waste dump boundary sample dataset, which includes remote sensing image data and associated labeled waste dump boundaries; create a driving factor feature library containing P driving factors, including spectral feature factors, texture feature factors, topographic feature factors, temporal feature factors, and radar feature factors. The spectral feature factors include normalized vegetation index, normalized water index, normalized building index, bare soil index, improved normalized water index, normalized combustion index, red-edge normalized vegetation index, and green-red vegetation index. The texture feature factors include texture homogeneity and texture heterogeneity. The topographic feature factors include slope, aspect, topographic curvature, profile curvature, planar curvature, and topographic roughness. The temporal feature factors include seasonal variation amplitude and seasonal variation rate. The radar feature factors include backscattering coefficient VV and backscattering coefficient VH. Calculate and obtain driving factor feature data from the remote sensing image data based on the driving factor feature library. S2. The DAMAT-U fusion network model is constructed with the U-Net network architecture as the core and trained using the waste dump boundary sample dataset. The DAMAT-U fusion network model includes an encoding unit and a decoding unit. The encoding unit performs multi-scale feature extraction on the remote sensing image data layer by layer and obtains feature maps A1 to An in sequence. S3. The decoding unit performs layer-by-layer decoding processing on the feature map An. The decoding unit includes a driving factor adaptive module DFAM. The driving factor adaptive module DFAM uses the driving factor feature data to encode and transform to obtain a multi-dimensional conditional vector, and uses the multi-dimensional conditional vector to generate scaling parameter γ and bias parameter β that match the number of channels in the feature layer. It uses a feature linear modulation mechanism to perform channel-by-channel affine transformation on the feature maps of each layer of the decoding unit and generates the modulation parameters of the model in real time. The encoding and decoding units of the fusion network model DAMAT-U are both five-layer structures. The first layer of the encoding unit is downsampled by a 7×7 convolutional layer to obtain the feature map A1. The second to fifth layers of the encoding unit each include a 3×3 convolutional layer, a 2×2 max pooling layer, and a multi-scale feature extraction module EASPPM. The second to fifth layers of the encoding unit perform multi-scale feature extraction layer by layer. During the decoding process, corresponding skip connections are made and the feature maps of the corresponding layers of the encoding unit are fused to obtain feature maps Bn to B1 in sequence. Feature map B1 is processed by a 1×1 convolution and a sigmoid function to output the predicted boundary of the spoil heap. S4. Obtain remote sensing image data of the study area. Calculate driving factor feature data based on the driving factor feature library for the remote sensing image data of the study area. Input the remote sensing image data and driving factor feature data of the study area into the trained fusion network model DAMAT-U. The fusion network model DAMAT-U outputs the predicted boundary of the spoil heap.

2. The method for identifying spoil heap boundaries by combining multiple driving factors and a fusion network model according to claim 1, characterized in that: In method S1, the remote sensing image data in the waste dump boundary sample dataset consists of several different types of multi-source remote sensing image data from the same geographical location. The waste dump boundary of each remote sensing image data is associated and labeled. The remote sensing image data also undergoes data preprocessing including atmospheric correction, radiometric calibration, topographic correction, and filtering and denoising.

3. The method for identifying spoil heap boundaries combining multiple driving factors and a fusion network model according to claim 1, characterized in that: The EASPPM multi-scale feature extraction module includes a 1×1 convolutional layer and five parallel processing branches. The 1×1 convolutional layer performs dimensionality reduction convolution on the input feature map and inputs it into the five parallel processing branches. The first processing branch uses 1×1 convolution to extract local detail features. The second, third, and fourth processing branches use 3×3 dilated convolution with dilation rates of 6, 12, and 18, respectively, to extract features. The fifth processing branch first obtains the image-level global context vector through global average pooling, and then restores it to its original spatial size through 1×1 convolution and upsampling operations. The EASPPM multi-scale feature extraction module concatenates the feature maps output by the five parallel processing branches in the channel dimension to form a comprehensive feature map of local to global multi-scale information, and then outputs the feature map through feature fusion processing of the 1×1 convolutional layer.

4. The method for identifying spoil heap boundaries by combining multiple driving factors and a fusion network model according to claim 1 or 3, characterized in that: The fifth layer output feature map A5 of the encoding unit of the DAMAT-U fusion network model is processed by convolution in the fifth layer of the decoding unit to obtain feature map B5. The fourth layer of the decoding unit performs 2×2 upsampling on feature map B5, and then performs skip connections and multi-scale attention mechanism (MAM) enhancement on feature map A4 output from the fourth layer of the encoding unit to obtain feature map B4. The third layer of the decoding unit performs upsampling on feature map B4 by dynamically adjusting the model parameters using the driving factor adaptive module (DFAM), and then performs skip connections and multi-scale attention mechanism (MAM) enhancement on feature map A3 output from the third layer of the encoding unit to obtain feature map B3. The second, first and third layers of the decoding unit are processed in the same way, and the second and first layers of the decoding unit output feature maps B2 and B1, respectively.

5. The method for identifying spoil heap boundaries by combining multiple driving factors and a fusion network model according to claim 4, characterized in that: The processing method of the Multi-Scale Attention Mechanism (MAM) is as follows: The Multi-Scale Attention (MAM) mechanism extracts channel attention and spatial attention from the input feature map. For channel attention, features are first compressed by global average pooling, and then processed by two fully connected layers and a sigmoid activation function to obtain channel weights. Spatial attention is processed by parallel max pooling and average pooling, and the results are concatenated. Then, it is processed by 7×7 convolution and sigmoid activation function to obtain spatial weights. Then, multi-scale fusion sampling and feature fusion splicing are performed to output a feature map.

6. The method for identifying spoil heap boundaries by combining multiple driving factors and a fusion network model according to claim 1, characterized in that: The total loss function expression of the fusion network model DAMAT-U is as follows: ,in For the total loss, To divide the loss, Mean square error, For domain adaptation loss, , , These represent the weights.

Images