CN-121214246-B - Refuse dump boundary identification method combining multiple driving factors and fusion network model

CN121214246BCN 121214246 BCN121214246 BCN 121214246BCN-121214246-B

Abstract

The invention discloses a method for identifying a boundary of a dumping site by combining multiple driving factors and a fusion network model, which comprises the steps of S1, constructing a dumping site boundary sample data set and manufacturing a driving factor feature library, S2, constructing a fusion network model DAMAT-U, extracting the characteristics of remote sensing image data layer by layer and multiple scales, sequentially obtaining characteristic diagrams A1-An, S3, decoding the characteristic diagrams An layer by a decoding unit, wherein the decoding unit comprises a driving factor self-adaptive module DFAM, the driving factor self-adaptive module DFAM generates modulation parameters of the model in real time by using the driving factor characteristic data, and correspondingly jump-connects and fuses the characteristic diagrams of the corresponding layers of the encoding unit in the decoding process, and S4, obtaining the fusion network model DAMAT-U after the remote sensing image data of a research area is input and trained, and predicting the boundary of the dumping site. The invention adopts depth fusion, multi-scale feature extraction, attention enhancement and driving factor self-adaptive modulation to realize the aim of researching the boundary segmentation of the dump of the mining area.

Inventors

ZHANG HAN
PU LIJUN
CHU TIANKUO
Xing Jianghe
LI JUN
ZHANG CHENGYE
YOU LIN

Assignees

中国矿业大学(北京)
北京数论科技有限公司

Dates

Publication Date: 20260512
Application Date: 20250925

Claims (6)

1. A method for identifying a boundary of a dumping site by combining multiple driving factors and a fusion network model is characterized by comprising the following steps: S1, constructing a soil discharge field boundary sample data set, wherein the soil discharge field boundary sample data set comprises remote sensing image data and associated marked soil discharge field boundaries, manufacturing a driving factor feature library comprising P driving factors, wherein the driving factors of the driving factor feature library comprise spectral feature factors, texture feature factors, topography feature factors, time sequence feature factors and radar feature factors, the spectral feature factors comprise normalized vegetation indexes, normalized water indexes, normalized building indexes, bare soil indexes, improved normalized water indexes, normalized combustion indexes, red edge normalized vegetation indexes and green and red vegetation indexes, the texture feature factors comprise texture homogeneity and texture difference, the topography feature factors comprise gradient, slope direction, topography curvature, profile curvature, plane curvature and topography roughness, the time sequence feature factors comprise seasonal variation amplitude and seasonal variation rate, the radar feature factors comprise backscattering coefficients VV and backscattering coefficients VH, and obtaining driving factor feature data based on the driving factor feature library calculation for remote sensing image data; S2, constructing a fusion network model DAMAT-U by taking a U-Net network architecture as a core, and performing learning training by utilizing a soil discharge field boundary sample data set, wherein the fusion network model DAMAT-U comprises a coding unit and a decoding unit, and the 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 carries out layer-by-layer decoding processing on a feature map An, the decoding unit comprises a driving factor self-adapting module DFAM, the driving factor self-adapting module DFAM carries out coding and transformation by utilizing driving factor feature data to obtain a multi-dimensional condition vector, and generates scaling parameters gamma and bias parameters beta matched with the channel number of the feature map layer by utilizing the multi-dimensional condition vector, and carries out channel affine transformation on each layer of feature map of the decoding unit by utilizing a feature linear modulation mechanism and generates modulation parameters of a model in real time; S4, acquiring remote sensing image data of the research area, calculating the remote sensing image data of the research area according to the driving factor feature library to acquire driving factor feature data, inputting the remote sensing image data of the research area and the driving factor feature data into a trained fusion network model DAMAT-U, and outputting a predicted dumping site boundary corresponding to the fusion network model DAMAT-U.
2. The method for recognizing the boundary of the dumping site by combining the multiple driving factors and the fusion network model according to claim 1, wherein in the method S1, remote sensing image data in a sample data set of the boundary of the dumping site are remote sensing image data of the same geographic position and different types and multiple sources, the dumping site of each remote sensing image data corresponds to the associated marked dumping site boundary, and the remote sensing image data are subjected to data preprocessing including atmosphere correction, radiometric calibration, topography correction and filtering denoising.
3. The method for recognizing the boundary of the dump by combining the multi-driving factors and the fusion network model according to claim 1, wherein the fusion multi-scale feature extraction module EASPPM comprises a1×1 convolution layer and five parallel processing branches, the 1×1 convolution layer performs dimension reduction convolution processing on the input feature map and inputs the five parallel processing branches respectively, the first processing branch extracts local detail features by using 1×1 convolution, the second, third and fourth processing branches extract features by using 3×3 cavity convolution with expansion rates of 6, 12 and 18 respectively, the fifth processing branch acquires global context vectors of image levels through global average pooling, and restores the global context vectors to original space dimensions through 1×1 convolution and up-sampling operations, the fusion multi-scale feature extraction module EASPPM splices feature maps output by the five parallel processing branches on channel dimensions 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 convolution layer.
4. The method for recognizing the boundary of the dumping field by combining the multi-driving factors and the fusion network model according to claim 1 or 3 is characterized in that a fifth layer output characteristic diagram A5 of an encoding unit of the fusion network model DAMAT-U is subjected to convolution processing of a fifth layer of a decoding unit to obtain a characteristic diagram B5, a fourth layer of the decoding unit carries out 2X 2 up-sampling processing on the characteristic diagram B5, then carries out jump connection with a characteristic diagram A4 output by the fourth layer of the encoding unit and enhancement key characteristic processing of a multi-scale attention mechanism MAM to obtain a characteristic diagram B4, a third layer of the decoding unit carries out up-sampling processing of dynamically adjusting model parameters by using a driving factor adaptive module DFAM, then carries out jump connection with a characteristic diagram A3 output by the third layer of the encoding unit and enhancement key characteristic processing of the multi-scale attention mechanism MAM to obtain a characteristic diagram B3, and the second layer, the first layer and the second layer of the decoding unit respectively output the characteristic diagrams B2 and B1 of the decoding unit are identical to the processing method of the third layer.
5. The method for identifying the boundary of the dump by combining the multiple driving factors and the fusion network model according to claim 4, wherein the method for processing the multiscale attention mechanism MAM is as follows: The multi-scale attention mechanism MAM extracts channel attention and space attention of an input feature map, firstly, compressed features are pooled through global average, channel weight is obtained through two layers of full-connection processing and Sigmoid activation function processing, result splicing processing is carried out on the space attention after parallel maximum pooling and average pooling processing, space weight is obtained through 7X 7 convolution and Sigmoid activation function processing, and then multi-scale fusion sampling and feature fusion splicing are carried out to output the feature map.
6. The method for identifying a boundary of a dump by combining multiple driving factors and a fusion network model according to claim 1, wherein the fusion network model DAMAT-U has a total loss function expression as follows: wherein In order to account for the total loss, In order to divide the loss of the device, In the form of a mean square error, In order to accommodate the loss of domain, 、、 Respectively, the weights.

Description

Refuse dump boundary identification method combining multiple driving factors and fusion network model Technical Field The invention relates to the field of boundary detection of a dumping site in an open-air mining area, in particular to a dumping site boundary identification method combining multiple driving factors and a fusion network model. Background The dumping sites (tailing ponds, waste rock piles and the like) generated in the surface mine exploitation process occupy a large amount of land resources, are easily subjected to geological disasters such as slope instability, landslide, debris flow and the like due to the special loose accumulation structure and the influence of external environmental factors (such as rainfall, weathering, earthquake and the like), and seriously threaten the production safety of mining areas, the surrounding ecological environment and the life and property safety of people. The dumping sites of different weather zones (such as arid areas and humid areas) show great differences in remote sensing images, for example, the vegetation coverage of the arid areas is less, the spectral characteristics of the dumping sites are prominent, but the dumping sites are easily confused with bare lands and rocks. Wet zone vegetation flourishes, the dump may be covered by a portion of the vegetation, the boundaries become spectrally blurred, but texture and topographical features may be more pronounced. Therefore, the high-precision and automatic interpretation of the boundaries of the dumping site is realized, and the method has great significance for safe production of mines, ecological environment protection and geological disaster prevention and control. Traditional dump monitoring mainly relies on field measurement and manual interpretation, is high in cost, time-consuming and labor-consuming, and is difficult to meet the requirements of large-scale dynamic real-time monitoring. Modern high-resolution remote sensing technology has macroscopic and dynamic advantages, and can effectively make up the defects of the traditional method. Semantic segmentation models (such as U-Net and improvement thereof) based on deep learning are widely applied to mining area target extraction and ground object recognition, and boundary detection is achieved through automatic learning of remote sensing image features. For example, U-Net models based on multi-scale sample sets have been used for mining industry solid waste and open stope identification, and significantly improve identification accuracy and generalization ability. The multi-source remote sensing data fusion model utilizes a single depth network to process multi-channel images simultaneously, and the multi-source remote sensing data fusion model also proves that the multi-source remote sensing data fusion model can effectively utilize multi-source information to enhance the segmentation effect. However, the prior art still has the defects in the monitoring of the soil discharge field (1) the prior model is mostly based on single-scale or single-image source training, the soil discharge pile bodies with different forms and sizes are not fully represented, the boundary segmentation is often blurred or missing, and the recognition accuracy is limited. (2) The environmental driving factors are ignored, different weather zones (such as drought, semi-drought, wetting and the like) are obviously different in vegetation coverage, soil moisture, lithologic weathering and the like, the weather factors (such as precipitation and temperature) are main driving of the change of surface vegetation and landform, and if the heterogeneous characteristics are not fully considered in model training, the cross-domain performance is reduced. Therefore, a new method capable of deeply fusing multi-source data, adapting environmental driving factors and optimizing boundary recognition tasks is needed. Disclosure of Invention The invention aims to solve the technical problems pointed out by the background technology, and provides a method for identifying the boundary of the dump by combining multiple driving factors and a fusion network model, which realizes deep fusion of multi-source remote sensing data and driving factors, multi-scale feature extraction, attention enhancement and driving factor self-adaptive modulation, and realizes self-adaptive feature weighting by fusion network model DAMAT-U through fusion network model DAMAT-U and embedding of driving factors and conditional modulation in fusion network model DAMAT-U, thereby finally realizing the purpose of studying the boundary segmentation of the dump in a mining area. The aim of the invention is achieved by the following technical scheme: A method for identifying a boundary of a dumping site by combining multiple driving factors and a fusion network model comprises the following steps: S1, constructing a soil discharge field boundary sample data set which comprises remote sensing image data and associa