The content of this article is from the "Surveying and Mapping Bulletin" No. 2, 2024, drawing review number: GS Jing (2024) No. 0287
Stope information extraction from Weining Beishan open-pit mine by combining deep learning and object-oriented analysis
LIU Li1
, LI Shiyao2
, WANG Run3,4, LIU Shaoyu1, SONG Yongfei1, NIU Ruiqing5
1. Ningxia Hui Autonomous Region Land and Resources Survey and Monitoring Institute, Yinchuan, Ningxia 750002, China;2. Wuhan Geological Survey Center, China Geological Survey, Wuhan, China 430205;3. Hubei Provincial General Station of Geological Environment, Wuhan, Hubei 430034;4. Hubei Provincial Key Laboratory of Resources and Eco-Environmental Geology, Wuhan, Hubei 430034;5. School of Geophysics and Spatial Information, China University of Geosciences, Wuhan 430074, Hubei, China
Funds: Ningxia Natural Science Foundation (2021AAC03432; 2021AAC03431); Key R&D Program of Ningxia Hui Autonomous Region(2021BEG03001); Geological Survey Project of China Geological Survey (DD20211391); Science and Technology Project of Hubei Provincial Key Laboratory of Resources and Eco-Environmental Geology (Hubei Provincial Bureau of Geology)(KJ2023-18)
Keywords: Weining Beishan, open-pit mine stope, deep learning, object-oriented, remote sensing monitoring
Citation format: LIU Li, LI Shiyao, WANG Run, et al. Stope Information Extraction of Weining Beishan Open-pit Mine Based on Deep Learning and Object-Oriented Analysis[J]. Bulletin of Surveying and Mapping,2024(2):51-57. DOI: 10.13474/j.cnki.11-2246.2024.0209
Abstract:The Beishan area of Weining is a key node area in Ningxia to carry out the restoration and management of mine ecological environment. It has become one of the important tasks of mine administration and management in the autonomous region to quickly and accurately obtain the spatial information of stopes in the region and monitor the construction progress of mine ecological restoration projects. In this paper, we take the Beishan area of Weining as the research area, and propose a method for extracting stope information from domestic high-resolution remote sensing satellite images of open-pit mines based on deep learning and object-oriented analysis. Firstly, the U-Net model supporting small-shot learning is used for the initial identification of open-pit stopes. Then, the object-oriented analysis and spatial analysis methods are combined to realize the fine extraction of the open-pit stope boundary. It is verified that the accuracy of the method is 0.71 for identifying the spatial location of the open-pit stope, and the extraction accuracy for the average spatial range is 0.78. On this basis, the restoration and management of open-pit mines in the Beishan area of Weining from 2019 to 2021 were identified and analyzed, and 43.2% of the 125 open-pit mine stopes identified had carried out ecological restoration projects, including 44 pit landfills and soil leveling, 6 redevelopment and utilization, and 4 artificial regreening. The results show that the proposed method can quickly extract the spatial information of the open-pit stope without the need for feature engineering, which can provide a technical reference for remote sensing monitoring of mines in Ningxia.
Ningxia is a typical resource-based region, and mining is the basic industry of its national economic and social development, and the rational exploration, development, utilization, and protection of mineral resources are related to the construction of ecological civilization and high-quality development [1]. The Beishan area of Weining is one of the key mining areas for mineral resources in Ningxia, and minerals such as iron ore, gypsum ore, clay ore, and construction sand and gravel are mainly mined in the open pit [2-3]. Weining Beishan is located at the southeast edge of the Tengger Desert, the ecological environment conditions are fragile, and the human disturbance of mineral resources development and engineering construction in the area is easy to cause damage to the ecological environment. In the past, the supervision of mineral resources development was mainly carried out by means of field investigation, step-by-step inspection and statistical reporting, and public supervision and reporting, which had low work efficiency and strong information subjectivity, and the work was affected by factors such as the environmental conditions of the mining area and the cooperation of the mining party, which had a long cycle and was easy to omission, and could not timely and accurately discover the environmental and geological problems of the mine, and the rectification effect of illegal and illegal mining behaviors was not satisfactory [4]. With the development of earth observation technology, especially the operational operation of domestic high-resolution remote sensing satellites in recent years, remote sensing technology has developed into an important means for the development and monitoring of mineral resources, the investigation and monitoring of mine geological environment, and the monitoring of ecological environment [5-11]. Since 2006, the China Geological Survey has deployed remote sensing monitoring of key mines, and since 2011, it has been transferred to the national remote sensing investigation and monitoring of terrestrial mineral resources development and mine geological environment [12], and experts and scholars from various provinces have successively carried out relevant research work, applying high-resolution optical remote sensing and radar remote sensing technologies to remote sensing surveys of mineral development, monitoring of illegal mining in mines, and ecological environment management in mining areas [13-18]. In the past, the acquisition of typical surface elements in mining areas mainly relied on expert visual interpretation and human-computer interaction information extraction, and the automation and intelligence of technical methods were still insufficient. Deep learning enables the hierarchical learning of the most representative and separable features of a dataset in an end-to-end manner [19-20]. Different from the classical machine learning algorithm, which requires expert experience to construct and screen target features, deep learning can independently learn sample features without manual feature construction or rule design, which effectively improves the degree of automation and intelligence. Some studies have explored the application of deep learning algorithms such as CNN, FCN, SegNet, and DeepLabv3+ to land use classification in mining areas [21-24]. However, the intelligent method represented by deep learning is still in the exploration stage of method applicability in the extraction of surface element information in mining areas, and it is necessary to combine regional characteristics such as mineral type, geological conditions, and surface environment to construct localized samples and models for extracting typical surface elements in mining areas, so as to promote the engineering application of deep learning in mine remote sensing monitoring. In this paper, taking Beishan, Weining, Ningxia as the research area, using domestic high-resolution remote sensing satellites as the main data source, combined with deep learning and object-oriented analysis methods, a localized extraction method for stope information in open-pit mines was established, so as to provide technical support for the autonomous region to crack down on illegal mining and monitor the restoration and governance of the ecological environment in mining areas. 1 Overview of the study areaThe study area is located in the Beishan area of Weining, north of Zhongwei City, Ningxia, as shown in Figure 1. The study area covers an area of about 290 km2, the altitude is about 1200~1600 m, the landform is mainly low, medium and hilly, and belongs to the temperate continental arid climate zone, with an average annual precipitation of about 180 mm, and the vegetation type is grassland desert.
Fig.1 Overview of the study areaThe strata in the Beishan area of Weining are mainly composed of Devonian and Carboniferous, and are composed of clastic-carbonate rocks with hydrocontinental alternating facies and gypsum-salts. Among them, the mineralization of quartz diorite intruded into the Upper Devonian and Lower Carboniferous strata during the Indosinian-Yanshan period is of great significance, and the Upper Devonian Laojunshan Formation and the Lower Carboniferous Pre-Montenegro Formation are the main ore-bearing strata in this area. The study area is dominated by copper and gold mineralization, with symbiotic or associated lead, zinc, iron, and cobalt mineralization, which is an important prospecting area for polymetallic ore in the autonomous region [25]. 2 Extraction of stope information from open-pit mines2.1 Data sourceThe remote sensing data used in this paper is the L1A data of GF-2 satellite imagery, and the spatial resolution of multispectral bands (B1: 0.45~0.52 μm, B2: 0.52~0.59 μm, B3: 0.63~0.69 μm, B4: 0.77~0.89 μm) is 4 m, and the spatial resolution of panchromatic bands is 1 m, and the imaging time is July 25, 2019. The L1A level data were preprocessed by radiometric calibration, atmospheric correction, orthorectification, multispectral and panchromatic image fusion, projection transformation, georeferencing and other preprocessing to obtain a digital orthophoto with a spatial resolution of 1 m covering the study area, with a coordinate system of CGCS2000 and a projection method of Gauss-Krüger 3-degree zone projection. Mineral resource development areas often have a variety of mine development land types, such as stopes, mine buildings, transit sites and solid wastes. In this paper, the land use/land cover data of Sentinel-2 10 m in 2019 [26] were used to assist in identifying the type of patches occupied by mine development. As shown in Figure 1, the main types of land cover in the study area are shrubs, buildings, cultivated land, bare land and a small amount of water. 2.2 Research MethodsThe main technical processes in this paper include the establishment of stope interpretation signs in open-pit mines, the extraction of stope spatial information, the object-oriented segmentation of remote sensing images, and the extraction of stope vector boundaries, and the technical route is shown in Figure 2. Secondly, the U-Net algorithm was used to classify the remote sensing images and obtain the spatial distribution of the stopes, and then the multi-scale segmentation algorithm was used to segment the remote sensing images and the spatial analysis of the image objects to obtain the stope vector boundary.
Fig. 2 Technical route 2.2.1 Open-pit mine stope interpretation markEstablishing an interpretation mark is the basis for interpreting open-pit mine stope and establishing model training samples. The mineral resources mined in the open pit in the study area are iron ore, clay ore (bentonite) and building stone. Based on the results of previous work and the mining management data in the area, this paper used GF-2 images in 2019 to interpret 138 typical open-pit mine stope patches, and randomly selected 75 to construct training samples, the samples were evenly distributed in the study area, and the rest of the patches were used as verification samples. The interpretation signs are shown in Table 1.
2.2.2 Stope spatial information extractionDeep learning avoids the task of relying on expert experience to construct the feature set of the dataset, and can provide an end-to-end classification model for information extraction. In this paper, the lightweight network model U-Net [27] is used to train the extraction model using the interpreted open-pit stope patches to obtain the spatial location and scope of the open-pit stope. The advantage of this model is that it supports model training under small sample conditions, and the network parameters are few, which is convenient for engineering applications. The U-Net structure is shown in Figure 3, the training sample size in the model is 256×256 pixels, and the tiles are generated in steps of 128 pixels in the interpretation area. In order to obtain sufficient model training samples, the angle rotation method was used to augment the training samples, that is, the original samples were rotated by 90°, 180° and 270° respectively, and a total of 4860 pairs of training samples were obtained. An example of a sample is shown in Figure 4.
Figure 3 U-Net structure
Fig. 4 Sample example of open-pit stope information extraction2.2.3 Object-oriented segmentation of remote sensing imagesObject-oriented segmentation of remote sensing images is a process of dividing the image scene into several sub-regions based on the homogeneity or heterogeneity criterion according to the differences in image features of ground objects. There are many types of land occupied by mine development in the study area, and the spectrum, texture and geometric characteristics of the patches are obviously different, and it is not easy to directly segment a single object that can fully express the contour information of the land occupied by mine development. In this paper, the segmentation scale, shape factor and compactness variables are controlled by using the interpretation patch as a reference, and the segmentation results under different parameter combinations are generated. When there is a high degree of overlap between one or more object combination boundaries and visual interpretation boundaries, and it meets the inflection point estimated by ESP, a quantitative evaluation tool for scale parameters, it is regarded as the optimal segmentation parameter [28-30]. In order to avoid the segmentation results being too "fragmented", the fewer segmented objects in the open-pit mining activity range are as few as possible under the optimal coincidence degree, that is, the more complete the objects, the better. Through ESP tools and human-computer interaction, the scale, shape, and compactness were finally selected as 50, 0.4, and 0.5 as the segmentation parameters, respectively. About 237,900 objects were obtained from the segmentation of the study area, and the local segmentation effect is shown in Figure 5.
Fig. 5 GF-2 image segmentation effect under optimal parameters2.2.4 Pixel-based deep learning model is used for stope vector boundary extraction, which is prone to fracture areas and voids, while object-oriented segmentation can provide objects with relatively consistent spectra, texture, and other characteristics. In this paper, the advantages of the two are combined to superimpose the results of image segmentation and pixel extraction, and the segmented objects of the intersecting regions are retained, as shown in Figure 6. In order to better summarize the stope boundary of open-pit mines, the objects that intersect with buildings, cultivated land and water bodies are removed by combining auxiliary data, and the objects with an area of less than 50 m2 or less than 10% of the area of the intersection area are removed by automatic filtering, as shown in Figure 6, the area of the intersection area in objects A, B, C, and D is less than 10%, and the objects A and D are only intersected at the edge, and the intersection area is 50~150 m2 or the area of the intersection area is 10%~20% by human-computer interaction retains the remaining intersecting objects and merges neighboring objects.
Fig. 6 Example of intersection analysis and result screening2.2.5 Accuracy verification of extraction results Accuracy evaluation of stope boundary extraction results in open-pit mines mainly adopts two methods: object-based evaluation and pixel-based evaluation. The former is used to analyze the accuracy of stope location extraction in open-pit mines, while the latter is used to evaluate the accuracy of stope boundary extraction in open-pit mines. Object-based evaluations are calculated using an F1-score:
(1) where p is the accuracy of the check, which is calculated from the true positive and false positive results, and r is the recall rate, which is calculated from the true positive and false negative results. Pixel-based evaluation uses area similarity and is calculated as follows:
(2) where inp and pre are the interpretation boundary and prediction boundary, respectively, and Area() is the area calculation function. When there is 1:N, N:1 or M∶N between the manual interpretation boundary and the extraction boundary, the total area of multiple objects is combined. 3 Result analysis3.1 Accuracy analysis of identification resultsA total of 152 open-pit stope patches were identified in this method. Among them, there are 77 newly identified stope patches, 27 wrong spots, and 13 missing spots, and the recognition accuracy F1 is 0.71. Some of the recognition results are shown in Figure 7. The proposed method combines the advantages of U-Net and object-oriented analysis, and can better identify and delineate the stope boundary of open-pit mines. Compared with the single U-Net network method, the stope boundary extracted by this method contains more complete stope information, which is more similar to the manual interpretation boundary, and has more engineering application prospects.
Fig. 7 Correct Spot Identification In order to quantify the accuracy of the identified stope boundary of the open-pit mine, this paper further calculates the area similarity between the correct identification and the manual interpretation of the patches. As shown in Figure 8, the average similarity of the patches was 0.78, 27.63% of the patches were more than 0.90, and only 6 mines had a spot similarity of less than 0.5. The difference in area is mainly due to the complexity of some large-scale mines with long mining duration, and some early mining faces have been weathered, resulting in the boundary location of the open-pit stope where it is located is not easy to define. The model has good accuracy in extracting stope information with obvious mining characteristics.
Fig. 8 The area similarity statistics in the study area only accounts for about 1.4% of the total area of the open-pit mine, and the misidentified patches are mainly the transit sites and bare mountains with similar characteristics of mine stopes. 3.2 Analysis of patch changes from 2019 to 2021 Weining Beishan is a key node area for ecological environment restoration and governance in Ningxia mining area. Based on the remote sensing monitoring base maps of mines in 2019 and 2021, this paper identifies and counts the restoration and management of open-pit mine stopes from 2019 to 2021. The changes in the open-pit mine stope include site leveling, development and construction, artificial regreening, continuous mining and no change, as shown in Figure 9.
Fig. 9 Statistics of stope changes in open-pit mines from 2019 to 2021Among the 125 open-pit mine stopes identified in this paper, 43.2% of the patches have been ecologically restored, 20.8% of the patches can be observed with continuous mining activities, and 36% of the patches have no change. Among them, 44 have been filled with pits and leveled with soil covering, 6 have been redeveloped and utilized, and infrastructure such as photovoltaic panels, roads, and factories have been constructed, and 4 have been observed to be artificially regreened. Some typical cases of ecological restoration are shown in Figure 10.
Fig. 10 Typical case of mine ecological restoration and management in the study area4 ConclusionIn this paper, a stope information extraction method for open-pit mines is constructed in the Beishan area of Weining, Ningxia. In this method, the U-Net model based on pixel classification is used to avoid the tedious construction of artificial feature engineering, locate the spatial location of the open-pit mine stope, and preliminarily extract the stope range of the open-pit mine. It is verified that the accuracy of the method in identifying the spatial location of the stope in the open-pit mine reaches 0.71, and the average spatial range recognition accuracy is 0.78. The experimental results prove the feasibility of the proposed method in the extraction of stope information in open-pit mines, and has the prospect of engineering application. The open-pit mine in the Beishan area of Weining is small in scale and scattered in layout, and the open-pit mine stope is easy to be confused with the surface cover type such as bare mountain and transit site, and there will be a certain error in manual labeling samples, and a large number of high-precision samples are required for more accurate extraction. The follow-up work will continue to improve the dataset to further improve the accuracy and efficiency of the open-pit mine stope information extraction method and verify the robustness of the method. About author:LIU Li (1986—), female, engineer, her main research direction is photogrammetry and remote sensing, mine geological environment monitoring. E-mail: [email protected] Corresponding author: Li Shiyao. E-mail: [email protected]
First instance: Yang Ruifang review: Song Qifan
Final Judge: Jin Jun
information