A mine image recognition method and device, a server and a storage medium

Abstract

The present application provides a kind of mining area image recognition method, device, server and storage medium, by obtaining the optical satellite remote sensing image to be identified;Using the mining area image recognition model that is constructed in advance, the mining area recognition result of the optical satellite remote sensing image to be identified is obtained;In the process of model training, for the shallow feature that is not easy to be directly learned and summarized by depth learning model, the related shallow feature is extracted using the preset feature extraction algorithm, training data is generated, which helps depth learning model to inherit empirical rules, so as to constrain the misclassification range of depth learning model training and identification, and then improve the recognition precision of mining area recognition.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to a method, apparatus, server, and storage medium for image recognition in mining areas. Background Technology

[0002] Mineral resources are an important component of natural resources, and a sound order in mining activities is a crucial prerequisite for ensuring safe mining production, maintaining fairness in the mining market, and steadily advancing ecological civilization construction. Remote sensing technology, as a vital technical means for natural resource supervision, is characterized by its wide coverage, non-contact nature, and high timeliness. It can quickly detect and identify the status of surface mineral resource extraction, serving as a keen "sky eye" for discerning abnormalities in mining activities.

[0003] Currently, the identification of mining patches in remote sensing images is mainly carried out using the traditional method of manual interpretation. This relies on a large number of experienced technicians to extract information such as the mining area and mining conditions from each image. By comparing relevant data, clues of abnormal mining activities are identified. The operation process is not only time-consuming and labor-intensive, but the final results are also easily affected by human subjective consciousness.

[0004] To achieve automatic classification and recognition of remote sensing image patches, scholars both domestically and internationally have conducted extensive research. Early studies commonly used methods such as backpropagation neural networks and support vector machines (SVMs). However, these classification methods are difficult to apply in practical work, mainly because their accuracy cannot meet production requirements, thus failing to replace manual interpretation. In recent years, deep learning technology has achieved remarkable results in computer vision, speech recognition, and information retrieval, exhibiting high accuracy and timeliness. It provides powerful technical means for multi-source data processing, feature mining, information extraction, and prediction simulation, and also offers new ideas for remote sensing image patch classification and recognition.

[0005] Based on deep learning models, Chen et al. applied stacked autoencoder networks to hyperspectral image classification, Luus et al. used deep convolutional neural networks for land use classification in remote sensing images, Li et al. proposed a semantic segmentation model, DeepUNet, to achieve land-sea segmentation in remote sensing images, and Zhao et al. applied the instance segmentation model Mask R-CNN to building extraction from remote sensing images. These deep learning methods have all achieved certain recognition results in research on land, ocean, and building applications.

[0006] However, mining patch classification and recognition is a relatively complex application scenario, which is more difficult than vegetation recognition, building recognition, and road segmentation. This is because the terrain features in mining areas are complex, with irregular geometric shapes, irregular spatial structures and spectral reflections, small differences in features between different mineral types, and a large range of mineralization scales. Extracting and summarizing their features is more difficult, thus making recognition more challenging and resulting in poor recognition accuracy of deep learning models. Summary of the Invention

[0007] The present invention provides a mining area image recognition method, device, server and storage medium, which mainly solves the technical problem that existing deep learning models have high learning difficulty and poor recognition accuracy in mining area image recognition.

[0008] To solve the above-mentioned technical problems, the present invention provides a method for mine area image recognition, comprising:

[0009] Acquire optical satellite remote sensing images to be identified;

[0010] Using a pre-built mining area image recognition model, the mining area recognition result of the optical satellite remote sensing image to be identified is obtained;

[0011] The mining area image recognition model was constructed in the following manner:

[0012] Acquire raw optical satellite remote sensing images, and preprocess the raw optical satellite remote sensing images to generate a first set of image patches;

[0013] For the first image block in the first image block set, a preset feature extraction algorithm is used to extract the texture features, spectral features, and exponential features of the first image block; and based on the texture features, spectral features, and exponential features of the first image block, corresponding texture feature maps, spectral feature maps, and exponential feature maps are generated.

[0014] Obtain the mineral type labeling information corresponding to the first image block;

[0015] The texture feature map, spectral feature map, exponential feature map, and mineral type labeling information corresponding to the first image block are input into a deep learning model for model training, and a mining area image recognition model is obtained through training.

[0016] The step of preprocessing the original optical satellite remote sensing image to generate the first image patch set includes:

[0017] Based on the outline of the mining area, delineate the mining area patches;

[0018] Obtain the labeled mineral type corresponding to the circled mining area patches;

[0019] Based on the selected mining area patches, a second set of image blocks is generated;

[0020] The second image block in the second image block set is cut into segments according to a set image size, and the first image block set is generated based on the segmented image blocks.

[0021] Optionally, the preset feature extraction algorithm includes a texture feature extraction algorithm, a spectral feature extraction algorithm, and an exponential feature extraction algorithm; and the texture feature extraction algorithm is used to extract the texture features of the first image patch, the spectral feature extraction algorithm is used to extract the spectral features of the first image patch, and the exponential feature extraction algorithm is used to extract the exponential features of the first image patch.

[0022] Optionally, before delineating the mining area patches, the method further includes:

[0023] The original optical satellite remote sensing image is converted into a set spatial coordinate system, and the color depth is converted into a set color depth.

[0024] Optionally, the mining area identification results include the mining area outline and the type of mineral.

[0025] Optionally, the mining area image recognition method further includes:

[0026] Obtain the image type and capture time of the optical satellite remote sensing image to be identified;

[0027] The mining area image corresponding to the mining area identification result of the optical satellite remote sensing image to be identified is converted into a vector layer, and the coordinates of the center point within the outline range of each mining patch are calculated as the coordinate information of the corresponding mining patch.

[0028] The vector map layer is overlaid with the administrative division to obtain the detailed address of the corresponding mining patch;

[0029] The vector layer is overlaid with the preset mining control zone layer to obtain the land occupation information corresponding to each mining patch.

[0030] The vector layer is overlaid with a preset mining rights layer to determine the mining behavior type corresponding to each mining patch.

[0031] The coordinate system of the mining area image corresponding to the mining area identification result of the optical satellite remote sensing image to be identified is converted into a projected coordinate system, and the area of ​​each mining patch is calculated to obtain the mining area.

[0032] The mineral type, coordinate information, detailed address, mining area, land occupation information, mining behavior type, image type, and shooting time corresponding to each mining patch are integrated to obtain the mining patch image and output as an information integration table.

[0033] The present invention also provides a mining area image recognition device, comprising:

[0034] The acquisition module is used to acquire optical satellite remote sensing images to be identified;

[0035] A mining area image recognition model is used to identify mining areas in the optical satellite remote sensing images to be identified and output the mining area identification results.

[0036] The model building module is used to build the mining area image recognition model in the following manner:

[0037] The process involves acquiring raw optical satellite remote sensing imagery and preprocessing the imagery to generate a first image patch set. This preprocessing includes: identifying mining area patches based on the mining area outline; obtaining the labeled mineral type corresponding to the identified mining area patches; generating a second image patch set based on the identified mining area patches; and cutting the corresponding second image patches in the second image patch set according to a set image size, and generating the first image patch set based on the cut image patches.

[0038] For the first image block in the first image block set, a preset feature extraction algorithm is used to extract the texture features, spectral features, and exponential features of the first image block; and based on the texture features, spectral features, and exponential features of the first image block, corresponding texture feature maps, spectral feature maps, and exponential feature maps are generated.

[0039] Obtain the mineral type labeling information corresponding to the first image block;

[0040] The texture feature map, spectral feature map, exponential feature map, and mineral type labeling information corresponding to the first image block are input into a deep learning model for model training, and a mining area image recognition model is obtained through training.

[0041] The present invention also provides a server, including a processor, a memory, and a communication bus;

[0042] The communication bus is used to enable communication between the processor and the memory;

[0043] The processor is used to execute one or more programs stored in the memory to implement the steps of the mining area image recognition method as described in any one of claims 1 to 7.

[0044] The present invention also provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the mining area image recognition method described above.

[0045] The beneficial effects of this invention are:

[0046] According to the present invention, a method, apparatus, server, and storage medium for identifying mining area images are provided. The method involves acquiring optical satellite remote sensing images to be identified; using a pre-constructed mining area image identification model, the mining area identification result of the optical satellite remote sensing images to be identified is obtained. The mining area image identification model is constructed as follows: acquiring original optical satellite remote sensing images; preprocessing the original optical satellite remote sensing images to generate a first image patch set; for the first image patch corresponding to the first image patch in the first image patch set, using a preset feature extraction algorithm to extract the texture features, spectral features, and exponential features of the first image patch; and generating corresponding texture feature maps, spectral feature maps, and exponential feature maps based on the texture features, spectral features, and exponential features of the first image patch; acquiring the mineral type labeling information corresponding to the first image patch; and inputting the texture feature map, spectral feature map, exponential feature map, and mineral type labeling information corresponding to the first image patch into a deep learning model for model training, thereby obtaining the mining area image identification model through training. During model training, for shallow features that are not easily learned and summarized by deep learning models, a preset feature extraction algorithm is used to extract relevant shallow features and generate training data. This helps deep learning models inherit empirical rules, thereby constraining the misclassification range of deep learning model training and recognition, and thus improving the accuracy of mining area recognition. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the mining area image recognition model construction method according to Embodiment 1 of the present invention;

[0048] Figure 2 This is a schematic diagram of the preprocessing method according to Embodiment 1 of the present invention;

[0049] Figure 3 This is a schematic diagram of the mining area image recognition method according to Embodiment 2 of the present invention;

[0050] Figure 4 This is a schematic diagram of the mining area image recognition model construction and image recognition method according to Embodiment 3 of the present invention;

[0051] Figure 5 This is a schematic diagram of a portion of the original sample from Embodiment 3 of the present invention;

[0052] Figure 6 This is a schematic diagram of the cut sample according to Embodiment 3 of the present invention;

[0053] Figure 7 This is a schematic diagram of feature fusion in Embodiment 3 of the present invention;

[0054] Figure 8 This is a schematic diagram of the mining area image recognition method according to Embodiment 3 of the present invention;

[0055] Figure 9This is a schematic diagram of the image cutting method according to Embodiment 3 of the present invention;

[0056] Figure 10 This is a schematic diagram of the prediction and annotation results in Embodiment 3 of the present invention;

[0057] Figure 11 This is a schematic diagram of the whole image prediction and annotation results in Embodiment 3 of the present invention;

[0058] Figure 12 This is a schematic diagram of the integrated table of image patch prediction and annotation results in Embodiment 3 of the present invention;

[0059] Figure 13 This is a schematic diagram of the structure of the mining area image recognition model construction device according to Embodiment 4 of the present invention;

[0060] Figure 14 This is a schematic diagram of the mining area image recognition device according to Embodiment 5 of the present invention;

[0061] Figure 15 This is a schematic diagram of the server structure according to Embodiment Six of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Example 1

[0063] To address the challenges of high learning difficulty and poor recognition accuracy in existing deep learning models for identifying mining area patches, this embodiment provides a method for constructing a mining area image recognition model. This method acquires raw optical satellite remote sensing images, preprocesses them, and then uses a preset feature extraction algorithm to extract relevant shallow features, generating training data to train a deep learning model for higher accuracy. This approach avoids directly using the deep learning model to learn shallow image features, thus preventing the difficulty in learning and summarizing complex mining area patch attributes (e.g., complex landform composition, inconsistent spatial structure and spectral representation, and extremely large scale), which leads to poor model accuracy. This method also helps the deep learning model inherit empirical rules, thereby constraining the misclassification range during training and recognition, and ultimately improving the accuracy of mining area identification.

[0064] Please see Figure 1 The mining area image recognition model construction method provided in this embodiment mainly includes the following steps:

[0065] S101. Acquire raw optical satellite remote sensing images;

[0066] Optical satellite remote sensing images are remote sensing images visible to the human eye, mainly including visible light and near-infrared bands, which can be obtained by the Gaofen-1 satellite. This embodiment does not limit this.

[0067] S102. Preprocess the original optical satellite remote sensing images to generate the first image patch set;

[0068] The first image patch set contains several image patches, which can be obtained by segmenting several original optical satellite remote sensing images. It should be understood that a single original optical satellite remote sensing image contains a large amount of data, potentially including numerous mining area patches. Segmenting the original optical satellite remote sensing image yields image patches with a larger data volume, which helps increase the amount of training sample data for the model, thereby improving training effectiveness.

[0069] In raw optical satellite remote sensing imagery, mining area patches constitute a relatively small proportion. If the raw imagery is directly segmented, a large number of invalid image patches without mining area patches exist. Too many invalid patches negatively impact image model training efficiency and, due to uneven sample distribution (a large number of invalid samples), may also affect the model's recognition accuracy. Therefore, in other optional embodiments of this invention, existing mining area patches can be first identified from the raw optical satellite remote sensing imagery. Then, training data can be generated based on these mining area patches for model training, improving both training efficiency and accuracy.

[0070] For details, please see Figure 2 As shown, the main preprocessing methods adopted are as follows:

[0071] S201. Based on the outline of the mining area, delineate the mining area patches;

[0072] The delineation of mining area patches can be done manually (the human eye can easily identify mining area patches in optical satellite remote sensing imagery). Specifically, closed curves can be used to enclose the mining area patches, such as rectangles or circles. It should be understood that the delineated area should at least encompass the mining area patches, but the delineated areas of two mining area patches should avoid overlap as much as possible. That is, the delineated area should not be too large, ensuring it remains within a reasonable range. For example, the delineated area can be 100 pixels outside the mining area outline. The specific outer range can be flexibly set.

[0073] Of course, in some embodiments of the present invention, an automatic delineation method can also be adopted to reduce the workload of researchers, since the number of mining area patches in the original optical satellite remote sensing image is quite large. For example, the contours of mining area patches can be extracted using existing arbitrary contour recognition algorithms, and then the corresponding contours of mining area patches can be delineated using a minimum bounding box. Furthermore, the minimum bounding box can be appropriately extended to delineate the mining area patches.

[0074] S202. Obtain the labeled mineral type corresponding to the circled mining area patches;

[0075] In this embodiment, the mineral type can be labeled manually.

[0076] Among them, the types of minerals include, but are not limited to, limestone, sandstone, shale and other types.

[0077] S203. Based on the selected mining area patches, generate a second set of image blocks;

[0078] Compared to direct splitting, this method greatly reduces the amount of data processing and improves model training efficiency.

[0079] S204. Cut the corresponding second image block in the second image block set according to the set image size, and generate the first image block set based on the cut image blocks.

[0080] The second image block is further segmented to generate several first image blocks. The main purpose is to better adapt to model requirements, such as better matching the hardware performance of the model training device, thereby improving training efficiency and recognition accuracy. Specifically, the second image block can be segmented by setting a sliding window size and a step size. The sliding window size and step size can be flexibly set without limitation. For example, a 256×256 pixel sliding window and a step size of 248 pixels can be used to segment the second image block.

[0081] In the preprocessing of raw optical satellite remote sensing images, before delineating mining area patches, raw optical satellite remote sensing images from different sources can be acquired to increase the amount of training data, which is beneficial for better model training. However, these images may differ in certain attributes, such as different spatial coordinate systems and color depths, which will affect the quality of the training data. Therefore, in other optional embodiments of this invention, all raw optical satellite remote sensing images are converted to a set spatial coordinate system and color depth to a set color depth, ensuring the uniformity of the spatial coordinate system and color depth of the raw optical satellite images. Specifically, the spatial coordinate system can be converted to the National 2000 coordinate system under Gaussian projection with 3-degree and 36-degree zones, and the color depth is uniformly converted to 8 bits. It should be understood that the specific spatial coordinate system and color depth can be flexibly set based on the actual situation. There are no restrictions on the specific spatial coordinate system and color depth used; the main purpose is to unify the coordinate system and color depth.

[0082] S103. For the first image block corresponding to the first image block in the first image block set, a preset feature extraction algorithm is used to extract the texture features, spectral features and exponential features of the first image block;

[0083] These features are not easily learned and summarized directly by deep learning models. Before training, relevant features are extracted using a pre-defined feature extraction algorithm and used as training input data. This helps deep learning models inherit empirical rules, thereby constraining the misclassification range of deep learning model training and recognition, and further improving recognition accuracy.

[0084] In this embodiment, the preset feature extraction algorithm includes a texture feature extraction algorithm, a spectral feature extraction algorithm, and an exponential feature extraction algorithm; and the texture feature extraction algorithm is used to extract the texture features of the first image block, the spectral feature extraction algorithm is used to extract the spectral features of the first image block, and the exponential feature extraction algorithm is used to extract the exponential features of the first image block.

[0085] The process of extracting texture features from the first image patch using a texture feature extraction algorithm includes:

[0086] A gray-level co-occurrence matrix transformation is performed on the first image block to obtain four texture features of the first image block: second moment, contrast, entropy, and inverse difference matrix.

[0087] Optionally, for each pixel of the first image block, a window of the gray-level co-occurrence matrix corresponding to the pixel is determined with the pixel itself as the center, and four texture features of the pixel are calculated based on the pixel values ​​within the window: the second moment, contrast, entropy, and inverse difference matrix.

[0088] Specifically, the window for determining the gray-level co-occurrence matrix corresponding to a pixel can be determined according to a set pixel range. For pixels located at the image boundary, a mirror filling method is used to make the pixels located at the image boundary have a window of a size corresponding to the set pixel range.

[0089] The four texture features can be calculated in the following ways:

[0090] The second moment ASM of a pixel is calculated using the following formula:

[0091] ,

[0092] In the formula, ASM represents the second moment, m represents the window pixel length, and n represents the window pixel width. This represents the pixel value at pixel coordinates (i,j);

[0093] The contrast ratio CON of a pixel is calculated using the following formula:

[0094] ,

[0095] In the formula, CON represents contrast.

[0096] The entropy ENT of a pixel is calculated using the following formula:

[0097] ,

[0098] In the formula, ENT represents entropy;

[0099] The inverse difference matrix (IDM) of a pixel is calculated using the following formula:

[0100] ,

[0101] In the formula, IDM represents the inverse difference matrix.

[0102] The extraction of spectral features from the first image patch using a spectral feature extraction algorithm includes:

[0103] The first image block is split into its original bands to obtain four spectral features: red band (R), green band (G), blue band (B), and near-infrared band (NIR).

[0104] The process of extracting the exponential features of the first image patch using the exponential feature extraction algorithm includes:

[0105] Remote sensing index transformation was performed using the spectral features of the first image patch to obtain four index features of the first image patch: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Ratio Vegetation Index (RVI), and Difference Environmental Vegetation Index (DVI).

[0106] Specifically, the Normalized Difference Vegetation Index (NDVI) of the first image patch is calculated using the following formula:

[0107] ,

[0108] The Normalized Difference Water Index (NDWI) for the first image patch is calculated using the following formula:

[0109] ,

[0110] The Ratio Vegetation Index (RVI) for the first image patch is calculated using the following formula:

[0111] ,

[0112] The difference in environmental vegetation index (DVI) for the first image patch is calculated using the following formula:

[0113] DVI = NIR-R

[0114] S104. Based on the texture features, spectral features, and exponential features of the first image patch, generate the corresponding texture feature map, spectral feature map, and exponential feature map;

[0115] In this embodiment, the texture feature map includes a second-order moment texture feature map, a contrast texture feature map, an entropy texture feature map, and an inverse difference matrix texture feature map.

[0116] Generating the corresponding texture feature map includes:

[0117] After obtaining the four texture features of all pixels in the first image block: second moment, contrast, entropy, and inverse difference matrix, a second moment texture feature map of the same size as the first image block is generated based on the second moment texture features of all pixels in the first image block.

[0118] Based on the contrast texture features of all pixels in the first image block, a contrast texture feature map of the same size as the first image block is generated.

[0119] Based on the entropy texture features of all pixels in the first image block, generate an entropy texture feature map with the same size as the first image block;

[0120] Based on the inverse difference matrix texture features of all pixels in the first image block, an inverse difference matrix texture feature map of the same size as the first image block is generated.

[0121] In this embodiment, the spectral feature map includes the red light band R spectral feature map, the green light band G spectral feature map, the blue light band B spectral feature map, and the near-infrared light band NIR spectral feature map.

[0122] The generation of the corresponding spectral feature map includes:

[0123] After obtaining the four spectral features of the red band R, green band G, blue band B, and near-infrared band NIR for all pixels of the first image block, a red band R spectral feature map of the same size as the first image block is generated based on the red band R of all pixels of the first image block.

[0124] Based on the green light band G of all pixels in the first image block, generate a spectral feature map of green light band G with the same size as the first image block;

[0125] Based on the blue light band B of all pixels of the first image block, generate a blue light band B spectral feature map with the same size as the first image block.

[0126] Based on the near-infrared (NIR) bands of all pixels in the first image block, a near-infrared NIR spectral feature map with the same size as the first image block is generated.

[0127] In this embodiment, the index feature map includes the Normalized Difference Vegetation Index (NDVI) feature map, the Normalized Difference Water Index (NDWI) feature map, the Ratio Vegetation Index (RVI) feature map, and the Difference Environmental Vegetation Index (DVI) feature map.

[0128] The generation of the corresponding exponential feature map includes:

[0129] After obtaining the four index features of all pixels in the first image block: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Ratio Vegetation Index (RVI), and Difference Environmental Vegetation Index (DVI), a Normalized Difference Vegetation Index (NDVI) feature map with the same size as the first image block is generated based on the Normalized Difference Vegetation Index (NDVI) of all pixels in the first image block.

[0130] Based on the Normalized Water Index (NDWI) of all pixels in the first image patch, generate a Normalized Water Index (NDWI) feature map with the same size as the first image patch.

[0131] Based on the ratio vegetation index (RVI) of all pixels in the first image patch, generate a ratio vegetation index (RVI) feature map with the same size as the first image patch.

[0132] Based on the difference environmental vegetation index (DVI) of all pixels in the first image patch, a difference environmental vegetation index (DVI) feature map with the same size as the first image patch is generated.

[0133] S105. Obtain the mineral type labeling information corresponding to the first image block;

[0134] Among them, the types of minerals include, but are not limited to, limestone, sandstone, shale and other types.

[0135] S106. Input the texture feature map, spectral feature map, exponential feature map and mineral type labeling information corresponding to the first image block into the deep learning model for model training, and obtain the mining area image recognition model through training.

[0136] It should be understood that the deep learning model used can be flexibly selected according to the actual situation, and this embodiment does not impose any restrictions on it. In addition, the process of training the deep learning model based on the training data can adopt any existing training method, which will not be described in detail here.

[0137] The mining area image recognition model construction method provided by this invention extracts relevant shallow features from shallow features that are not easily learned and summarized by deep learning models during the model training process, and generates training data. This helps deep learning models inherit empirical rules, thereby constraining the misclassification range of deep learning model training and recognition, and thus improving the accuracy of mining area recognition. Example 2

[0138] This implementation, based on the mining area image recognition model obtained through training in Example 1, provides a method for implementing mining area image recognition using this model. Please refer to [link to relevant documentation]. Figure 3 The image recognition methods for this mining area mainly include:

[0139] S301. Acquire the optical satellite remote sensing image to be identified;

[0140] S301. Using the mining area image recognition model, obtain the mining area recognition results of the optical satellite remote sensing image to be identified.

[0141] The mining area image recognition model trained based on Example 1 can effectively improve the accuracy of mining area recognition.

[0142] Before inputting the optical satellite remote sensing image to be identified into the mining area image recognition model, it may be necessary to preprocess it to better meet the model's prediction requirements. Therefore, in some optional embodiments of the present invention, the optical satellite remote sensing image to be identified may be preprocessed before being input into the mining area image recognition model.

[0143] Optionally, the optical satellite remote sensing image to be identified can be segmented according to a set image size to obtain image blocks to be identified;

[0144] The texture features, spectral features, and exponential features of the image patch to be identified are extracted using a preset feature extraction algorithm; and based on the texture features, spectral features, and exponential features of the image patch to be identified, corresponding texture feature maps, spectral feature maps, and exponential feature maps are generated.

[0145] The texture feature map, spectral feature map, and exponential feature map corresponding to the image block to be identified are used as inputs to the mining area image recognition model, so as to obtain the mining area recognition result of the image block to be identified using the mining area image recognition model.

[0146] Since the optical satellite remote sensing image to be identified is divided into several image blocks, the mining area identification result of the image can be obtained by integrating the mining area identification results of each image block. Specifically, the mining area identification results of each image block (a small image carrying the mining area outline and mineral type information) can be stitched together according to the cutting relationship to obtain the mining area identification result of the entire optical satellite remote sensing image (i.e., how many mining area patches exist, the distribution, outline, and corresponding mineral type information of each mining area patch, etc.).

[0147] In some optional embodiments of the present invention, in order to better display the mining area identification results and facilitate browsing by more people, after obtaining the mining area identification results of each image block to be identified corresponding to the optical satellite remote sensing image to be identified, the mining area identification results of each image block to be identified can be processed and displayed in the form of an information table.

[0148] Optionally, obtain the image type and capture time of the optical satellite remote sensing image to be identified;

[0149] The mining area image corresponding to the mining area identification result of the optical satellite remote sensing image to be identified is converted into a vector layer, and the coordinates of the center point within the outline range of each mining patch are calculated as the coordinate information of the corresponding mining patch.

[0150] By overlaying the vector map layer with the administrative division, the detailed address of the corresponding mining patch is obtained; taking Chongqing as an example, the detailed address is represented in the form of "Chongqing City - a certain district (county) - a certain township (street) - a certain village (community)".

[0151] Overlay the vector layer with the preset mining control zone layer to obtain the land occupation information corresponding to each mining patch; that is, whether it encroaches on the key control zone and the name of the encroached control zone. The key control zone includes permanent basic farmland, nature reserves, ecological protection red lines, national / local public welfare forests, and a 1km radius on both sides of highways / railways.

[0152] The vector layer is overlaid with a preset mining rights layer to determine the mining activity type corresponding to each mining patch; this includes, but is not limited to, legal mining, unlicensed mining, and mining beyond the designated boundaries. The mining rights scope is determined using the National Mining Rights Holders Exploration and Mining Information Disclosure System. The mining activity is determined as follows: if a mining patch is completely outside the mining rights scope, it is considered unlicensed mining; if a mining patch is partially outside the mining rights scope, it is considered mining beyond the designated boundaries; if a mining patch is completely within the mining rights scope, it is considered legal mining.

[0153] The coordinate system of the mining area image corresponding to the mining area identification result of the optical satellite remote sensing image to be identified is converted into a projected coordinate system, and the area of ​​each mining patch is calculated to obtain the mining area.

[0154] The information for each mining patch, including mineral type, coordinates, detailed address, mining area, land occupation, mining activity type, image type, and capture time, is integrated and the mining patch image is obtained, outputting an integrated information table. The surrounding image of the mining patch can be captured by taking a screenshot using a rectangular window that extends the patch by a specified number of pixels. Example 3

[0155] This embodiment, based on Embodiment 1 and / or Embodiment 2 above, provides a method for mine area image recognition, mainly comprising two parts: training a mine area image recognition model and performing mine area image recognition using the trained model. Please refer to [link to relevant documentation]. Figure 4 Mainly includes:

[0156] 1. Sample preparation

[0157] 1-1 Sample Collection

[0158] Optical satellite remote sensing imagery was acquired, and by outlining the contours of the mining patches and labeling the mineral types, a large number of original mining patch samples were obtained, with a sample size exceeding 5000. See also... Figure 5 The image shows a portion of the samples.

[0159] The optical satellite remote sensing images can be obtained through the Gaofen-1 satellite, with a resolution of 2m. They are all composite images after correction, registration and fusion, containing four bands: red (R), green (G), blue (B) and near-infrared (Nir). The coordinate system is uniformly converted to the Gaussian projection of the National 2000 coordinate system with 3 degrees and 36 degrees.

[0160] The optical satellite remote sensing images consist of approximately 96 images successfully captured by the Gaofen-1 satellite within the Chongqing municipality from 2018 to 2020.

[0161] The mineral types are classified into four categories: limestone, sandstone, shale, and others.

[0162] 1-2 Creating a dataset

[0163] Due to the varying sizes of the mining areas, each original sample set was cut to a specific size to obtain a new sample dataset of uniform dimensions. (See [link]). Figure 6 As shown, the new sample data volume is larger than the original sample volume.

[0164] The cutting method involves cutting the original sample with the pixels within the mining area as the center, using a sliding window of 256×256 pixels and a step size of 248 pixels.

[0165] 2. Feature Extraction and Fusion

[0166] 2-1 Feature Extraction

[0167] 2-1-1 Extracting Spectral Features

[0168] The original bands of each new sample image are split to obtain four spectral features: R, G, B, and Nir.

[0169] 2-1-2 Extracting Texture Features

[0170] For each new sample image, a gray-level co-occurrence matrix transformation is performed to obtain four texture features: Angular Second Moment (ASM), Contrast (CON), Entropy (ENT), and Inverse Difference Matrix (IDM).

[0171] ASM is a measure of the uniformity of gray-level distribution and the coarseness of texture in an image. It reflects the uniformity of gray-level distribution and the coarseness of texture. When the image texture is uniform and regular, the energy value is larger; conversely, when the element values ​​of the gray-level co-occurrence matrix are similar, the energy value is smaller.

[0172] CON reflects the sharpness of an image and the depth of its texture. The sharper the texture, the greater the contrast.

[0173] Entropy (ENT) measures the randomness of the information contained in an image, representing the complexity of the image. Entropy is maximized when all values ​​in the co-occurrence matrix are equal, or when pixel values ​​exhibit maximum randomness.

[0174] IDM, also known as homogeneity, reflects the clarity and regularity of textures. Textures that are clear, regular, and easy to describe have larger values.

[0175] The Gray-Level Co-occurrence Matrix Transform (GLCM) uses a 5×5 pixel neighborhood as the window size for each pixel's GLCM. The ASM, CON, ENT, and IDM values ​​of the window patch are calculated as the four texture feature values ​​corresponding to that point. For pixels located at image boundaries, the image boundaries are first mirrored to ensure that pixels near the boundaries are divided into effective patches within the 5×5 neighborhood window. By traversing all pixels in the original image, a texture feature map with the same size as the original image is obtained.

[0176] 2-1-3 Extracting Index Features

[0177] For each new sample image, a remote sensing index transformation was performed to obtain four index features: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Ratio Vegetation Index (RVI), and Difference Environmental Vegetation Index (DVI).

[0178] Since vegetation and water bodies are the most common land features near mining areas, while houses are less common, the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Ratio Vegetation Index (RVI), and Difference Environmental Vegetation Index (DVI) were selected as quantitative indicators of index characteristics.

[0179] NDVI uses a combination of red (R) and near-infrared (Nir) bands for calculation and is one of the important parameters reflecting vegetation growth and nutrient information.

[0180] NDWI extracts water information from remote sensing images by performing difference processing on the green band (G) and near-infrared band (Nir).

[0181] The RVI ratio, or vegetation index, is a sensitive indicator parameter for green plants. It is highly correlated with LAI, leaf dry biomass (DM), and chlorophyll content, and can be used to detect and estimate plant biomass.

[0182] Similar to NDVI and the Normalized Difference Vegetation Index, DVI is also used to reflect vegetation cover.

[0183] 2-2 Feature Fusion

[0184] 2-2-1 Normalization treatment

[0185] To eliminate the dimensional differences among the 12 types of shallow features extracted above (four texture features: second moment, contrast, entropy, and inverse difference matrix; four spectral features: red band R, green band G, blue band B, and near-infrared band NIR; and four index features: normalized vegetation index NDVI, normalized water index NDWI, ratio vegetation index RVI, and difference environmental vegetation index DVI), they are normalized to ensure comparability among the features.

[0186] As a preferred method, the feature vector is linearly transformed using the linear normalization method, mapping it to the same gray value range [0, 255] as the original image.

[0187] 2-2-2 Feature Fusion

[0188] After normalization, a fused data set of 12 feature vectors is formed, which is used as the input to the subsequent deep neural network (see [link]). Figure 7 (As shown).

[0189] 3. Model Training

[0190] The fused data of 12-dimensional feature vectors and the corresponding mineral type labels are input into a deep learning model for training, resulting in a trained model. This trained model will then be used for subsequent classification and recognition.

[0191] The deep learning model can be DeepLabV3+.

[0192] 4. Mining Plot Classification and Recognition (see...) Figure 8 (As shown)

[0193] 4-1 Image Preprocessing

[0194] The input images to be tested are preprocessed by standardizing the spatial coordinate system and color depth, and then cutting them into multiple uniformly sized image blocks for identification. The coordinate system is uniformly converted to the National 2000 coordinate system with a Gaussian projection of 3 degrees and 36 degrees, and the color depth is uniformly 8 bits. The cutting method involves using pixels within the mining area as the center, employing a 256×256 pixel sliding window, and a step size of 248 pixels to cut the original samples, i.e., using an effective area overlapping and covering method for cutting (see [link to relevant documentation]). Figure 9 (As shown).

[0195] 4-2 Shallow Feature Extraction

[0196] The shallow features of the image patch to be identified are extracted in a 2-1 manner, and the features are fused in a 2-2 manner.

[0197] 4-3 Mining Plot Prediction

[0198] 4-3-1 Input Model Recognition

[0199] The fused features of all image patches to be identified are input into (3. Model Training) the trained deep learning model for identification, obtaining the predicted annotations of mining patches (see...). Figure 10 (As shown).

[0200] 4-3-2 Prediction Result Assembly

[0201] A stitching operation, the reverse of the 4-1 cutting process, is adopted to stitch together the predicted annotations of each image block to be identified into a complete image sheet with the same size as the original input image to be tested (see [link]). Figure 11 (As shown).

[0202] 4-4 Information Classification and Integration

[0203] The above prediction results include the outline shape and mineral type information of the mining patches. A series of spatial analyses are performed on the outline shape to obtain spatial classification information.

[0204] 4-4-1 Based on the original input image to be tested, read the image type and shooting time information.

[0205] The image types include, but are not limited to, images from the Gaofen-1 satellite. The capture time is listed in the format of "year and month".

[0206] 4-4-2 The prediction results are converted into a vector layer, and the coordinates of the center point of each mining patch's surface outline are calculated to obtain coordinate information. The center point coordinates can be calculated using the average coordinates of all pixels within the surface outline of the mining patch. The center point is then overlaid with the administrative division for spatial analysis to determine the address information of the center point, which serves as the detailed address of the mining patch.

[0207] Vector layers can use the shapfile file format.

[0208] Transform its coordinate system into a projected coordinate system, calculate the area of ​​each patch, and obtain the mining scale (area).

[0209] Coordinate information is expressed in latitude and longitude. The projected coordinate system adopts the Gauss projection under the National Geodetic Coordinate System 2000.

[0210] For example, the detailed address in Chongqing is expressed in the form of "Chongqing City - District (County) - Township (Street) - Village (Community) (Group)".

[0211] 4-4-3 The prediction results are overlaid with the key control areas (i.e., areas where mining is not allowed) for spatial analysis to obtain the land occupation information of mining plots, i.e. whether they encroach on the key control areas and the name of the control areas they encroach on.

[0212] The key control area includes permanent basic farmland, nature reserves, ecological protection red lines, national / local public welfare forests, and a 1km radius on both sides of highways / railways.

[0213] 4-4-4 The prediction results are overlaid with the mining rights area (i.e. the area where legal mining is permitted) for spatial analysis to obtain information on the types of mining activities, namely legal mining activities, unlicensed mining, and mining beyond the boundaries.

[0214] The scope of mining rights is based on data published by the National Mining Rights Holders Exploration and Mining Information Disclosure System.

[0215] The methods for determining mining activities are as follows:

[0216] If a mining plot is completely outside the mining rights area, it is determined that the plot is being mined without a license.

[0217] If a mining plot is located outside the mining rights area, the plot is deemed to be engaged in cross-border mining.

[0218] If a mining plot is entirely within the mining rights area, then mining on that plot is deemed legal.

[0219] 4-4-5 Information Integration

[0220] This paper integrates eight categories of information: mineral type, coordinates, detailed address, mining scale (area), land occupation information, mining activity type, image type, and shooting time. It also extracts images surrounding the mining area and outputs a single integrated information table. Please refer to [link to table]. Figure 12 As shown.

[0221] Among them, the image surrounding the mining patch can be captured by taking a rectangular window that expands the patch by 100 pixels. Example 4

[0222] This embodiment, based on Embodiment 1 and / or Embodiment 3, provides a mining area image recognition model construction device, mainly used to implement at least some or all of the steps of the mining area image recognition model construction method described in Embodiment 1 and / or Embodiment 3. Please refer to... Figure 13 The image recognition model construction device for this mining area mainly includes:

[0223] The preprocessing module 131 is used to acquire raw optical satellite remote sensing images and preprocess the raw optical satellite remote sensing images to generate a first image patch set.

[0224] Optical satellite remote sensing images are remote sensing images visible to the human eye, mainly including visible light and near-infrared bands, which can be obtained by the Gaofen-1 satellite. This embodiment does not limit this.

[0225] The first image patch set contains several image patches, which can be obtained by segmenting several original optical satellite remote sensing images using the preprocessing module 131. It should be understood that a single original optical satellite remote sensing image contains a large amount of data, potentially including multiple mining area patches. Segmenting the original optical satellite remote sensing image yields image patches with larger data volumes, which helps increase the amount of training sample data for the model, thereby improving training effectiveness.

[0226] In raw optical satellite remote sensing imagery, mining area patches constitute a relatively small proportion. If the raw imagery is directly segmented, a large number of invalid image patches without mining area patches exist. Too many invalid patches negatively impact image model training efficiency and, due to uneven sample distribution (a large number of invalid samples), may also affect the model's recognition accuracy. Therefore, in other optional embodiments of this invention, existing mining area patches can be first identified from the raw optical satellite remote sensing imagery. Then, training data can be generated based on these mining area patches for model training, improving both training efficiency and accuracy.

[0227] The preprocessing module 131 is used to delineate mining area patches based on the mining area outline; obtain the labeled mineral type corresponding to the delineated mining area patches; generate a second image block set based on the delineated mining area patches; cut the corresponding second image blocks in the second image block set according to a set image size; and generate a first image block set based on the cut image blocks.

[0228] The second image block is further segmented to generate several first image blocks. The main purpose is to better adapt to model requirements, such as better matching the hardware performance of the model training device, thereby improving training efficiency and recognition accuracy. Specifically, the second image block can be segmented by setting a sliding window size and a step size. The sliding window size and step size can be flexibly set without limitation. For example, a 256×256 pixel sliding window and a step size of 248 pixels can be used to segment the second image block.

[0229] In the preprocessing of raw optical satellite remote sensing images, before delineating mining area patches, raw optical satellite remote sensing images from different sources can be acquired to increase the amount of training data, which is beneficial for better model training. However, these images may differ in certain attributes, such as different spatial coordinate systems and color depths, which will affect the quality of the training data. Therefore, in other optional embodiments of the present invention, the preprocessing module 131 is also used to convert all raw optical satellite remote sensing images into a set spatial coordinate system and color depth, ensuring the uniformity of the spatial coordinate system and color depth of the raw optical satellite images. Specifically, the spatial coordinate system can be converted to the National 2000 coordinate system under Gaussian projection with 3-degree and 36-degree zones, and the color depth is uniformly converted to 8 bits. It should be understood that the specific spatial coordinate system and color depth can be flexibly set based on the actual situation. There are no restrictions on the specific spatial coordinate system and color depth used; the main purpose is to unify the coordinate system and color depth.

[0230] The feature extraction module 132 is used to extract the texture features, spectral features and exponential features of the first image block corresponding to the first image block in the first image block set using a preset feature extraction algorithm; and to generate the corresponding texture feature map, spectral feature map and exponential feature map based on the texture features, spectral features and exponential features of the first image block.

[0231] These features are not easily learned and summarized directly by deep learning models. Before training, the feature extraction module 132 extracts relevant features using a preset feature extraction algorithm as training input data. This helps the deep learning model inherit empirical rules, thereby constraining the misclassification range of the deep learning model during training and recognition, and further improving recognition accuracy.

[0232] In this embodiment, the preset feature extraction algorithm includes a texture feature extraction algorithm, a spectral feature extraction algorithm, and an exponential feature extraction algorithm; and the texture feature extraction algorithm is used to extract the texture features of the first image block, the spectral feature extraction algorithm is used to extract the spectral features of the first image block, and the exponential feature extraction algorithm is used to extract the exponential features of the first image block.

[0233] The process of extracting texture features from the first image patch using a texture feature extraction algorithm includes:

[0234] A gray-level co-occurrence matrix transformation is performed on the first image block to obtain four texture features of the first image block: second moment, contrast, entropy, and inverse difference matrix.

[0235] Optionally, for each pixel of the first image block, a window of the gray-level co-occurrence matrix corresponding to the pixel is determined with the pixel itself as the center, and four texture features of the pixel are calculated based on the pixel values ​​within the window: the second moment, contrast, entropy, and inverse difference matrix.

[0236] Specifically, the window for determining the gray-level co-occurrence matrix corresponding to a pixel can be determined according to a set pixel range. For pixels located at the image boundary, a mirror filling method is used to make the pixels located at the image boundary have a window of a size corresponding to the set pixel range.

[0237] The four texture features can be calculated in the following ways:

[0238] The second moment ASM of a pixel is calculated using the following formula:

[0239] ,

[0240] In the formula, ASM represents the second moment, m represents the window pixel length, and n represents the window pixel width. This represents the pixel value at pixel coordinates (i,j);

[0241] The contrast ratio CON of a pixel is calculated using the following formula:

[0242] ,

[0243] In the formula, CON represents contrast.

[0244] The entropy ENT of a pixel is calculated using the following formula:

[0245] ,

[0246] In the formula, ENT represents entropy;

[0247] The inverse difference matrix (IDM) of a pixel is calculated using the following formula:

[0248] ,

[0249] In the formula, IDM represents the inverse difference matrix.

[0250] The extraction of spectral features from the first image patch using a spectral feature extraction algorithm includes:

[0251] The first image block is split into its original bands to obtain four spectral features: red band (R), green band (G), blue band (B), and near-infrared band (NIR).

[0252] The process of extracting the exponential features of the first image patch using the exponential feature extraction algorithm includes:

[0253] Remote sensing index transformation was performed using the spectral features of the first image patch to obtain four index features of the first image patch: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Ratio Vegetation Index (RVI), and Difference Environmental Vegetation Index (DVI).

[0254] Specifically, the Normalized Difference Vegetation Index (NDVI) of the first image patch is calculated using the following formula:

[0255] ,

[0256] The Normalized Difference Water Index (NDWI) for the first image patch is calculated using the following formula:

[0257] ,

[0258] The Ratio Vegetation Index (RVI) for the first image patch is calculated using the following formula:

[0259] ,

[0260] The difference in environmental vegetation index (DVI) for the first image patch is calculated using the following formula:

[0261] DVI = NIR-R

[0262] The feature extraction module 132 is also used to generate corresponding texture feature maps, spectral feature maps and exponential feature maps based on the texture features, spectral features and exponential features of the first image patch;

[0263] In this embodiment, the texture feature map includes a second-order moment texture feature map, a contrast texture feature map, an entropy texture feature map, and an inverse difference matrix texture feature map.

[0264] The feature extraction module 132 is used to, after obtaining four texture features of all pixels of the first image block: second moment, contrast, entropy, and inverse difference matrix; generate a second moment texture feature map of the same size as the first image block based on the second moment texture features of all pixels of the first image block; generate a contrast texture feature map of the same size as the first image block based on the contrast texture features of all pixels of the first image block; generate an entropy texture feature map of the same size as the first image block based on the entropy texture features of all pixels of the first image block; and generate an inverse difference matrix texture feature map of the same size as the first image block based on the inverse difference matrix texture features of all pixels of the first image block.

[0265] In this embodiment, the spectral feature map includes the red light band R spectral feature map, the green light band G spectral feature map, the blue light band B spectral feature map, and the near-infrared light band NIR spectral feature map.

[0266] The feature extraction module 132 is used to, after obtaining the four spectral features of the red light band R, green light band G, blue light band B, and near-infrared light band NIR of all pixels of the first image block, generate a red light band R spectral feature map of the same size as the first image block based on the red light band R of all pixels of the first image block; generate a green light band G spectral feature map of the same size as the first image block based on the green light band G of all pixels of the first image block; generate a blue light band B spectral feature map of the same size as the first image block based on the blue light band B of all pixels of the first image block; and generate a near-infrared light band NIR spectral feature map of the same size as the first image block based on the near-infrared light band NIR of all pixels of the first image block.

[0267] In this embodiment, the index feature map includes the Normalized Difference Vegetation Index (NDVI) feature map, the Normalized Difference Water Index (NDWI) feature map, the Ratio Vegetation Index (RVI) feature map, and the Difference Environmental Vegetation Index (DVI) feature map.

[0268] The feature extraction module 132, after obtaining four index features—Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Ratio Vegetation Index (RVI), and Difference Environmental Vegetation Index (DVI)—for all pixels of the first image patch, generates a Normalized Difference Vegetation Index (NDVI) feature map of the same size as the first image patch based on the Normalized Difference Vegetation Index (NDVI) of all pixels of the first image patch; generates a Normalized Difference Water Index (NDWI) feature map of the same size as the first image patch based on the Normalized Difference Vegetation Index (NDWI) of all pixels of the first image patch; generates a Ratio Vegetation Index (RVI) feature map of the same size as the first image patch based on the Ratio Vegetation Index (RVI) of all pixels of the first image patch; and generates a Difference Environmental Vegetation Index (DVI) feature map of the same size as the first image patch based on the Difference Environmental Vegetation Index (DVI) of all pixels of the first image patch.

[0269] The first acquisition module 133 is used to acquire the mineral type labeling information corresponding to the first image block.

[0270] Among them, the types of minerals include, but are not limited to, limestone, sandstone, shale and other types.

[0271] The training module 134 is used to input the texture feature map, spectral feature map, exponential feature map and mineral type labeling information corresponding to the first image block into the deep learning model for model training, and obtain the mining area image recognition model through training.

[0272] It should be understood that the deep learning model used can be flexibly selected according to the actual situation, and this embodiment does not impose any restrictions on it. In addition, the process of training the deep learning model based on the training data can adopt any existing training method, which will not be described in detail here.

[0273] The mining area image recognition model construction device provided by this invention extracts relevant shallow features from shallow features that are not easily learned and summarized by deep learning models during the model training process, and generates training data. This helps deep learning models inherit empirical rules, thereby constraining the misclassification range of deep learning model training and recognition, and thus improving the accuracy of mining area recognition. Example 5

[0274] This embodiment, based on Embodiments 2 and / or 3 above, provides a mining area image recognition device, mainly used to implement some or all of the steps of the mining area image recognition model construction method described in Embodiments 2 and / or 3. Please refer to [link / reference]. Figure 14 The image recognition device in this mining area mainly includes:

[0275] The second acquisition module 141 is used to acquire the optical satellite remote sensing image to be identified;

[0276] Mining area image recognition model 142 is used to identify mining areas in the optical satellite remote sensing image to be identified and output the mining area recognition results.

[0277] Model building module 143 is used to build a mining area image recognition model in the following manner:

[0278] Acquire raw optical satellite remote sensing images and preprocess them to generate the first set of image patches;

[0279] For the first image block in the first image block set, a preset feature extraction algorithm is used to extract the texture features, spectral features, and exponential features of the first image block; and based on the texture features, spectral features, and exponential features of the first image block, corresponding texture feature maps, spectral feature maps, and exponential feature maps are generated.

[0280] Obtain the mineral type labeling information corresponding to the first image block;

[0281] The texture feature map, spectral feature map, exponential feature map, and mineral type labeling information corresponding to the first image patch are input into the deep learning model for model training, thereby obtaining the mining area image recognition model. For the specific process of constructing the mining area image recognition model, please refer to the description in Example 1, which will not be repeated here. Example 6

[0282] This embodiment, based on Embodiments 2 and / or 3 above, provides a server primarily used to implement at least some or all of the steps in the mining area image recognition model construction method described in Embodiments 2 and / or 3. Please refer to [link to documentation]. Figure 15 The mining server mainly includes:

[0283] Processor 151, memory 152 and communication bus 153;

[0284] Communication bus 153 is used to realize the connection and communication between processor 151 and memory 152;

[0285] The processor 151 executes one or more programs stored in the memory 152 to implement the steps of the mining area image recognition method as described in Embodiments 2 and / or 3. Please refer to the descriptions in Embodiments 2 and / or 3 for details, which will not be repeated here. Example 7

[0286] This embodiment, based on Embodiments 2 and / or 3 above, provides a computer-readable storage medium storing one or more programs, which can be executed by one or more processors to implement the steps of the mining area image recognition method as described in Embodiments 2 and / or 3. Please refer to the descriptions in Embodiments 2 and / or 3 for details, which will not be repeated here.

[0287] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a computer storage medium (ROM / RAM, magnetic disk, optical disk) for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, the present invention is not limited to any particular hardware and software combination.

[0288] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for image recognition in mining areas, characterized in that, include: Acquire optical satellite remote sensing images to be identified; Using a pre-built mining area image recognition model, the mining area recognition result of the optical satellite remote sensing image to be identified is obtained; The mining area image recognition model was constructed in the following manner: Acquire raw optical satellite remote sensing images, and preprocess the raw optical satellite remote sensing images to generate a first set of image patches; For the first image block in the first image block set, a preset feature extraction algorithm is used to extract the texture features, spectral features, and exponential features of the first image block; and based on the texture features, spectral features, and exponential features of the first image block, corresponding texture feature maps, spectral feature maps, and exponential feature maps are generated. Obtain the mineral type labeling information corresponding to the first image block; The texture feature map, spectral feature map, exponential feature map, and mineral type labeling information corresponding to the first image block are input into a deep learning model for model training, and a mining area image recognition model is obtained through training; The step of preprocessing the original optical satellite remote sensing image to generate the first image patch set includes: Based on the outline of the mining area, delineate the mining area patches; Obtain the labeled mineral type corresponding to the circled mining area patches; Based on the selected mining area patches, a second set of image blocks is generated; The second image block in the second image block set is cut into segments according to a set image size, and the first image block set is generated based on the segmented image blocks.

2. The mining area image recognition method as described in claim 1, characterized in that, The preset feature extraction algorithm includes a texture feature extraction algorithm, a spectral feature extraction algorithm, and an exponential feature extraction algorithm; and the texture feature extraction algorithm is used to extract the texture features of the first image patch, the spectral feature extraction algorithm is used to extract the spectral features of the first image patch, and the exponential feature extraction algorithm is used to extract the exponential features of the first image patch.

3. The mining area image recognition method as described in claim 1, characterized in that, Before delineating the mining area patches, the following is also included: The original optical satellite remote sensing image is converted into a set spatial coordinate system, and the color depth is converted into a set color depth.

4. The mining area image recognition method as described in claim 1, characterized in that, The mining area identification results include the mining area outline and the type of mineral.

5. The mining area image recognition method as described in claim 1, characterized in that, The mining area image recognition method also includes: Obtain the image type and capture time of the optical satellite remote sensing image to be identified; The mining area image corresponding to the mining area identification result of the optical satellite remote sensing image to be identified is converted into a vector layer, and the coordinates of the center point within the outline range of each mining patch are calculated as the coordinate information of the corresponding mining patch. The vector map layer is overlaid with the administrative division to obtain the detailed address of the corresponding mining patch; The vector layer is overlaid with the preset mining control zone layer to obtain the land occupation information corresponding to each mining patch. The vector layer is overlaid with a preset mining rights layer to determine the mining behavior type corresponding to each mining patch. The coordinate system of the mining area image corresponding to the mining area identification result of the optical satellite remote sensing image to be identified is converted into a projected coordinate system, and the area of ​​each mining patch is calculated to obtain the mining area. The mineral type, coordinate information, detailed address, mining area, land occupation information, mining behavior type, image type, and shooting time corresponding to each mining patch are integrated to obtain the mining patch image and output as an information integration table.

6. A mining area image recognition device, characterized in that, include: The acquisition module is used to acquire optical satellite remote sensing images to be identified; A mining area image recognition model is used to identify mining areas in the optical satellite remote sensing images to be identified and output the mining area identification results. The model building module is used to build the mining area image recognition model in the following manner: The process involves acquiring raw optical satellite remote sensing imagery and preprocessing the imagery to generate a first image patch set. This preprocessing includes: identifying mining area patches based on the mining area outline; obtaining the labeled mineral type corresponding to the identified mining area patches; generating a second image patch set based on the identified mining area patches; and cutting the corresponding second image patches in the second image patch set according to a set image size, and generating the first image patch set based on the cut image patches. For the first image block in the first image block set, a preset feature extraction algorithm is used to extract the texture features, spectral features, and exponential features of the first image block; and based on the texture features, spectral features, and exponential features of the first image block, corresponding texture feature maps, spectral feature maps, and exponential feature maps are generated. Obtain the mineral type labeling information corresponding to the first image block; The texture feature map, spectral feature map, exponential feature map, and mineral type labeling information corresponding to the first image block are input into a deep learning model for model training, and a mining area image recognition model is obtained through training.

7. A server, characterized in that, The server includes a processor, a memory, and a communication bus; The communication bus is used to enable communication between the processor and the memory; The processor is used to execute one or more programs stored in the memory to implement the steps of the mining area image recognition method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps of the mining area image recognition method as described in any one of claims 1 to 5.

Images