A mineral resource low-altitude monitoring method based on unmanned aerial vehicle vision

Abstract

This invention relates to the field of image processing technology, specifically to a method for low-altitude monitoring of mineral resources based on UAV vision. The method involves: obtaining feature change sensitivity based on the differences and variations in feature vectors of pixels in different receptive fields; obtaining a first weighting coefficient for pixels in different receptive fields based on the feature change sensitivity; obtaining a difference vector and principal components based on the differences in feature vectors of pixels within the same receptive field; obtaining directional consistency based on the directional distribution characteristics of the difference vector and principal components; obtaining stability based on the feature vectors of other pixels in the receptive field of a pixel; obtaining a second weighting coefficient for pixels in different receptive fields based on directional consistency and stability; and obtaining a fusion weight based on the first and second weighting coefficients. This invention obtains a weighted feature vector based on the fusion weight and feature vector; and obtains a mining area segmentation map based on the weighted feature vector, thus improving the accuracy of mineral resource monitoring.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for low-altitude monitoring of mineral resources based on UAV vision. Background Technology

[0002] Low-altitude remote sensing technology using unmanned aerial vehicles (UAVs) is widely used in mineral resource monitoring due to its advantages such as flexibility, low acquisition cost, and ability to acquire high-resolution image data. By planning low-altitude flight paths and acquiring multi-view images of the mining area, high-precision image data covering the target area is obtained. Semantic segmentation models such as DeepLabv3+ are used to segment the target mining area in the images at the pixel level to obtain the distribution results of the mining area. Further extraction of the mining area boundary information, combined with area calculation and spatial overlay analysis, enables quantitative monitoring of the mining area's mining scope and changes.

[0003] The segmentation results of the target mining area serve as the foundation for subsequent boundary extraction and area calculation, and their accuracy directly affects the accuracy of mining area range identification and change monitoring. Existing DeepLabv3+ models primarily rely on the Availability Spatial Pyramid Pooling (ASPP) module to acquire multi-scale features. Different porosity branches correspond to different receptive field ranges, and the features from each branch are stitched and fused to form multi-scale features at the same spatial location. However, in the current scenario, the mining area boundary region is composed of various landforms such as ore, bare soil, and vegetation, resulting in chaotic boundary textures. This leads to significant differences in the adaptability of different scale branches in the boundary region. Features extracted by some branches can well characterize the boundary structure, while others may introduce information from non-mining areas or only reflect localized incomplete features. Traditional methods, in the process of multi-scale feature fusion, fail to distinguish the effectiveness of different branch features in the same region, allowing branch features with low adaptability to participate in the fusion, which interferes with the final feature expression. This leads to the weakening or submergence of effective features by ineffective features, reducing the discrimination ability of boundary regions and causing problems such as blurring, offsetting or breaking of segmentation results. The accuracy and stability of boundary region segmentation results decrease, ultimately affecting the accuracy of mineral resource monitoring. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a low-altitude monitoring method for mineral resources based on UAV vision, and the specific technical solution adopted is as follows: Acquire monitoring images of the mining area; The feature vectors of pixels in receptive fields of different scales are obtained through ASPP; the feature change sensitivity is obtained based on the difference and change features of the feature vectors of the pixels in different receptive fields; and the first weight coefficient of the pixels in different receptive fields is obtained based on the feature change sensitivity. Based on the difference features of the feature vectors of the pixel and other pixels in the same receptive field, different difference vectors and principal components are obtained; based on the directional distribution features of the difference vectors and principal components, directional consistency is obtained; based on the difference features of the feature vectors of other pixels in the receptive field of the pixel, stability is obtained; based on the directional consistency and stability, a second weighting coefficient of the pixel in different receptive fields is obtained. The fusion weight of the pixel under different receptive fields is obtained based on the first weight coefficient and the second weight coefficient; the weighted feature vector of the pixel is obtained based on the fusion weight and the feature vector; and the mining area segmentation map is obtained through the semantic segmentation model based on the weighted feature vector.

[0005] Furthermore, the step of obtaining the feature change sensitivity based on the difference and change features of the feature vectors of the pixel in different receptive fields includes: In the formula, D represents the sensitivity of a pixel to feature changes. This indicates normalization, and N represents the number of receptive fields at different scales. This represents the feature vector of a pixel in the nth receptive field. This represents the average feature vector of a pixel across all receptive fields. This indicates the calculated modulus; the modulus is sorted in ascending order of receptive field scale. This indicates that the pixel is in the nth receptive field and the nth receptive field. The magnitude of the difference between the feature vectors of each receptive field. Indicates satisfaction The quantity.

[0006] Further, the step of obtaining the first weighting coefficient of the pixel in different receptive fields based on the feature change sensitivity includes: Sort the receptive fields in ascending order of scale, calculate the ratio of the position of any receptive field to the sum of the positions of all receptive fields, and obtain the position feature value of any receptive field; when the sensitivity of the feature change is not lower than a preset threshold, calculate the difference between a constant 1 and the position feature value, and obtain the first weight coefficient of the pixel in any receptive field; when the sensitivity of the feature change is lower than the preset threshold, use the position feature value as the first weight coefficient of the pixel in any receptive field.

[0007] Furthermore, the step of obtaining different difference vectors and principal components based on the difference features between the feature vectors of the pixel and other pixels under the same receptive field includes: Calculate the difference between the feature vector of the pixel and other pixels within the corresponding local receptive field, and calculate the unit vector of the difference result to obtain the difference vector between the pixel and other pixels; obtain the principal components of all difference vectors of the pixel under the same receptive field by principal component analysis.

[0008] Further, the step of obtaining the directional consistency degree based on the difference vector and the directional distribution characteristics of the principal components includes: In the formula, C represents the directional consistency of the pixel. The unit vector representing the principal component. R represents the unit vector representing the average difference vector, and R represents the number of other pixels within the receptive field of the pixel. This represents the difference vector between the pixel and the r-th other pixel.

[0009] Further, the step of obtaining stability based on the difference features of feature vectors of other pixels in the receptive field of the pixel includes: In the formula, Q represents the stability of the pixel. Indicates normalization. This represents the maximum value of the feature vector magnitude of other pixels in the receptive field. This represents the minimum feature vector magnitude of other pixels in the receptive field.

[0010] Further, the step of obtaining the second weighting coefficients of the pixel in different receptive fields based on the directional consistency and the stability includes: In the formula This represents the second weighting coefficient of the pixel in the nth receptive field. This indicates the directional consistency of the pixel in the nth receptive field. This represents the stability of the pixel in the nth receptive field.

[0011] Further, the step of obtaining the fusion weight of the pixel in different receptive fields based on the first weight coefficient and the second weight coefficient includes: In the formula This represents the fusion weight of the pixel in the nth receptive field, where N represents the number of receptive fields. This represents the first weight coefficient of the pixel in the nth receptive field. This represents the second weighting coefficient of the pixel in the nth receptive field.

[0012] Further, the step of obtaining the weighted feature vector of the pixel based on the fusion weight and the feature vector includes: In the formula This represents the weighted feature vector of the pixel, where N represents the number of receptive fields. This represents the fusion weight of the pixel in the nth receptive field. This represents the feature vector of the pixel in the nth receptive field.

[0013] The present invention has the following beneficial effects: In this invention, feature change sensitivity is obtained based on the differences and variations in feature vectors of pixels in different receptive fields. This reflects the adaptability and sensitivity of pixels to receptive fields of different scales. Furthermore, a first weighting coefficient for pixels in different receptive fields is obtained based on this feature change sensitivity, which preliminarily weights the feature vectors of pixels under different receptive fields, improving the accuracy of multi-scale feature fusion results. Directional consistency is obtained based on the directional distribution characteristics of the difference vectors and principal components, reflecting the stability and consistency of local spatial characterization under any receptive field. Stability is obtained based on the differences in feature vectors of other pixels within a pixel's receptive field, reflecting the uniformity and stability of feature expression under any receptive field. A second weighting coefficient for pixels in different receptive fields is obtained based on directional consistency and stability, allowing weighting based on the quality of feature extraction from pixels under different receptive fields, further improving the accuracy of multi-scale feature fusion results. Obtaining fusion weights comprehensively improves the accuracy of feature vector weighting under different receptive fields. Finally, a weighted feature vector of pixels is obtained, and a mining area segmentation map is generated based on this weighted feature vector, improving the accuracy of mining resource monitoring. Attached Figure Description

[0014] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a low-altitude monitoring method for mineral resources based on UAV vision, provided as an embodiment of the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a low-altitude monitoring method for mineral resources based on UAV vision proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] The following description, in conjunction with the accompanying drawings, details a specific scheme for a low-altitude monitoring method for mineral resources based on UAV vision provided by the present invention.

[0019] Please see Figure 1 The diagram illustrates a flowchart of a low-altitude monitoring method for mineral resources based on UAV vision, according to an embodiment of the present invention. The method includes the following steps: Step S1: Obtain monitoring images of the mining area.

[0020] Based on the mining area's scope and terrain features, low-altitude flight paths are planned for the target mining area. UAVs continuously collect impact data according to preset flight paths and intervals, acquiring multiple images covering the mining area. These images are then stitched together based on their positional relationships to generate a complete stitched image covering the mining area, unified to the same geographic coordinate system. Since the model has fixed input size requirements, the stitched images need to be segmented. A sliding window of a preset size (e.g., 512 pixels on each side) is used to crop the images, with a sliding step of half the window's side length, ensuring overlapping areas between images. An adaptive histogram equalization method is used to enhance image contrast and improve boundary discernibility. The cropped and preprocessed images are then used as different monitoring images.

[0021] Step S2: Obtain the feature vectors of pixels in receptive fields of different scales using ASPP; obtain the feature change sensitivity based on the difference and change features of the feature vectors of pixels in different receptive fields; obtain the first weight coefficient of pixels in different receptive fields based on the feature change sensitivity.

[0022] Because the features extracted from a pixel under different void ratios in the ASPP void space pyramid pooling module may have varying adaptability to that pixel, and this difference is not considered in the subsequent multi-scale feature fusion process, it can easily lead to the weakening of effective features or their submersion by ineffective features, reducing the ability to distinguish boundary regions in mining scenes. Therefore, to improve the accuracy of the multi-scale feature fusion process for pixels, it is necessary to analyze the features extracted under different void ratios in the receptive field. In mining areas, various land features such as ore, bare soil, and vegetation are spatially irregularly mixed, and the distribution of these features exhibits spatial heterogeneity. Receptive fields of different scales may focus on completely different local structural regions at the same pixel. For example, smaller receptive fields mainly focus on local details, such as the boundary texture changes between ore and bare soil; while larger receptive fields may introduce more background or neighborhood information, diluting the salience of boundary structural features to some extent. Therefore, in areas with complex terrain features, such as mixed areas of minerals, bare soil, and vegetation, large-scale receptive fields introduce additional interference information, affecting the feature representation of pixels in that area. Conversely, in areas with relatively homogeneous structures, smaller-scale receptive fields struggle to fully capture the semantic context of the surrounding neighborhood, potentially leading to a lack of global consistency in feature representation. Larger receptive fields can effectively aggregate more homogeneous information, resulting in more stable feature representation. The preference for receptive field scale varies significantly across pixels at different locations; therefore, it is possible to analyze the appropriate receptive field scale for each pixel and assign different weights to it.

[0023] Furthermore, firstly, feature vectors of pixels within receptive fields of different scales are obtained using ASPP. It should be noted that ASPP void space pyramid pooling is an existing technology, and its specific steps will not be elaborated upon. Different void ratios set in the model correspond to receptive fields of different scales; implementers can determine the void ratio according to the implementation scenario. The extracted feature vectors can characterize the semantic expression of the pixel within the corresponding receptive field. For example, data of different dimensions in the feature vector can reflect multi-dimensional features such as the probability that the location is in an ore area, the probability that it is in bare soil, and the probability that it is on a boundary. Then, feature change sensitivity can be obtained based on the differences and changes in the feature vectors of pixels in different receptive fields. Preferably, in this embodiment of the invention, the step of obtaining feature change sensitivity includes:

[0024] In the formula, D represents the sensitivity to feature changes of a pixel. This indicates normalization, and N represents the number of receptive fields at different scales. This represents the feature vector of a pixel in the nth receptive field. This represents the average feature vector of a pixel across all receptive fields. Indicates the calculation of the modulus; This value reflects the dispersion of feature vectors across different receptive fields. A larger value indicates more significant differences in feature vectors across different receptive fields, meaning the pixel is more sensitive to changes in receptive field scale. Conversely, a smaller value indicates that features across multiple receptive fields tend to be consistent, suggesting structural stability at that pixel. The pixels are sorted in ascending order of receptive field scale. This indicates that the pixel is in the nth receptive field and the nth receptive field. The magnitude of the difference in the feature vectors of each receptive field reflects the feature differences between receptive fields at adjacent scales. Indicates satisfaction The number of times the decreasing relationship between adjacent scales is satisfied; a larger value means that the feature change gradually decreases as the receptive field scale increases, showing a clear scale smoothing trend, indicating that the pixel is more sensitive to scale changes. Conversely, a smaller value indicates that the feature change lacks a monotonic pattern, and the scale change of the receptive field has a weaker impact on it. Therefore, a higher sensitivity to feature change means that the feature response of the pixel is more significantly different at different receptive field scales, and shows a strong structural change trend with the change of the receptive field; conversely, a lower sensitivity to feature change means that the feature response of the pixel tends to be consistent at different scales, and the scale change has a weaker impact on it, indicating that the pixel has strong spatial structural stability.

[0025] Furthermore, the feature change sensitivity of a pixel can reflect the difference in the pixel's adaptation to receptive fields of different scales. Therefore, the first weight coefficient of a pixel in different receptive fields can be obtained based on the feature change sensitivity. Preferably, in this embodiment, the step of obtaining the first weight coefficient includes: sorting the receptive fields in ascending order of scale, calculating the ratio of the position of any receptive field to the sum of the positions of all receptive fields, and obtaining the position feature value of any receptive field. The smaller the position feature value, the smaller the scale of the receptive field. When the feature change sensitivity is not lower than a preset threshold, it means that the pixel exhibits a strong structural change trend with the change of the receptive field, indicating that the pixel is sensitive to changes in the scale of the receptive field. The pixel relies more on smaller-scale receptive fields for fine feature characterization to preserve boundary detail information. Therefore, the difference between the constant 1 and the position feature value is calculated to obtain the first weight coefficient of the pixel in any receptive field. The smaller the scale of the arbitrary receptive field, the larger the first weight coefficient of its feature vector. When the sensitivity to feature changes is below a preset threshold, it means that the pixel has strong spatial structural stability. Therefore, it relies more on a larger receptive field for stability expression to enhance global semantic consistency. Thus, the positional feature value is used as the first weight coefficient of the pixel in any receptive field. The larger the scale of the arbitrary receptive field, the larger the first weight coefficient of its feature vector. In this embodiment of the invention, the preset threshold is taken as 1 in the range of feature change sensitivity, and the implementer can determine it according to the implementation scenario. The first weight coefficient can initially assign weights to the feature vectors of the pixel in different receptive fields, so that the feature vectors in the receptive fields with better fit to the pixel have higher weights.

[0026] Step S3: Obtain different difference vectors and principal components based on the difference features of the feature vectors of pixels in the same receptive field and other pixels; obtain directional consistency based on the directional distribution features of the difference vectors and principal components; obtain stability based on the difference features of the feature vectors of other pixels in the receptive field of a pixel; obtain the second weight coefficient of the pixel in different receptive fields based on directional consistency and stability.

[0027] In complex spatial structures within mining areas, receptive fields of different scales correspond to different spatial information aggregation ranges. The feature representation of the same pixel at different scales may be simultaneously influenced by local detail information and global context information, resulting in differences in structural consistency and response stability across multi-scale feature representations. Therefore, it is necessary to analyze the feature representation quality of the same pixel under different receptive fields to distinguish the effectiveness of features at different scales in representing the target semantics. First, different difference vectors and principal components are obtained based on the difference features between the feature vectors of the pixel and other pixels within the same receptive field. Preferably, in this embodiment, the steps of obtaining the difference vectors and principal components include: calculating the difference between the feature vectors of the pixel and other pixels within the corresponding local area of ​​the receptive field and calculating the unit vector of the difference result to obtain the difference vector between the pixel and other pixels; obtaining the principal components of all difference vectors of the pixel within the same receptive field through principal component analysis, where the principal components reflect the dominant direction of the difference vectors. It should be noted that principal component analysis is an existing technology, and the specific steps are not described in detail here. Then, the directional consistency is obtained based on the directional distribution characteristics of the difference vectors and principal components. Preferably, in this embodiment, the steps of obtaining the directional consistency include:

[0028] In the formula, C represents the directional consistency of the pixel. The unit vector representing the principal component. R represents the unit vector of the average difference vector, which is the average of all difference vectors of this pixel within the same receptive field. R represents the number of other pixels within the receptive field of this pixel. This represents the difference vector between the current pixel and the r-th other pixel. The dot product between unit vectors reflects the similarity and alignment between the dominant direction and the average difference vector direction. The larger the result, the more consistent the dominant direction and the average difference vector direction are, which indicates that the region is more likely to present a relatively consistent geographical structure under the current receptive field. This reflects that the receptive field has good stability and consistency in its characterization of local spatial structure, and its corresponding feature vector has higher reliability. It reflects the consistency of the feature change direction of other pixels in the neighborhood of this pixel, when A larger value indicates a more concentrated and consistent direction of feature changes within the neighborhood, signifying good stability and continuity of the features extracted within the receptive field in the local area. Therefore, a higher degree of directional consistency means that the feature change direction of this pixel is highly consistent with that of other pixels in the neighborhood within that receptive field. This receptive field does not introduce significant heterogeneous information interference within the current spatial region, and the features extracted through this receptive field can more stably and centrally reflect the structural change characteristics of the region, exhibiting higher reliability and quality.

[0029] Furthermore, the spatial distribution stability of feature response intensity can reflect the uniformity and stability of feature expression within the current receptive field. When the feature response intensity distribution is relatively concentrated and fluctuates little, it indicates that the feature expression in that region has high stability, and the reliability of the features extracted within that receptive field is higher. Therefore, stability is obtained based on the difference features of the feature vectors of other pixels in the receptive field of a pixel; preferably, in this embodiment of the invention, the step of obtaining stability includes:

[0030] In the formula, Q represents the stability of the pixel. Indicates normalization. This represents the maximum value of the feature vector magnitude of other pixels in the receptive field. This represents the minimum value of the feature vector magnitude of other pixels in the receptive field. The smaller the value, the greater the stability, meaning a smaller fluctuation range in the feature response intensity of each pixel within the receptive field. This indicates better stability and consistency during feature extraction, without being affected by significant local anomalies or multi-structure interference, resulting in higher feature extraction quality for that pixel. After obtaining the orientation consistency and stability of pixels within the same receptive field, a second weighting coefficient for the pixel in different receptive fields can be obtained based on the orientation consistency and stability. Preferably, in this embodiment of the invention, the step of obtaining the second weighting coefficient includes:

[0031] In the formula, This represents the second weight coefficient of the pixel in the nth receptive field. This indicates the directional consistency of the pixel in the nth receptive field. The range of values ​​is from -2 to 2. This represents the stability of the pixel in the nth receptive field. A larger second weight coefficient means better quality feature extraction for the pixel within that receptive field, and consequently, a higher weight for its feature vector in multi-scale fusion.

[0032] Step S4: Obtain the fusion weight of the pixel under different receptive fields based on the first weight coefficient and the second weight coefficient; obtain the weighted feature vector of the pixel based on the fusion weight and the feature vector; obtain the mining area segmentation map through the semantic segmentation model based on the weighted feature vector.

[0033] After obtaining the first weight coefficient and the second weight coefficient of the pixel under different receptive fields, the fusion weight of the pixel under different receptive fields can be obtained according to the first weight coefficient and the second weight coefficient; preferably, in this embodiment of the invention, the step of obtaining the fusion weight includes:

[0034] In the formula, This represents the fusion weight of the pixel in the nth receptive field, where N represents the number of receptive fields. This represents the first weight coefficient of the pixel in the nth receptive field. This represents the second weight coefficient of the pixel in the nth receptive field. A larger fusion weight means that the feature vector of the corresponding receptive field has a higher weight in the multi-scale feature fusion process.

[0035] After obtaining the fusion weights for receptive fields at different scales, the weighted feature vector of each pixel can be obtained based on the fusion weights and the feature vector, specifically including: In the formula This represents the weighted feature vector of that pixel, where N represents the number of receptive fields. This represents the fusion weight of the pixel in the nth receptive field. This represents the feature vector of the pixel in the nth receptive field. By calculating the weights of different receptive fields, we can make the feature vectors with better adaptability and higher feature quality contribute more during the fusion process, thus making the fusion result more accurately reflect the structural features at that pixel.

[0036] Furthermore, a segmentation map of the mining area can be obtained using a semantic segmentation model based on the weighted feature vectors. Specifically, this involves obtaining the weighted feature vectors of each pixel in the monitoring image, upsampling the weighted and fused high-level semantic features using the DeepLabv3+ standard decoding structure, and concatenating and convolving them with the shallow features to form a decoded feature representation. It should be noted that the DeepLabv3+ semantic segmentation model is an existing technology, and the specific steps will not be elaborated further. A pixel-by-pixel Softmax classification operation is performed on the decoded feature map to obtain the probability distribution of each pixel in different categories. The category of each pixel is determined using the maximum probability principle, thereby generating an initial semantic segmentation result map and obtaining the segmentation result of the mining area, improving the accuracy and reliability of the segmentation. Finally, based on the segmentation result map, area calculation and spatial overlay analysis are combined to achieve effective monitoring of the mining area's mining scope and changes. Thus, by adaptively weighting the feature vectors of pixels under different receptive fields in the ASPP, the semantic segmentation result is more accurate, improving the accuracy of mining resource monitoring.

[0037] In summary, this invention provides a method for low-altitude monitoring of mineral resources based on UAV vision. It obtains feature change sensitivity based on the differences and variations in feature vectors of pixels in different receptive fields; acquires a first weight coefficient for pixels in different receptive fields based on the feature change sensitivity; obtains a difference vector and principal components based on the differences in feature vectors of pixels within the same receptive field; obtains directional consistency based on the directional distribution characteristics of the difference vector and principal components; obtains stability based on the feature vectors of other pixels in the receptive field of a pixel; acquires a second weight coefficient for pixels in different receptive fields based on directional consistency and stability; and obtains a fusion weight based on the first and second weight coefficients. This invention obtains a weighted feature vector based on the fusion weight and feature vector; and obtains a mining area segmentation map based on the weighted feature vector, thus improving the accuracy of mineral resource monitoring.

[0038] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0039] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for low-altitude monitoring of mineral resources based on UAV vision, characterized in that, The method includes the following steps: Acquire monitoring images of the mining area; The feature vectors of pixels in receptive fields of different scales are obtained through ASPP; the feature change sensitivity is obtained based on the difference and change features of the feature vectors of the pixels in different receptive fields; and the first weight coefficient of the pixels in different receptive fields is obtained based on the feature change sensitivity. Based on the difference features of the feature vectors of the pixel and other pixels in the same receptive field, different difference vectors and principal components are obtained; based on the directional distribution features of the difference vectors and principal components, directional consistency is obtained; based on the difference features of the feature vectors of other pixels in the receptive field of the pixel, stability is obtained; based on the directional consistency and stability, a second weighting coefficient of the pixel in different receptive fields is obtained. The fusion weight of the pixel under different receptive fields is obtained based on the first weight coefficient and the second weight coefficient; the weighted feature vector of the pixel is obtained based on the fusion weight and the feature vector; and the mining area segmentation map is obtained through a semantic segmentation model based on the weighted feature vector. The step of obtaining the feature change sensitivity based on the difference and change features of the feature vectors of the pixel in different receptive fields includes: In the formula, D represents the sensitivity of a pixel to feature changes. This indicates normalization, and N represents the number of receptive fields at different scales. This represents the feature vector of a pixel in the nth receptive field. This represents the average feature vector of a pixel across all receptive fields. This indicates the calculated modulus; the modulus is sorted in ascending order of receptive field scale. This indicates that the pixel is in the nth receptive field and the nth receptive field. The magnitude of the difference between the feature vectors of each receptive field. Indicates satisfaction Quantity; The step of obtaining the directional consistency degree based on the difference vector and the directional distribution characteristics of the principal components includes: In the formula, C represents the directional consistency of the pixel. The unit vector representing the principal component. R represents the unit vector representing the average difference vector, and R represents the number of other pixels within the receptive field of the pixel. This represents the difference vector between the pixel and the r-th other pixel. The step of obtaining stability based on the difference features of feature vectors of other pixels in the receptive field of the pixel includes: In the formula, Q represents the stability of the pixel. Indicates normalization. This represents the maximum value of the feature vector magnitude of other pixels in the receptive field. This represents the minimum value of the feature vector magnitude of other pixels in the receptive field; The step of obtaining the weighted feature vector of a pixel based on the fusion weight and the feature vector includes: In the formula This represents the weighted feature vector of the pixel, where N represents the number of receptive fields. This represents the fusion weight of the pixel in the nth receptive field. This represents the feature vector of the pixel in the nth receptive field.

2. The method for low-altitude monitoring of mineral resources based on UAV vision according to claim 1, characterized in that, The step of obtaining the first weighting coefficient of the pixel in different receptive fields based on the feature change sensitivity includes: Sort the receptive fields in ascending order of scale, calculate the ratio of the position of any receptive field to the sum of the positions of all receptive fields, and obtain the position feature value of any receptive field; when the sensitivity of the feature change is not lower than a preset threshold, calculate the difference between a constant 1 and the position feature value, and obtain the first weight coefficient of the pixel in any receptive field; when the sensitivity of the feature change is lower than the preset threshold, use the position feature value as the first weight coefficient of the pixel in any receptive field.

3. The method for low-altitude monitoring of mineral resources based on UAV vision according to claim 1, characterized in that, The step of obtaining different difference vectors and principal components based on the difference features between the feature vectors of the pixel and other pixels under the same receptive field includes: Calculate the difference between the feature vector of the pixel and other pixels within the corresponding local receptive field, and calculate the unit vector of the difference result to obtain the difference vector between the pixel and other pixels; obtain the principal components of all difference vectors of the pixel under the same receptive field by principal component analysis.

4. The method for low-altitude monitoring of mineral resources based on UAV vision according to claim 1, characterized in that, The step of obtaining the second weighting coefficient of the pixel in different receptive fields based on the orientation consistency and the stability includes: In the formula This represents the second weighting coefficient of the pixel in the nth receptive field. This indicates the directional consistency of the pixel in the nth receptive field. This represents the stability of the pixel in the nth receptive field.

5. The method for low-altitude monitoring of mineral resources based on UAV vision according to claim 1, characterized in that, The step of obtaining the fusion weight of a pixel under different receptive fields based on the first weight coefficient and the second weight coefficient includes: In the formula This represents the fusion weight of the pixel in the nth receptive field, where N represents the number of receptive fields. This represents the first weight coefficient of the pixel in the nth receptive field. This represents the second weighting coefficient of the pixel in the nth receptive field.

Images