Karst landform image analysis method and system based on deep learning

Abstract

The invention provides a karst landform image analysis method and system based on deep learning, relates to the technical field of deep learning, and accurately determines candidate karst landform semantic features corresponding to each semantic space by extracting basic semantic features of a karst landform image and utilizing a pre-learned karst landform image analysis model. And the candidate karst landform semantic features are loaded to the corresponding sub-models, sub-landform description feature parameters are generated, and then a detailed karst landform description knowledge chain is constructed. According to the method, the accuracy and efficiency of karst landform image recognition are remarkably improved, the images are effectively clustered through the karst landform description knowledge chain, and a data analysis means is provided for the fields of scientific research and exploration of karst landforms, geological disaster prediction, environmental protection and the like.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a karst landform image analysis method and system based on deep learning. Background Art

[0002] Karst landform is a unique and complex landform type on the earth's surface. Its formation and evolution process involves a variety of geological processes, which are of great significance to geological science research, development and protection of natural resources, and prediction and prevention of geological disasters. However, due to the complexity and diversity of karst landforms, traditional image analysis methods are often difficult to accurately and efficiently identify and classify them.

[0003] In the past, image analysis of karst landforms mainly relied on manual visual interpretation and simple image processing techniques. These methods are not only time-consuming and labor-intensive, but also greatly affected by subjective factors, making it difficult to ensure the objectivity and accuracy of the analysis results. With the continuous development of remote sensing technology and image processing technology, although some image analysis methods based on pixels or simple features have emerged, these methods are still incapable of processing complex karst landform images and cannot fully explore and utilize the deep semantic information in the images. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a karst landform image analysis method based on deep learning, the method comprising:

[0005] Acquire a karst landform image to be subjected to image recognition, and extract basic karst landform semantic features corresponding to the karst landform image;

[0006] Determine, based on the basic karst landform semantic features, candidate karst landform semantic features corresponding to each semantic space of the karst landform image analysis model for which parameter learning has been completed in advance;

[0007] The candidate karst landform semantic features corresponding to each semantic space are loaded into the karst landform image analysis sub-model existing in the semantic space corresponding to the karst landform image analysis model, so as to determine the karst landform description knowledge chain based on the sub-landform description feature parameters generated based on each karst landform image analysis sub-model;

[0008] A karst landform description knowledge chain generated by the karst landform image analysis model is obtained, and the karst landform image is clustered based on the karst landform description knowledge chain.

[0009] On the other hand, an embodiment of the present invention also provides a karst landform image analysis system based on deep learning, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present application extracts the basic semantic features of karst landform images and uses the pre-learned karst landform image analysis model to accurately determine the candidate karst landform semantic features corresponding to each semantic space. These candidate karst landform semantic features are loaded into the corresponding sub-model to generate sub-landform description feature parameters, and then a detailed karst landform description knowledge chain is constructed. This method not only significantly improves the accuracy and efficiency of karst landform image recognition, but also effectively clusters images through the karst landform description knowledge chain, providing data analysis means for the fields of scientific research exploration of karst landforms, geological disaster prediction and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the karst landform image analysis method based on deep learning provided in an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of the karst landform image analysis system based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a karst landform image analysis method based on deep learning provided by an embodiment of the present invention. The karst landform image analysis method based on deep learning is introduced in detail below.

[0014] Step S110, obtaining a karst landform image to be subjected to image recognition, and extracting basic karst landform semantic features corresponding to the karst landform image.

[0015] In this embodiment, for a currently ongoing project on karst landform research in a certain area, karst landform images to be image recognized can be obtained from an aerial photography image library of the area. These karst landform images cover karst landform areas at different angles and under different lighting conditions.

[0016] First, the karst landform image is decomposed to generate multiple target landform feature blocks. For example, in a large-area karst landform image, the karst landform image may be decomposed into several smaller areas according to the natural structure of the landform, such as peaks, valleys, caves and other obvious terrain features. These areas become target landform feature blocks.

[0017] Then, for each target geomorphic feature block, extract its corresponding block semantic features. For example, for a target geomorphic feature block containing a cave entrance, first unify its image format. Assuming that the original image is a special aerial photography format, convert it into a universal image format, such as JPEG format. Then standardize the image size and adjust the length and width of the image to a preset standard size, such as 500×500 pixels. Then perform normalization so that the pixel value of the image is within a specific range so that the subsequent algorithm can better process it.

[0018] Then, a multi-scale image analysis algorithm is used to construct a multi-scale image representation. For example, the scale range to be used is determined to be from 1 / 2 to 1 / 8 of the original size. For each scale in the scale range, the scale feature images at different scales are obtained by downsampling (the mean pooling method is used here, that is, the average pixel value of each small area is calculated to reduce the number of pixels) and upsampling (the bilinear interpolation method is used to calculate the new pixel value based on the surrounding pixel values ​​to increase the image size). Then, these scale feature images at different scales are combined to generate a multi-scale image representation of the target landform feature block.

[0019] Then, a local feature description algorithm is used to construct a Gaussian difference pyramid at each scale of the multi-scale image representation to detect local extreme points. For the local extreme points of the rock texture detected around the cave entrance, the corresponding directional histogram is calculated. For example, with the local extreme point as the center, the grayscale value changes of pixels in a certain range around (such as a 10×10 pixel range) are taken to calculate the directional histogram. The pixels in the surrounding preset range are counted according to the predefined directional interval (such as 0-360 degrees divided into 8 intervals) to obtain a histogram representing the directional distribution, which can describe the texture features around the local extreme point.

[0020] The directional histograms of all local extreme points are combined to generate the local feature description of the target geomorphic feature block. For example, the texture characteristics of the rocks around the cave entrance are reflected through this local feature description.

[0021] A sorting algorithm based on feature importance is used to sort local features according to their importance in describing the target geomorphic feature block. For example, for the image of the cave entrance, local features that reflect the texture of the rock edge may be more important than local features in other smooth areas. After removing local features whose importance is lower than the set importance (based on the threshold determined by previous experience or experiments), similar local features in the retained local features are fused through the feature fusion algorithm to generate an optimized local feature description.

[0022] Then, a shape-based global feature extraction method is used to extract shape features from the multi-scale image representation. For the target geomorphic feature block of the cave entrance, its edge curve is obtained through the edge detection algorithm, and then the edge curve is characterized, such as calculating the curvature of the curve. At the same time, the regional shape characteristics are described by calculating the area, perimeter, aspect ratio and other parameters of the block.

[0023] A color-based global feature extraction method is used to extract color features from multi-scale image representations. For example, the color histogram of the rocks around the cave entrance is statistically analyzed, that is, the distribution frequency of different colors is statistically analyzed, and color moments such as color mean, variance, skewness and other parameters are calculated to describe color features.

[0024] The shape features and color features are combined to generate a global feature description of the target geomorphic feature block, which can reflect the overall characteristics of the target geomorphic feature block.

[0025] The global feature description is mapped to a high-dimensional feature space, and the low-dimensional global feature description is converted into a high-dimensional feature representation through a nonlinear mapping function (such as a mapping function based on a neural network).

[0026] Finally, based on the relative importance of global features and local features in describing the target geomorphic feature blocks, a feature fusion algorithm is used to fuse the global feature description and the local feature description to obtain the fused block semantic features. The fused block semantic features of all target geomorphic feature blocks are integrated to generate the basic karst landform semantic features corresponding to the karst landform image.

[0027] Step S120, determining candidate karst landform semantic features corresponding to each semantic space of a karst landform image analysis model for which parameter learning has been completed in advance based on the basic karst landform semantic features.

[0028] In this embodiment, a karst landform image analysis model with pre-completed parameter learning is configured. This karst landform image analysis model has multiple semantic spaces, and each semantic space is used to process different characteristic aspects of karst landforms.

[0029] First, the basic karst landform semantic features are dimensionally normalized to generate the target karst landform semantic features. For example, the basic karst landform semantic features may contain multi-dimensional information such as texture features, shape features, and color features at different scales. Through a specific normalization algorithm (such as mapping the value of each dimension to between 0 and 1), the target karst landform semantic features are obtained.

[0030] According to the pre-defined semantic space of the karst landform image analysis model, the feature dimension or feature combination corresponding to each semantic space is determined. For example, one semantic space may correspond to the dimension related to the terrain undulation characteristics of the karst landform, and another semantic space may correspond to the dimension related to the water distribution characteristics of the karst landform. And the correspondence between each feature item in the semantic feature of the target karst landform and the semantic space is determined according to the pre-established feature mapping table. For example, a feature item related to the terrain height difference in the semantic feature of the target karst landform will be mapped to the semantic space related to the terrain undulation characteristics according to the feature mapping table.

[0031] According to the feature mapping table, each feature item in the target karst landform semantic feature is loaded into the corresponding semantic space of the karst landform image analysis model to generate candidate karst landform semantic features corresponding to each semantic space. For example, the feature items related to the shape of the cave are loaded into the candidate karst landform semantic features corresponding to the semantic space that specifically processes the shape of the cave.

[0032] Step S130, loading the candidate karst landform semantic features corresponding to each semantic space into the karst landform image analysis sub-model existing in the semantic space corresponding to the karst landform image analysis model, so as to determine the karst landform description knowledge chain based on the sub-landform description feature parameters generated based on each karst landform image analysis sub-model.

[0033] It is assumed that the karst landform image analysis model contains multiple karst landform image analysis sub-models, and these karst landform image analysis sub-models process the features of different semantic spaces respectively.

[0034] First, identify the different semantic spaces in the karst landform image analysis model and the karst landform image analysis sub-models corresponding to the semantic spaces. For example, there is a semantic space about the vegetation coverage characteristics of karst landforms, and the corresponding karst landform image analysis sub-model is a model specifically used to analyze vegetation texture, color and other features. For each semantic space, determine the position and functional characteristics of the semantic space in the overall architecture of the karst landform image analysis model, and establish a mapping relationship table. For example, the semantic space of vegetation coverage characteristics is in the part of surface feature analysis in the entire model architecture, and its function is to provide vegetation-related feature descriptions. In the mapping relationship table, this semantic space is used as an entry, and the corresponding karst landform image analysis sub-model is used as the associated content.

[0035] For each semantic space, the candidate karst landform semantic features corresponding to the semantic space are loaded into the corresponding karst landform image analysis sub-model one by one in batches or all at once according to the input interface specification of the corresponding karst landform image analysis sub-model. For example, for the candidate karst landform semantic features corresponding to the semantic space of vegetation coverage features, the data may be loaded in batches according to the requirements of the vegetation analysis sub-model. The loaded candidate karst landform semantic features are processed by each karst landform image analysis sub-model to generate sub-landform description feature parameters corresponding to the input candidate karst landform semantic features. For example, after the vegetation analysis sub-model processes the candidate karst landform semantic features of the vegetation coverage features, sub-landform description feature parameters such as vegetation type and vegetation density are generated. These parameters reflect the feature description of karst landforms in different semantic spaces by the karst landform image analysis sub-model.

[0036] Arrange the characteristic parameters of each sub-geomorphology description according to the order of semantic space or the predefined logical order of the karst landform image analysis sub-model. For example, arrange the characteristic parameters of sub-geomorphology description related to terrain features first, then arrange the characteristic parameters of sub-geomorphology description related to vegetation features, etc. And analyze the potential relationship between the characteristic parameters of each sub-geomorphology description to identify the existing target association points.

[0037] A feature identification label is constructed for each sub-geomorphology description feature parameter. For example, for a sub-geomorphology description feature parameter representing the height of a mountain, its feature identification label may be "peak-height". Through the predefined karst landform semantic knowledge base, the feature identification label is mapped to a further semantic description, and the mapped implementation semantic description is attached to the corresponding sub-geomorphology description feature parameter. For example, "peak-height" may be mapped to "the vertical height of the peak in the karst landform has an important influence on the undulation of the landform".

[0038] The semantic similarity between each sub-geomorphology description feature parameter is calculated based on the basic semantic distance calculation strategy, and the sub-geomorphology description feature parameters are grouped according to the semantic similarity to obtain an initial grouping sequence. For example, sub-geomorphology description feature parameters related to peak height and peak slope may be grouped into the same group due to semantic similarity, because they are both related to the topographic features of the peak.

[0039] The initial grouping sequence is optimized to obtain an optimized target grouping sequence. The number of sub-geomorphology description feature parameters in each initial group is checked. If the number of sub-geomorphology description feature parameters in any initial group is greater than the first set number (assuming 5), the advanced semantic distance calculation strategy is re-adopted to calculate the semantic similarity between the sub-geomorphology description feature parameters in the initial group to further subdivide the initial group. For example, if there are 8 sub-geomorphology description feature parameters related to peaks in a group, it may be subdivided into more detailed groups such as peak height, peak slope, and peak surface area through the advanced semantic distance calculation strategy. If the number of sub-geomorphology description feature parameters in any initial group is less than the second set number (assuming 2), the initial group is merged with the adjacent group, and the semantic similarity of the sub-geomorphology description feature parameters in the merged group is re-evaluated. If it meets the preset semantic similarity requirements, the merged group is retained, otherwise the original group status is maintained.

[0040] For each group in the optimized target grouping sequence, the relative position of the landform features corresponding to the sub-landform description feature parameters in the group in the karst landform image is analyzed. For example, for a group containing peak height and peak slope, if the landform features corresponding to them are adjacent in the image, then after marking these sub-landform description feature parameters as sub-landform description feature parameters with spatial association, the corresponding spatial association information is added to the corresponding group to generate a group combining spatial relations.

[0041] Analyze the feature dependency between sub-geomorphology description feature parameters in each group combined with spatial relationship. In each group, if the geomorphology feature corresponding to a sub-geomorphology description feature parameter is a dependent condition of the geomorphology feature corresponding to another sub-geomorphology description feature parameter, then the corresponding sub-geomorphology description feature parameter is marked as having feature dependency, and the corresponding feature dependency information is added to the group to generate a group combined with feature dependency. For example, in a group containing the size of the cave entrance and the length of the cave internal channel, the size of the cave entrance may be a dependent condition of the length of the cave internal channel (a large entrance may have a relatively long internal channel), so their feature dependency is marked.

[0042] In each grouping combined with feature dependencies, the sub-geomorphology description feature parameter with the most association relationships is searched and determined as the target association point. For example, in a group containing multiple sub-geomorphology description feature parameters related to mountain peaks, the mountain peak height may be determined as the target association point because it has more semantic similarity associations, spatial relationship associations, and feature dependency associations with other parameters.

[0043] Each target association point is used as a node in the target association structure, and for each target association point, according to the association relationship between the target association point and other sub-landform description feature parameters, the relevant sub-landform description feature parameters are connected to the target association point to construct the target association structure. For example, taking the peak height as the target association point, the sub-landform description feature parameters related to it, such as the peak slope and the peak surface area, are connected to the node.

[0044] Determine the starting point of the karst landform description knowledge chain, which is the sub-landform description feature parameter related to the most significant feature of the karst landform. For example, if a huge sinkhole is the most significant feature in the entire karst landform image, then the sub-landform description feature parameter related to the depth or diameter of the sinkhole may be determined as the starting point.

[0045] According to the predetermined construction rules, other sub-geomorphological description characteristic parameters are gradually connected to the karst landform description knowledge chain based on the target association structure until all sub-geomorphological description characteristic parameters are included in the karst landform description knowledge chain to generate the final karst landform description knowledge chain. For example, starting from the starting point of the tiankeng depth, the sub-geomorphological description characteristic parameters related to the surrounding terrain and vegetation of the tiankeng are sequentially connected according to the target association structure, and a complete karst landform description knowledge chain is gradually constructed.

[0046] Step S140, obtaining a karst landform description knowledge chain generated by the karst landform image analysis model, and clustering the karst landform image based on the karst landform description knowledge chain.

[0047] In this embodiment, the karst landform description knowledge chain includes the relationship between various features in the karst landform image and a detailed landform description.

[0048] Based on this karst landform description knowledge chain, karst landform images are clustered. For example, if the karst landform description knowledge chain contains different types of karst landform features, such as a group of sub-landform description feature parameters with tiankeng as the main feature, and another group of sub-landform description feature parameters with cave groups as the main feature, etc., karst landform images can be clustered according to these different feature combinations. Images with similar karst landform description knowledge chains (i.e., similar landform feature combinations and relationships) are classified into one category. For example, karst landform images with large tiankengs and sparse vegetation as the main features are clustered together, while karst landform images with dense cave groups and abundant underground rivers as the main features are clustered into another category. In this way, by clustering based on the karst landform description knowledge chain, karst landform images can be better classified and analyzed, which is helpful for further research on karst landforms, such as studying the distribution laws and formation mechanisms of different types of karst landforms.

[0049] Based on the above steps, the embodiment of the present application extracts the basic semantic features of karst landform images and uses the pre-learned karst landform image analysis model to accurately determine the candidate karst landform semantic features corresponding to each semantic space. These candidate karst landform semantic features are loaded into the corresponding sub-model to generate sub-landform description feature parameters, and then a detailed karst landform description knowledge chain is constructed. This method not only significantly improves the accuracy and efficiency of karst landform image recognition, but also effectively clusters images through the karst landform description knowledge chain, providing data analysis means for the fields of scientific research exploration of karst landforms, geological disaster prediction and environmental protection.

[0050] In a possible implementation, step S110 specifically includes:

[0051] Step S111, decomposing the karst landform image to generate a plurality of target landform feature blocks.

[0052] Step S112, extracting the block semantic features corresponding to each target geomorphic feature block, and determining the basic karst landform semantic features corresponding to the karst landform image based on the block semantic features corresponding to each target geomorphic feature block.

[0053] In a possible implementation, after step S111, the method further includes:

[0054] Step S1111, obtaining a karst landform feature screening map sequence, and when there is no landform feature block matching the karst landform feature screening map sequence in the multiple target landform feature blocks, the karst landform image is screened out.

[0055] Step S1112: when any one of the plurality of target geomorphic feature blocks matches a geomorphic feature block in the karst geomorphic feature screening map sequence, jump to step S112.

[0056] In this embodiment, for the process of extracting the basic karst landform semantic features corresponding to the karst landform image, for a karst landform image from a specific area, this karst landform image contains rich karst landform information, such as various shapes of peaks, caves and surface signs of underground rivers. First, the karst landform image is decomposed to generate multiple target landform feature blocks. For example, based on obvious natural boundaries such as the outline of the peak, the entrance of the cave, and the winding trend of the underground river on the surface, the image is divided into different target landform feature blocks. One of the target landform feature blocks may be a separate peak area, and the other may be an area containing the entrance of the cave and its surrounding rock structure. Next, the corresponding block semantic features are extracted for each target landform feature block, and the basic karst landform semantic features corresponding to the karst landform image are determined based on the block semantic features corresponding to each target landform feature block.

[0057] Among them, after the karst landform image is decomposed to generate multiple target geomorphic feature blocks, a karst landform feature screening map sequence will be obtained. This karst landform feature screening map sequence contains some specific geomorphic feature patterns, which may be geomorphic features that are determined by previous studies to be unnecessary for further analysis or that do not meet the current research focus. For example, this screening map sequence may contain some small geomorphic feature patterns that have been fully studied and are not closely related to the current research purpose, such as some isolated, very small image patterns of karst pits. Search in multiple target geomorphic feature blocks, if there is no geomorphic feature block matching the karst landform feature screening map sequence, the karst landform image is screened out, because this karst landform image may not contain the karst landform features of interest in this embodiment or the features it contains have been excluded from the research scope. If there is any geomorphic feature block matching the karst landform feature screening map sequence in the multiple target geomorphic feature blocks, it jumps to the step of extracting the block semantic features corresponding to each target geomorphic feature block, and continues the subsequent analysis and processing.

[0058] In a possible implementation, step S112 specifically includes:

[0059] Step S1121, performing image format unification processing, image size standardization processing and normalization processing on each target geomorphic feature block in turn to generate a pre-processed target geomorphic feature block.

[0060] In this embodiment, taking the target geomorphic feature block containing the peak area mentioned above as an example, it is assumed that the format of the original acquired image is a special format that is only applicable to a specific acquisition device. This format may have compatibility issues in the subsequent processing algorithm. Through the unified processing of the image format, it is converted into a universal format that is widely supported by the image processing algorithm, such as the TIFF format. The standardization of the image size is to ensure that the different target geomorphic feature blocks have consistency in size during the subsequent processing process, so as to facilitate unified algorithm operation. For the target geomorphic feature block in the peak area, if its original size is large and irregular, such as 800 pixels long and 600 pixels wide, it is adjusted to a preset standard size, such as 400 pixels long and wide. The normalization process is to map the pixel value of the image to a specific interval, such as normalizing the pixel value to between 0 and 1, which helps to improve the accuracy and efficiency of the subsequent algorithm processing. After this series of preprocessing operations, the preprocessed target geomorphic feature block suitable for subsequent analysis is obtained.

[0061] Step S1122, using a multi-scale image analysis algorithm to construct a multi-scale image representation of the pre-processed target landform feature block, wherein specifically, firstly determine the scale range to be used, and for each scale in the scale range, obtain scale feature images at different scales through downsampling and upsampling algorithms, wherein downsampling uses a mean pooling method and upsampling uses a bilinear interpolation method, and then combine the scale feature images at different scales to generate the multi-scale image representation.

[0062] Taking the target geomorphic feature block containing the cave entrance and its surrounding rock structure as an example, the scale range to be used is first determined. According to the experience and requirements of karst landform image analysis, the scale range from 1 / 3 to 1 / 9 of the original size is determined to be appropriate. For each scale in this scale range, the scale feature images at different scales are obtained by downsampling and upsampling algorithms. Downsampling uses the mean pooling method. For example, at a certain scale, the target geomorphic feature block is divided into several small areas, and the average value of the pixels in each small area is calculated as the pixel value after downsampling, thereby reducing the number of pixels in the image. Upsampling uses the bilinear interpolation method. When it is necessary to restore the low-resolution downsampled image to a higher resolution, the new pixel value is calculated by bilinear interpolation according to the values ​​of the surrounding pixels, thereby obtaining scale feature images at different scales. Finally, these scale feature images at different scales are combined to generate a multi-scale image representation of the target geomorphic feature block of the cave entrance and its surrounding rock structure. This multi-scale image representation can comprehensively reflect the feature information of the target geomorphic feature block at different scales, which is conducive to more accurate feature extraction in the future.

[0063] Step S1123, using a local feature description algorithm, at each scale of the multi-scale image representation, local extreme points are detected by constructing a Gaussian difference pyramid, for each detected local extreme point, a corresponding directional histogram is calculated to describe the texture features around the local extreme point, the directional histogram is calculated based on the change in pixel grayscale values ​​within a preset range around the local extreme point, the pixels within the surrounding preset range are counted according to a predefined directional interval, and a histogram representing the directional distribution is obtained.

[0064] Step S1124, combining the directional histograms of all local extreme value points to generate a local feature description of the target landform feature block, wherein the local feature description is used to reflect the local texture and structural features of the target landform feature block.

[0065] In this embodiment, a local feature description algorithm is used to construct a Gaussian difference pyramid at each scale of the multi-scale image representation to detect local extreme points. For the target geomorphic feature block containing surface signs of the underground river, a Gaussian difference pyramid is constructed at each scale of its multi-scale image representation. During the construction process, the image is convolved by Gaussian kernel functions of different scales to detect local extreme points. For each detected local extreme point, the corresponding directional histogram is calculated to describe the texture features around the local extreme point. For example, for the rock texture around the underground river, a certain detected local extreme point is taken as the center, and the pixel grayscale value changes in the preset range (such as 10×10 pixel range) around it are selected to calculate the directional histogram. According to the predefined direction interval (such as 0-360 degrees divided into 8 intervals), the pixels within this range are classified into the corresponding interval according to the direction of their grayscale value change, and a histogram representing the direction distribution is obtained. This directional histogram can describe the texture features around the local extreme point in detail, such as the direction and change trend of the rock texture around the underground river. The directional histograms of all local extreme points are combined to generate a local feature description of the target geomorphic feature block, which is the surface sign of the underground river. This local feature description effectively reflects the local texture and structural characteristics of the target geomorphic feature block.

[0066] Step S1125, using a sorting algorithm based on feature importance, sorting the local features in the local feature description according to their importance in describing the target geomorphic feature block, removing local features whose importance is lower than the set importance, and then fusing similar local features in the retained local features through a feature fusion algorithm to generate an optimized local feature description, wherein the evaluation of feature importance is determined based on the frequency of occurrence of the local feature in the entire local feature description and its correlation with other local features.

[0067] Taking the target geomorphic feature block containing the peak area as an example, in its local feature description, those local features related to the peak contour shape, such as the local features corresponding to the sharp shape of the peak top, may have a higher importance in describing the peak area. However, some local features caused by noise or local minor changes may be relatively unimportant. The evaluation of feature importance is determined based on the frequency of occurrence of local features in the entire local feature description and the correlation with other local features. For example, if a local feature appears frequently in the local feature description and has a strong correlation with other local features that can reflect the main features of the peak, then the importance of this local feature is higher. After removing local features whose importance is lower than the set importance (this set importance is a threshold determined based on a large amount of analysis and experience of karst landform images), similar local features in the retained local features are fused through the feature fusion algorithm to generate an optimized local feature description. For example, some local features describing similar directions of rock texture in the peak area are fused to obtain a more representative optimized local feature description.

[0068] Step S1126, using a shape-based global feature extraction method to extract shape features from the multi-scale image representation, the shape features include contour features and regional shape features, wherein the contour features obtain the edge curve of the target geomorphic feature block through an edge detection algorithm, and perform feature description on the edge curve, and the regional shape features are described by calculating the area, perimeter, and aspect ratio parameters of the target geomorphic feature block.

[0069] In this embodiment, for the target geomorphic feature block including the cave entrance and its surrounding rock structures, the shape features include contour features and regional shape features. The contour feature obtains the edge curve of the target geomorphic feature block through an edge detection algorithm, such as using the Canny edge detection algorithm to accurately detect the edge curve of the cave entrance, and then the edge curve is characterized. The curvature of the edge curve, the length of the curve and other features can be calculated, and these features can reflect the shape characteristics of the cave entrance. The regional shape feature is described by calculating parameters such as the area, perimeter, and aspect ratio of the target geomorphic feature block. For example, the area, perimeter and aspect ratio occupied by the cave entrance and its surrounding rock structures are calculated, and these parameters can describe the shape characteristics of the area as a whole.

[0070] Step S1127, using a color-based global feature extraction method to extract color features from the multi-scale image representation, the color features include a color histogram and a color moment, the color histogram is a statistical distribution frequency of different colors in the target landform feature block, and the color moment is a description of the color feature by calculating the color mean, variance, and skewness parameters of the target landform feature block.

[0071] Taking the target geomorphic feature block containing the peak area as an example, the color features include color histogram and color moment. The color histogram is a statistical distribution frequency of different colors in the target geomorphic feature block. For example, the number of pixels with different gray values ​​or different color channels (if it is a color image) in the peak area is counted as the proportion of the total number of pixels to obtain a color histogram. The color moment describes the color features by calculating the color mean, variance, and skewness parameters of the target geomorphic feature block. Calculating the mean of the color of the peak area can understand the average color level of the area, the variance can reflect the degree of color dispersion, and the skewness can reflect the asymmetry of the color distribution.

[0072] Step S1128, combining the shape feature and the color feature to generate a global feature description of the target geomorphic feature block, wherein the global feature description is used to reflect the overall feature of the target geomorphic feature block.

[0073] For the target geomorphic feature blocks containing surface signs of underground rivers, the shape features (such as the shape features of the contour curve and the area, perimeter and other features of the region) and color features (color histogram and color moment) of the previously extracted areas surrounding the underground river are combined. This global feature description can comprehensively reflect the overall characteristics of the target geomorphic feature blocks, such as the overall shape, color distribution and other comprehensive features of the surface signs of the underground river.

[0074] Step S1129, mapping the global feature description to a high-dimensional feature space, converting the low-dimensional global feature description into a high-dimensional feature representation through a nonlinear mapping function, and based on the relative importance of global features and local features in describing the target geomorphic feature blocks, using a feature fusion algorithm to fuse the global feature description and the local feature description to obtain fused block semantic features, integrating the fused block semantic features of all target geomorphic feature blocks, and generating basic karst landform semantic features corresponding to the karst landform image.

[0075] In this embodiment, taking the target geomorphic feature block including the cave entrance and its surrounding rock structure as an example, a predetermined nonlinear mapping function (such as a mapping function based on a neural network) is used to convert the original low-dimensional global feature description (including a feature description composed of shape features and color features) into a high-dimensional feature representation. This high-dimensional feature representation can more richly represent the feature information of the target geomorphic feature block, which helps to better distinguish different target geomorphic feature blocks in subsequent analysis.

[0076] Further, taking the target geomorphic feature block containing the peak area as an example, if it is found through evaluation that the global feature is of high importance in describing the overall shape and color distribution of the peak, while the local feature is more critical in describing the local texture and special structure of the peak, the global feature description and the local feature description are fused according to their relative importance according to a predetermined fusion algorithm (such as a weighted average fusion algorithm). After the global feature description (including shape and color features) and the local feature description (such as the local texture and structure features of the peak) of the peak area are fused, the fused block semantic feature is obtained, which can comprehensively reflect the various feature information of the peak area.

[0077] Finally, in a karst landform image, there are multiple target geomorphic feature blocks, including peak areas, cave entrance areas, underground river surface sign areas, etc., and the fused block semantic features are obtained for these areas respectively. The fused block semantic features of these different target geomorphic feature blocks are integrated. The integration process may include combining and arranging different features according to certain rules, and finally generating basic karst landform semantic features that can fully reflect the characteristics of the entire karst landform image. This basic karst landform semantic feature covers various geomorphic feature information in the karst landform image, providing a basis for subsequent analysis and processing.

[0078] In a possible implementation, step S120 includes:

[0079] Step S121, performing dimension normalization processing on the basic karst landform semantic features to generate target karst landform semantic features.

[0080] In this embodiment, taking the basic karst landform semantic features extracted and integrated from the karst landform image as an example, these basic karst landform semantic features may contain various feature information from different target landform feature blocks, such as the shape, texture, and color features of the peak area, the relevant features of the cave entrance area, and the features of the underground river surface sign area, etc. These features may have different value ranges and data distributions in terms of dimension. Dimension normalization processing aims to perform unified standardization operations on these different dimensional features so that each feature dimension is within a predetermined standard range. For example, for the feature dimension representing the height of the peak, its original value may vary within a large numerical range. Through dimensional normalization processing, it is mapped to a numerical range between 0 and 1; for the feature dimension representing the mean color of the cave entrance, a similar normalization operation is also performed. In this way, each feature dimension in the entire basic karst landform semantic feature is processed, and finally a target karst landform semantic feature is generated. This target karst landform semantic feature is consistent in dimension, which is convenient for subsequent interaction with the karst landform image analysis model.

[0081] Step S122, according to the predefined semantic space of the karst landform image analysis model, determine the feature dimension or feature combination corresponding to each semantic space, and determine the correspondence between each feature item in the target karst landform semantic feature and the semantic space according to a pre-established feature mapping table.

[0082] Next, it is assumed that the karst landform image analysis model has multiple pre-defined semantic spaces, one of which is defined as a space specifically processing the karst landform terrain undulation related features. The feature dimensions corresponding to this semantic space may include feature dimensions such as the height of the peak, the slope, and the depth of the valley, or a combination of these feature dimensions; another semantic space may be a space specifically processing the karst landform water body related features, and its corresponding feature dimensions may be feature dimensions such as the width of the underground river, the area of ​​the water body in the cave, or a combination of these feature dimensions. The pre-established feature mapping table records in detail the mapping relationship between each feature item in the target karst landform semantic feature and these semantic spaces. For example, in the target karst landform semantic feature, the feature item of the peak height will be mapped to the semantic space processing the terrain undulation related features according to the feature mapping table; and the feature item of the water body area in the cave will be mapped to the semantic space processing the water body related features. In this way, the corresponding relationship between each feature item and the semantic space is accurately determined.

[0083] Step S123, according to the feature mapping table, each feature item in the target karst landform semantic feature is loaded into the corresponding semantic space of the karst landform image analysis model, and candidate karst landform semantic features corresponding to each semantic space are generated.

[0084] Finally, for the semantic space that processes the features related to terrain undulation, the feature items related to the peak height, slope and valley depth in the target karst landform semantic features are loaded into this semantic space according to the feature mapping table. These loaded feature items constitute the candidate karst landform semantic features corresponding to this semantic space; for the semantic space that processes the features related to water bodies, the feature items related to the width of underground rivers and the area of ​​water bodies in caves in the target karst landform semantic features are loaded into the semantic space, thereby generating the candidate karst landform semantic features corresponding to this semantic space. For other semantic spaces in the karst landform image analysis model, in the same way, the corresponding feature items are loaded into their respective semantic spaces according to the feature mapping table, and finally the candidate karst landform semantic features corresponding to each semantic space are generated. These candidate karst landform semantic features will serve as the basis for subsequent operations and will be further loaded into the karst landform image analysis sub-model in the semantic space corresponding to the karst landform image analysis model for processing, thereby promoting the progress of the entire karst landform image analysis process.

[0085] In a possible implementation manner, before step S130, the method further includes:

[0086] Step A110, obtaining a sample karst landform image, and extracting basic sample karst landform semantic features corresponding to the sample karst landform image.

[0087] In this embodiment, these sample karst landform images are selected to represent karst landforms of different types and characteristics. For example, multiple sample images are selected from a specific karst landform research area, some of which mainly show large-scale peak forest landforms, some of which focus on the complex internal structure of caves, and some of which reflect the relationship between surface rivers and karst landforms. These sample karst landform images are processed to extract the corresponding basic sample karst landform semantic features. Taking the sample image of the peak forest landform as an example, the method for extracting basic karst landform semantic features for karst landform images described above is used. First, the peak forest landform sample image is decomposed to generate multiple target landform feature blocks, which may include single peaks, valleys in peak groups, and other areas. Then, each target landform feature block is subjected to image format unification, image size standardization, and normalization in turn to generate a pre-processed target landform feature block. Then, a multi-scale image analysis algorithm is used to construct its multi-scale image representation, and a local feature description algorithm is used to obtain a local feature description. A sorting algorithm based on feature importance is used to perform feature screening and fusion to obtain an optimized local feature description. A global feature extraction method based on shape and color is used to obtain a global feature description, which is mapped to a high-dimensional feature space. Finally, based on the relative importance of global and local features, the fused block semantic features are obtained. The fused block semantic features of all target landform feature blocks are integrated to obtain the basic sample karst landform semantic features corresponding to the peak forest landform sample image.

[0088] Step A120, determining the input sample karst landform semantic features corresponding to each semantic space of the karst landform image analysis model based on the basic sample karst landform semantic features.

[0089] It is assumed that the karst landform image analysis model has a special processing terrain structure semantic space, vegetation cover semantic space, water body distribution semantic space, etc. For the basic sample karst landform semantic features of the peak forest landform sample image, the features related to the peak height, slope, etc. will be determined as the input sample karst landform semantic features corresponding to the terrain structure semantic space; the features related to the type and density of vegetation in the peak forest area will be determined as the input sample karst landform semantic features corresponding to the vegetation cover semantic space; and the features related to the surface rivers or water bodies in the caves will be determined as the input sample karst landform semantic features corresponding to the water body distribution semantic space. In this way, according to the definitions and requirements of different semantic spaces, the input sample karst landform semantic features corresponding to each semantic space are determined from the basic sample karst landform semantic features.

[0090] Step A130, obtaining a sample karst landform description knowledge chain corresponding to the sample karst landform image.

[0091] For the sample image of the peak forest landform mentioned above, its sample karst landform description knowledge chain contains the description of the relationship between the various features of the peak forest landform and the overall description of the landform characteristics. For example, in this knowledge chain, the height and slope of the peak may have a certain correlation with the vegetation coverage, and this correlation may be based on field surveys or laws derived from previous studies. At the same time, the distribution of surface rivers may also have a certain relationship with the formation and evolution of the peaks, and these relationships will be recorded in the sample karst landform description knowledge chain. For the sample image of the internal structure of the cave, its sample karst landform description knowledge chain will describe the relationship between the size and shape of the cave and the internal rock structure, water distribution, and possible signs of biological activity.

[0092] Step A140, loading the input sample karst landform semantic features corresponding to each semantic space into the semantic space corresponding to the karst landform image analysis model, determining the sample karst landform description knowledge chain as the expected learning output result of the karst landform image analysis model, performing parameter learning on the karst landform image analysis model, and generating a target karst landform image analysis model, wherein the target karst landform image analysis model includes multiple karst landform image analysis sub-models, and the neuron weight parameters in each karst landform image analysis sub-model are different.

[0093] Taking the processing of terrain structure semantic space as an example, the input sample karst landform semantic features corresponding to the semantic space previously determined are loaded into the module corresponding to the terrain structure semantic space in the karst landform image analysis model. The same operation is performed for other semantic spaces such as vegetation coverage semantic space and water body distribution semantic space. After loading, the karst landform image analysis model uses the sample karst landform description knowledge chain as the expected learning output result for parameter learning. During the learning process, the weight parameters of each neuron in the model will be continuously adjusted. Since the target karst landform image analysis model includes multiple karst landform image analysis sub-models, the neuron weight parameters in each karst landform image analysis sub-model are different. For example, in the karst landform image analysis sub-model processing terrain structure semantic space, the neuron weight parameters will be adjusted according to the input sample karst landform semantic features related to terrain structure and the corresponding parts in the sample karst landform description knowledge chain, so that the sub-model can better analyze and process the features related to terrain structure. For other karst landform image analysis sub-models, such as the sub-models for processing vegetation coverage and water body distribution, similar parameter adjustments will be made according to the corresponding input sample karst landform semantic characteristics and sample karst landform description knowledge chain, and finally a target karst landform image analysis model will be generated. This model can be used for more accurate processing and analysis in the subsequent analysis of karst landform images.

[0094] In a possible implementation manner, the step S130 specifically includes:

[0095] Step S131, identifying different semantic spaces in the karst landform image analysis model and the karst landform image analysis sub-models corresponding to the semantic spaces, and for each semantic space, determining the position and functional characteristics of the semantic space in the overall architecture of the karst landform image analysis model, and establishing a mapping relationship table, in which each semantic space is an entry and the corresponding karst landform image analysis sub-model is the associated content.

[0096] Step S132: for each semantic space, the candidate karst landform semantic features corresponding to the semantic space are loaded into the corresponding karst landform image analysis sub-model one by one in batches or all at once according to the input interface specification of the corresponding karst landform image analysis sub-model; the loaded candidate karst landform semantic features are processed by each karst landform image analysis sub-model to generate sub-landform description feature parameters corresponding to the input candidate karst landform semantic features; the sub-landform description feature parameters reflect the feature description of karst landform in different semantic spaces by the karst landform image analysis sub-model.

[0097] Step S133, arrange the characteristic parameters describing each sub-topography according to the order of the semantic space or the predefined logical order of the karst landform image analysis sub-model, analyze the potential relationship between the characteristic parameters describing each sub-topography to identify the existing target association points, and establish a target association structure based on the target association points.

[0098] Step S134, determining the starting point of the karst landform description knowledge chain, wherein the starting point is the sub-landform description characteristic parameter related to the most significant feature of the karst landform.

[0099] Step S135, according to the predetermined construction rules, gradually based on the target association structure, other sub-geomorphology description feature parameters are connected to the karst landform description knowledge chain until all sub-geomorphology description feature parameters are incorporated into the karst landform description knowledge chain to generate the final karst landform description knowledge chain.

[0100] In a possible implementation, step S133 includes:

[0101] Step S1331, constructing a feature identification tag for each sub-geomorphology description feature parameter, wherein the feature identification tag is used to describe the feature of the karst landform represented by the sub-geomorphology description feature parameter.

[0102] Step S1332, mapping the feature identification label to a further semantic description through a predefined karst landform semantic knowledge base, and attaching the mapped implementation semantic description to the corresponding sub-landform description feature parameter.

[0103] Step S1333, based on the basic semantic distance calculation strategy, the semantic similarity between each sub-geomorphology description feature parameter is calculated, and the sub-geomorphology description feature parameters are grouped according to the semantic similarity to obtain an initial grouping sequence, wherein each initial group in the initial grouping sequence reflects the potential relationship between the sub-geomorphology description feature parameters based on semantics, and the sub-geomorphology description feature parameters within the same initial group are correlated at some feature levels of the karst landform.

[0104] Step S1334, optimizing the initial grouping sequence to obtain an optimized target grouping sequence, wherein the number of sub-topography description feature parameters in each initial group is specifically checked; if the number of sub-topography description feature parameters in any initial group is greater than a first set number, the advanced semantic distance calculation strategy is re-adopted to calculate the semantic similarity between the sub-topography description feature parameters in the initial group to further subdivide the initial group; if the number of sub-topography description feature parameters in any initial group is less than a second set number, the initial group is merged with the adjacent group, and the semantic similarity of the sub-topography description feature parameters in the merged group is re-evaluated; if the preset semantic similarity requirement is met, the merged group is retained, otherwise the original grouping state is maintained; wherein the second set number is less than the first set number.

[0105] Step S1335, for each group in the optimized target grouping sequence, analyze the relative positions of the landform features corresponding to the sub-landform description feature parameters within the group in the karst landform image; if the landform features corresponding to any two or more sub-landform description feature parameters are adjacent or containment relationships in the karst landform image, then after marking any two or more sub-landform description feature parameters as sub-landform description feature parameters with spatial association, add the corresponding spatial association information to the corresponding group to generate a group combining spatial relationships.

[0106] Step S1336, analyze the feature dependency relationship between the sub-topography description feature parameters within each group combined with spatial relationships. Within each group, if the topography feature corresponding to a sub-topography description feature parameter is a dependent condition of the topography feature corresponding to another sub-topography description feature parameter, then the corresponding sub-topography description feature parameter is marked as having a feature dependency relationship, and the corresponding feature dependency information is added to the group to generate a group combined with feature dependencies.

[0107] Step S1337, in each group combined with feature dependencies, search for the sub-landform description feature parameters with the most association relationships, and determine the sub-landform description feature parameters with the most association relationships as the target association points, wherein the association relationships include semantic similarity associations, spatial relationship associations, and feature dependency associations.

[0108] Step S1338, each of the target association points is used as a node in the target association structure, and for each target association point, according to the association relationship between the target association point and other sub-landform description feature parameters, the relevant sub-landform description feature parameters are connected to the target association point to construct a target association structure.

[0109] In this embodiment, the karst landform image analysis model contains multiple semantic spaces. For example, there is a semantic space specifically for processing karst landform terrain features, such as semantic spaces related to features such as peak height, valley depth, and slope gradient; there is also a semantic space related to water body features in karst landforms, including the flow of underground rivers, the area of ​​water bodies in caves, etc.; and a semantic space related to vegetation coverage, involving the type, density, and distribution of vegetation. Each semantic space has a corresponding karst landform image analysis sub-model. Taking the terrain feature semantic space as an example, its corresponding karst landform image analysis sub-model is a module specifically used to process terrain-related data, and its internal algorithm and structure are optimized for the analysis and processing of terrain features.

[0110] For each semantic space, determine its position and functional characteristics in the overall architecture of the karst landform image analysis model. The terrain feature semantic space may be located in the part of the model architecture that analyzes the landform infrastructure, and its function is to accurately extract and analyze the terrain-related features of the karst landform. The water body feature semantic space may be under the branch related to water resource analysis, and its main function is to describe the water body conditions in the karst landform. The vegetation cover semantic space is in the area of ​​surface cover analysis and is responsible for processing vegetation-related information. Then establish a mapping relationship table, in which the terrain feature semantic space is used as an entry, and the corresponding karst landform image analysis sub-model (module for processing terrain data) is used as the associated content. Similarly, the water body feature semantic space and its corresponding sub-model, the vegetation cover semantic space and its corresponding sub-model, etc. are all associated one by one in the mapping relationship table.

[0111] Specifically, for each semantic space, the candidate karst landform semantic features corresponding to the semantic space are loaded according to the input interface specification of the corresponding karst landform image analysis sub-model. Taking the terrain feature semantic space as an example, if the input interface specification of the corresponding karst landform image analysis sub-model requires that the data be arranged in a specific format and order, then the candidate karst landform semantic features corresponding to the terrain feature semantic space are processed according to this requirement. Assuming that the candidate karst landform semantic features contain data such as peak height values ​​and valley depth values, it may be necessary to convert these data into a specific data type (such as floating point numbers) and arrange them in a specified order. For the water body feature semantic space, if the corresponding sub-model input interface specification requires data to be input in batches, then the candidate karst landform semantic features related to the water body (such as underground river flow values, water area values ​​in caves, etc.) are loaded into the corresponding karst landform image analysis sub-model one by one in batches according to this requirement.

[0112] The loaded candidate karst landform semantic features are processed by each karst landform image analysis sub-model. The terrain feature sub-model performs complex calculations and analyses on the input candidate karst landform semantic features such as peak height and valley depth. For example, it may calculate sub-landform description feature parameters such as terrain undulation and terrain complexity based on the internal algorithm. These sub-landform description feature parameters reflect the feature description of karst landforms in different semantic spaces by the karst landform image analysis sub-model. After processing the candidate karst landform semantic features such as underground river flow and water area in caves, the water body feature sub-model generates sub-landform description feature parameters related to water bodies, such as the dynamic stability of water bodies, the degree of interaction between water bodies and surrounding landforms, etc.

[0113] Arrange the characteristic parameters of each sub-geomorphology description according to the order of semantic space or the predefined logical order of the karst landform image analysis sub-model. For example, arrange the characteristic parameters of sub-geomorphology description related to terrain features, such as terrain undulation, slope gradient, etc., then arrange the characteristic parameters of sub-geomorphology description related to water features, such as water body dynamic stability, water area in karst caves, etc., and finally arrange the characteristic parameters of sub-geomorphology description related to vegetation coverage, such as vegetation density, vegetation type, etc.

[0114] Analyze the potential relationship between each sub-geomorphology description feature parameter to identify the existing target association points. First, construct a feature identification label for each sub-geomorphology description feature parameter. For the sub-geomorphology description feature parameter of terrain undulation, its feature identification label may be "terrain-undulation degree". Through the predefined karst landform semantic knowledge base, this feature identification label is mapped to a further semantic description. In the karst landform semantic knowledge base, "terrain-undulation degree" may be mapped to "the height change amplitude of the terrain surface relative to the sea level in the karst landform, reflecting the result of karst action on the terrain shaping", and then attach the implementation semantic description of this mapping to the corresponding sub-geomorphology description feature parameter.

[0115] The semantic similarity between each sub-geomorphology description feature parameter is calculated based on the basic semantic distance calculation strategy. For example, the two sub-geomorphology description feature parameters, terrain undulation and slope gradient, may have high semantic similarity because they are both related to the shape and change of the terrain. According to the semantic similarity, the sub-geomorphology description feature parameters are grouped to obtain an initial grouping sequence. Terrain undulation and slope gradient may be grouped into the same group because they are related at the topographic feature level of karst landforms. However, water body dynamic stability and vegetation density may be grouped into different groups because they represent different aspects of geomorphic characteristics.

[0116] Optimize the initial grouping sequence. Check the number of sub-geomorphology description feature parameters in each initial group. Assuming the first set number is 5, if the number of sub-geomorphology description feature parameters in a certain initial group is greater than 5, such as a group containing 8 sub-geomorphology description feature parameters related to terrain, the advanced semantic distance calculation strategy is re-adopted to calculate the semantic similarity between the sub-geomorphology description feature parameters in the initial group to further subdivide the initial group. For example, this group may be subdivided into more detailed sub-groups such as peak height, valley depth, slope gradient, etc. If the number of sub-geomorphology description feature parameters in any initial group is less than the second set number (assuming the second set number is 2), the initial group is merged with the adjacent group, and the semantic similarity of the sub-geomorphology description feature parameters in the merged group is re-evaluated. If the preset semantic similarity requirements are met, the merged group is retained, otherwise the original group status is maintained.

[0117] For each group in the optimized target grouping sequence, the relative position of the landform features corresponding to the sub-landform description feature parameters in the group in the karst landform image is analyzed. For example, in a group containing peak height and slope gradient, if the landform features corresponding to the peak height and the slope gradient are adjacent in the karst landform image, then after marking the two sub-landform description feature parameters of the peak height and the slope gradient as sub-landform description feature parameters with spatial association, the corresponding spatial association information (such as adjacent relationship) is added to the corresponding group to generate a group combining spatial relationships.

[0118] The characteristic dependency relationship between the sub-geomorphological description characteristic parameters in each group combined with spatial relationship is analyzed. In the group containing the water body area and the cave shape in the cave, if the cave shape is a dependent condition of the water body area in the cave (because the shape of the cave will limit the area of ​​the water body), the corresponding sub-geomorphological description characteristic parameter is marked as having a characteristic dependency relationship, and the corresponding characteristic dependency relationship information (such as the restriction relationship of the cave shape on the water body area) is added to the group to generate a group combined with characteristic dependency relationship.

[0119] In each grouping combined with feature dependencies, search for the sub-geomorphology description feature parameters with the most associations and determine them as target association points. For example, in a grouping containing multiple sub-geomorphology description feature parameters related to a peak (such as peak height, peak slope, peak surface area, etc.), the peak height may be determined as a target association point because it has more semantic similarity associations, spatial relationship associations, and feature dependency associations with other parameters. Each target association point is taken as a node in the target association structure. For each target association point, the relevant sub-geomorphology description feature parameters are connected to the target association point according to the association relationship between the target association point and other sub-geomorphology description feature parameters to construct a target association structure. For example, with the peak height as the target association point, the sub-geomorphology description feature parameters related to it, such as the peak slope and the peak surface area, are connected to the node.

[0120] Determine the starting point of the karst landform description knowledge chain, which is the sub-landform description characteristic parameter related to the most significant feature of the karst landform. For example, in a karst landform area, if there is a huge tiankeng, then the sub-landform description characteristic parameter related to the depth or diameter of the tiankeng may be determined as the starting point. Because the tiankeng is the most significant feature in this karst landform area, building a knowledge chain from this most significant feature can better reflect the characteristic relationship of the entire karst landform.

[0121] According to the predetermined construction rules, other sub-geomorphological description characteristic parameters are gradually connected to the karst geomorphological description knowledge chain based on the target association structure until all sub-geomorphological description characteristic parameters are included in the karst geomorphological description knowledge chain to generate the final karst geomorphological description knowledge chain. Starting from the starting point related to the depth of the tiankeng, according to the target association structure, if the depth of the tiankeng is related to the terrain undulation around the tiankeng, the sub-geomorphological description characteristic parameter of the terrain undulation is connected to the knowledge chain. Then, if the terrain undulation has a certain relationship with the vegetation distribution, the sub-geomorphological description characteristic parameter related to the vegetation distribution is connected to the knowledge chain, and so on, all sub-geomorphological description characteristic parameters are gradually included in the karst geomorphological description knowledge chain, and finally a complete knowledge chain that can fully describe the relationship of karst geomorphological characteristics is obtained.

[0122] Figure 2 The hardware structure of the karst landform image analysis system 100 based on deep learning for implementing the karst landform image analysis method based on deep learning provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the deep learning-based karst landform image analysis system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0123] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the deep learning-based karst landform image analysis system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0124] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the karst landform image analysis method based on deep learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0125] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned deep learning-based karst landform image analysis system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0126] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned karst landform image analysis method based on deep learning is implemented.

[0127] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A karst landform image analysis method based on deep learning, characterized in that: The method comprises: Acquire a karst landform image to be subjected to image recognition, and extract basic karst landform semantic features corresponding to the karst landform image; Determine, based on the basic karst landform semantic features, candidate karst landform semantic features corresponding to each semantic space of the karst landform image analysis model for which parameter learning has been completed in advance; The candidate karst landform semantic features corresponding to each semantic space are loaded into the karst landform image analysis sub-model existing in the semantic space corresponding to the karst landform image analysis model, so as to determine the karst landform description knowledge chain based on the sub-landform description feature parameters generated based on each karst landform image analysis sub-model; A karst landform description knowledge chain generated by the karst landform image analysis model is obtained, and the karst landform image is clustered based on the karst landform description knowledge chain.

2. The karst landform image analysis method based on deep learning according to claim 1 is characterized in that: The extracting of basic karst landform semantic features corresponding to the karst landform image specifically includes: Decomposing the karst landform image to generate a plurality of target landform feature blocks; The block semantic features corresponding to each target geomorphic feature block are extracted, and the basic karst landform semantic features corresponding to the karst landform image are determined based on the block semantic features corresponding to each target geomorphic feature block.

3. The karst landform image analysis method based on deep learning according to claim 2 is characterized in that: After the step of decomposing the karst landform image to generate a plurality of target landform feature blocks, the method further comprises: Acquire a karst landform feature screening map sequence, and when there is no landform feature block matching the karst landform feature screening map sequence in the plurality of target landform feature blocks, screen out the karst landform image; When there is any one of the multiple target geomorphic feature blocks that matches the geomorphic feature block in the karst geomorphic feature screening map sequence, the process jumps to the step of extracting the block semantic features corresponding to each target geomorphic feature block.

4. The karst landform image analysis method based on deep learning according to claim 2 is characterized in that: The extracting the block semantic features corresponding to each target geomorphic feature block, and determining the basic karst landform semantic features corresponding to the karst landform image based on the block semantic features corresponding to each target geomorphic feature block specifically includes: Performing image format unification processing, image size standardization processing and normalization processing on each target geomorphic feature block in turn to generate a pre-processed target geomorphic feature block; A multi-scale image analysis algorithm is used to construct a multi-scale image representation of the pre-processed target landform feature block, wherein the scale range to be used is first determined, and for each scale in the scale range, scale feature images at different scales are obtained by downsampling and upsampling algorithms, wherein the downsampling adopts a mean pooling method and the upsampling adopts a bilinear interpolation method, and then the scale feature images at different scales are combined to generate the multi-scale image representation; Using a local feature description algorithm, at each scale of the multi-scale image representation, a Gaussian difference pyramid is constructed to detect local extreme points, and for each detected local extreme point, a corresponding directional histogram is calculated to describe the texture features around the local extreme point, wherein the directional histogram is calculated based on the grayscale value change of pixels in a preset range around the local extreme point, and the pixels in the preset range around the local extreme point are counted according to a predefined directional interval to obtain a histogram representing directional distribution; Combining the directional histograms of all local extreme value points to generate a local feature description of the target geomorphic feature block, wherein the local feature description is used to reflect the local texture and structural features of the target geomorphic feature block; A sorting algorithm based on feature importance is adopted to sort the local features in the local feature description according to their importance in describing the target geomorphic feature block, and after removing the local features whose importance is lower than the set importance, similar local features in the retained local features are fused by a feature fusion algorithm to generate an optimized local feature description, wherein the evaluation of feature importance is determined based on the frequency of occurrence of the local feature in the entire local feature description and its correlation with other local features; A shape-based global feature extraction method is adopted to extract shape features from the multi-scale image representation, wherein the shape features include contour features and regional shape features, wherein the contour features obtain edge curves of target geomorphic feature blocks through edge detection algorithms, and feature descriptions are performed on the edge curves, and the regional shape features are described by calculating area, perimeter, and aspect ratio parameters of the target geomorphic feature blocks; A color-based global feature extraction method is used to extract color features from the multi-scale image representation, wherein the color features include a color histogram and a color moment. The color histogram is a statistical distribution frequency of different colors in the target landform feature block, and the color moment is a description of the color feature by calculating the color mean, variance, and skewness parameters of the target landform feature block. The shape feature and the color feature are combined to generate a global feature description of the target geomorphic feature block, wherein the global feature description is used to reflect the overall feature of the target geomorphic feature block; Mapping the global feature description to a high-dimensional feature space, and converting the low-dimensional global feature description into a high-dimensional feature representation through a nonlinear mapping function; Based on the relative importance of global features and local features in describing the target geomorphic feature block, a feature fusion algorithm is used to fuse the global feature description and the local feature description to obtain a fused block semantic feature; The fused block semantic features of all target geomorphic feature blocks are integrated to generate basic karst landform semantic features corresponding to the karst landform image.

5. The karst landform image analysis method based on deep learning according to claim 1 is characterized in that: The step of determining candidate karst landform semantic features corresponding to each semantic space of the karst landform image analysis model that has completed parameter learning in advance based on the basic karst landform semantic features comprises: Perform dimension normalization on the basic karst landform semantic features to generate the target karst landform semantic features; Determine the feature dimension or feature combination corresponding to each semantic space according to the predefined semantic space of the karst landform image analysis model, and determine the correspondence between each feature item in the target karst landform semantic feature and the semantic space according to a pre-established feature mapping table; According to the feature mapping table, each feature item in the target karst landform semantic feature is loaded into the corresponding semantic space of the karst landform image analysis model to generate candidate karst landform semantic features corresponding to each semantic space.

6. The karst landform image analysis method based on deep learning according to claim 1, characterized in that: Before the step of loading the candidate karst landform semantic features corresponding to each semantic space into the semantic space corresponding to the karst landform image analysis model, the method further includes: Acquire a sample karst landform image, and extract basic sample karst landform semantic features corresponding to the sample karst landform image; Determining the input sample karst landform semantic features corresponding to each semantic space of the karst landform image analysis model based on the basic sample karst landform semantic features; Acquire a sample karst landform description knowledge chain corresponding to the sample karst landform image; The input sample karst landform semantic features corresponding to each semantic space are loaded into the semantic space corresponding to the karst landform image analysis model, the sample karst landform description knowledge chain is determined as the expected learning output result of the karst landform image analysis model, and parameter learning is performed on the karst landform image analysis model to generate a target karst landform image analysis model, wherein the target karst landform image analysis model includes multiple karst landform image analysis sub-models, and the neuron weight parameters in each karst landform image analysis sub-model are different.

7. The karst landform image analysis method based on deep learning according to claim 1, characterized in that: The method of loading the candidate karst landform semantic features corresponding to each semantic space into the karst landform image analysis sub-model existing in the semantic space corresponding to the karst landform image analysis model, so as to determine the karst landform description knowledge chain based on the sub-landform description feature parameters generated based on each karst landform image analysis sub-model, specifically includes: Identify different semantic spaces in the karst landform image analysis model and the karst landform image analysis sub-models corresponding to the semantic spaces, determine the position and functional characteristics of each semantic space in the overall architecture of the karst landform image analysis model, and establish a mapping relationship table, in which each semantic space is used as an entry and the corresponding karst landform image analysis sub-model is used as the associated content; For each semantic space, the candidate karst landform semantic features corresponding to the semantic space are loaded into the corresponding karst landform image analysis sub-model one by one in batches or all at once according to the input interface specification of the corresponding karst landform image analysis sub-model, and the loaded candidate karst landform semantic features are processed by each karst landform image analysis sub-model to generate sub-landform description feature parameters corresponding to the input candidate karst landform semantic features, and the sub-landform description feature parameters reflect the feature description of the karst landform by the karst landform image analysis sub-model in different semantic spaces; Arranging the sub-landform description feature parameters according to the order of the semantic space or the predefined logical order of the karst landform image analysis sub-model, analyzing the potential relationship between the sub-landform description feature parameters to identify the existing target association points, and establishing a target association structure according to the target association points; Determine the starting point of the karst landform description knowledge chain, wherein the starting point is a sub-landform description characteristic parameter related to the most significant feature of the karst landform; According to the predetermined construction rules, other sub-geomorphology description characteristic parameters are gradually connected to the karst landform description knowledge chain based on the target association structure, until all sub-geomorphology description characteristic parameters are incorporated into the karst landform description knowledge chain, thereby generating the final karst landform description knowledge chain.

8. The karst landform image analysis method based on deep learning according to claim 7 is characterized in that: The step of analyzing the potential relationship between the characteristic parameters describing each sub-landform to identify the existing target association points, and establishing the target association structure according to the target association points includes: Constructing a feature identification tag for each sub-geomorphology description feature parameter, wherein the feature identification tag is used to describe the feature of the karst landform represented by the sub-geomorphology description feature parameter; Mapping the feature identification label to a further semantic description through a predefined karst landform semantic knowledge base, and attaching the mapped implementation semantic description to the corresponding sub-landform description feature parameter; The semantic similarity between each sub-geomorphology description feature parameter is calculated based on the basic semantic distance calculation strategy, and the sub-geomorphology description feature parameters are grouped according to the semantic similarity to obtain an initial grouping sequence, wherein each initial group in the initial grouping sequence reflects the potential relationship between the sub-geomorphology description feature parameters based on semantics, and the sub-geomorphology description feature parameters in the same initial group have correlation at some feature levels of karst landforms; The initial grouping sequence is optimized to obtain an optimized target grouping sequence, wherein the number of sub-geomorphology description feature parameters in each initial group is specifically checked; if the number of sub-geomorphology description feature parameters in any initial group is greater than a first set number, the advanced semantic distance calculation strategy is re-adopted to calculate the semantic similarity between the sub-geomorphology description feature parameters in the initial group to further subdivide the initial group; if the number of sub-geomorphology description feature parameters in any initial group is less than a second set number, the initial group is merged with an adjacent group, and the semantic similarity of the sub-geomorphology description feature parameters in the merged group is re-evaluated; if the preset semantic similarity requirement is met, the merged group is retained, otherwise the original grouping state is maintained, wherein the second set number is less than the first set number; For each group in the optimized target grouping sequence, the relative position of the landform features corresponding to the sub-landform description feature parameters in the group in the karst landform image is analyzed; if the landform features corresponding to any two or more sub-landform description feature parameters are adjacent or contain each other in the karst landform image, then after marking the any two or more sub-landform description feature parameters as sub-landform description feature parameters with spatial association, the corresponding spatial association information is added to the corresponding group to generate a group combining the spatial relationship; Analyze the feature dependency relationship between the sub-geomorphology description feature parameters in each group combined with the spatial relationship. In each group, if the geomorphology feature corresponding to a sub-geomorphology description feature parameter is a dependent condition of the geomorphology feature corresponding to another sub-geomorphology description feature parameter, then the sub-geomorphology description feature parameter is marked as having a feature dependency relationship, and the corresponding feature dependency relationship information is added to the group to generate a group combined with the feature dependency relationship; In each grouping combined with feature dependencies, searching for sub-landform description feature parameters with the most association relationships, and determining the sub-landform description feature parameters with the most association relationships as target association points, wherein the association relationships include semantic similarity association, spatial relationship association, and feature dependency association; Each of the target association points is taken as a node in the target association structure, and for each target association point, according to the association relationship between the target association point and other sub-landform description feature parameters, the relevant sub-landform description feature parameters are connected to the target association point to construct the target association structure.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store programs, instructions or codes. When the programs, instructions or codes are executed by the processor, the karst landform image analysis method based on deep learning as described in any one of claims 1 to 8 is implemented.

10. A karst landform image analysis system based on deep learning, characterized in that: The deep learning-based karst landform image analysis system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the deep learning-based karst landform image analysis method described in any one of claims 1 to 8.

Images