A Multi-Scene Hazard Identification System for Natural Resources Based on Deep Learning from Remote Sensing Images

Abstract

This invention relates to the field of image recognition technology and discloses a multi-scene hazard identification system for natural resources based on deep learning of remote sensing images. The system includes: constructing beach length coordinates; generating a wet response field; generating a surface smoothing response field; outputting the beach length wet texture reflection weakening amount; outputting pixel-level hazard probability; generating scene-level hazard energy density; and outputting hazard identification results. This invention transforms remote sensing image analysis from a conventional image plane to an analysis space adapted to the physical partitioning of the tailings dam by constructing beach length coordinates that fit the tailings dam engineering structure. It constructs a continuous wet response field and an illumination-corrected surface smoothing response field to distinguish the appearance and internal performance differences of the surface cementation layer. The mechanism characteristics of tailings dam hazard formation are integrated into the multi-scale guidance stage of the deep learning network, reducing unfounded identification bias and thus balancing the engineering characteristics of tailings dams with the application requirements of remote sensing identification.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a multi-scenario hazard identification system for natural resources based on deep learning of remote sensing images. Background Technology

[0002] Tailings dams are critical engineering facilities for storing tailings after mineral processing in mines, and their safety directly impacts regional ecology and production safety. Using remote sensing imagery and deep learning for hazard identification is currently a major technological approach for natural resource supervision. Existing identification methods mostly focus on the segmentation and detection of conventional surface targets, only capable of dividing simple areas such as water bodies and dams, without addressing the unique structural zoning patterns of tailings dams. Within a tailings dam, there is a continuous gradient of physical properties and sedimentary sedimentation from the clarifier to the dam body; conventional methods only extract features within the image plane and cannot construct a spatial analysis benchmark that aligns with engineering logic.

[0003] Tailings dam surfaces undergo long-term wet-dry cycles, causing fine particles and cementing materials to migrate to the surface and form a surface cemented layer. This structure exhibits a smooth appearance but weakened mechanical properties, a key indicator of potential tailings dam hazards. Existing methods lack the construction of continuous wet response fields and illumination-corrected smooth response fields, failing to eliminate texture misjudgments caused by terrain shadows and unable to extract anomalous variations in wetness and smoothness. Most methods employ unconstrained black-box networks, easily misclassifying ordinary smooth dam surfaces or conventional wet areas as hazard zones, resulting in identification results lacking physical basis.

[0004] Most existing tailings dam hazard identification methods output pixel-level probability maps or multiple dispersed indicators, failing to aggregate the identification results into a single, quantified, scene-level numerical value. This type of output cannot be directly used for hazard classification and cross-dam area comparison, and does not meet the standardized application requirements of natural resource supervision. Furthermore, these methods fail to achieve deep coupling between mechanistic features and deep learning, resulting in insufficient stability of identification results under complex imaging conditions and variable tailings dam operating conditions, making it difficult to support the routine and accurate supervision of tailings dam hazards. Summary of the Invention

[0005] This invention provides a multi-scenario hazard identification system for natural resources based on deep learning of remote sensing images, solving the technical problems mentioned in the background.

[0006] This invention provides a multi-scene hazard identification system for natural resources based on deep learning of remote sensing images, comprising:

[0007] The first module inputs multispectral images and elevation models into the structural analysis network, outputs the clarifier mask, the beach mask, and the dam mask, and uses the distance ratio between the boundary of the clarifier mask and the boundary of the dam mask to construct the beach distance coordinates within the beach mask;

[0008] The second module generates a wet response field by using the normalized difference of reflectance values ​​in a specified band in the multispectral image.

[0009] The third module performs illumination normalization and local gradient energy calculation on the multispectral image to generate a smooth surface response field.

[0010] The fourth module extracts the local rate of change of the wetting response field and the surface smooth response field along the beach length coordinate direction, which are used as the wetting attenuation term and the smooth return term, respectively. The beach surface mid-section enhancement term, wetting attenuation term and smooth return term based on the beach length coordinate are multiplied together to output the beach length wet texture return weakening amount.

[0011] The fifth module inputs multispectral images, wetting response field, surface smoothing response field, beach path coordinates, and beach path wet texture return weakening amount into the guiding network, and outputs pixel-level hazard probability.

[0012] The sixth module normalizes the sum of the product of the pixel-level hazard probability and the beach wetting texture return weakening amount by using the beach mask area, and aggregates the results to generate scene-level hazard energy density.

[0013] The seventh module maps the scene-level hazard energy density to a specified numerical range and outputs the hazard identification results.

[0014] The beneficial effects of this invention are as follows: By constructing a beach-length coordinate system that conforms to the tailings dam engineering structure, this invention transforms remote sensing image analysis from a conventional image plane into an analysis space adapted to the physical partitioning of the tailings dam, ensuring that feature extraction is consistent with the actual evolution of the tailings dam. This invention constructs a continuous wetting response field and an illumination-corrected surface smoothness response field, which can extract the variation characteristics of wetting and smoothness along the beach-length direction, distinguishing the appearance and internal performance differences of the surface cementation layer, and reducing analysis bias caused by terrain shadows and texture interference. This invention integrates the mechanism characteristics of tailings dam hazard formation into the multi-scale guidance stage of a deep learning network, enabling the network learning process to be based on physical logic, reducing unfounded identification bias. This invention aggregates pixel-level identification results into scene-level quantified values, forming a unified hazard identification output. The output results can be directly used for hazard comparison and classification judgment of different tailings dams. Thus, it takes into account both the engineering characteristics of tailings dams and the application needs of remote sensing identification, making the process and results of hazard identification more consistent with the actual use scenarios of natural resource supervision. Attached Figure Description

[0015] Figure 1 This is a flowchart of the calculation process of the natural resource multi-scene hidden danger identification system based on remote sensing image deep learning of the present invention;

[0016] Figure 2 This is a computational scene diagram of the natural resource multi-scene hidden danger identification system based on remote sensing image deep learning according to the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0018] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0019] like Figures 1-2 As shown, the multi-scene hazard identification system for natural resources based on deep learning of remote sensing images includes:

[0020] The first module inputs multispectral images and elevation models into the structural analysis network, outputs the clarifier mask, the beach mask, and the dam mask, and uses the distance ratio between the boundary of the clarifier mask and the boundary of the dam mask to construct the beach distance coordinates within the beach mask;

[0021] The second module generates a wet response field by using the normalized difference of reflectance values ​​in a specified band in the multispectral image.

[0022] The third module performs illumination normalization and local gradient energy calculation on the multispectral image to generate a smooth surface response field.

[0023] The fourth module extracts the local rate of change of the wetting response field and the surface smooth response field along the beach length coordinate direction, which are used as the wetting attenuation term and the smooth return term, respectively. The beach surface mid-section enhancement term, wetting attenuation term and smooth return term based on the beach length coordinate are multiplied together to output the beach length wet texture return weakening amount.

[0024] The fifth module inputs multispectral images, wetting response field, surface smoothing response field, beach path coordinates, and beach path wet texture return weakening amount into the guiding network, and outputs pixel-level hazard probability.

[0025] The sixth module normalizes the sum of the product of the pixel-level hazard probability and the beach wetting texture return weakening amount by using the beach mask area, and aggregates the results to generate scene-level hazard energy density.

[0026] The seventh module maps the scene-level hazard energy density to a specified numerical range and outputs the hazard identification results.

[0027] In one embodiment of the present invention, multispectral imagery and an elevation model are input into a structural analytical network, which outputs a clarifier mask, a beach mask, and a dam mask. The beach distance coordinates are then constructed within the beach mask using the distance ratio between the boundary of the clarifier mask and the boundary of the dam mask. This includes:

[0028] Multispectral images With elevation model Channel splicing is performed to obtain joint input data. Combined input data Input Structure Parsing Network The category probability field is obtained through the following formula. :

[0029] ;

[0030] in For multispectral images, For elevation model, This is for channel splicing operations. For structural analysis networks, For the set of structural analysis network parameters, For normalization function, For category probability fields;

[0031] Based on category probability field For any pixel Generate a mask based on the probability of the highest category. A clarification pool mask was obtained. Beach surface mask and dam body cover :

[0032] ;

[0033] in For pixels In category The mask values ​​below, For pixels Category The probability, , The operation is performed to retrieve the category corresponding to the maximum value;

[0034] Based on beach mask Construct the set of pixels within the beach mask :

[0035] ;

[0036] in The set of pixels within the beach surface mask. For any pixel, Represents a cell Located inside the beach cover;

[0037] Masking of clarification pond With dam body shield Perform boundary extraction operations separately The boundary of the clarification pool mask is obtained. Boundary between the dam body and the membrane ;

[0038] For any pixel Using distance calculation function Calculate the pixels separately to the boundary of the clarification pool mask distance With pixels to the boundary of the dam body shield distance :

[0039] ;

[0040] ;

[0041] in For pixels to the boundary of the clarification pool mask distance, For pixels to the boundary of the dam body shield The distance;

[0042] Based on distance With distance The set of pixels within the beach mask Internal construction of beach process coordinates : ;

[0043] in For pixels Beach coordinates, To prevent constants with a denominator of zero.

[0044] It should be noted that multispectral imagery is remote sensing imagery data used to characterize multi-band surface reflectance information of a target area, and can be acquired through satellite multispectral imaging, airborne multispectral imaging, or UAV multispectral imaging. Elevation models are data used to characterize the surface elevation undulations and topographic relationships of a target area, and can be acquired through surveying results, photogrammetric reconstruction, or laser measurement results. Joint input data is input data formed by merging the channels of multispectral imagery and elevation models at the same pixel location. The structural parsing network is a network used to receive the joint input data and output the response results of each category of pixels. The structural parsing network parameter set is all the learnable parameters fixed after the structural parsing network is trained. The category probability field is a probability expression of the likelihood that each pixel belongs to each category of the clarifier mask, beach mask, and dam mask. A pixel is the smallest positional unit of the multispectral imagery and elevation model on a unified grid. Mask values ​​are binary results used to indicate whether a pixel belongs to the target category. The clarifier mask is a binary result used to mark the extent of the clarifier area. The beach mask is a binary result used to mark the extent of the beach area. The dam mask is a binary result used to mark the extent of the dam area. The set of pixels within the beach mask is the set of all pixels that satisfy the valid values ​​of the beach mask. The clarifier mask boundary is the set of boundary pixels at the boundary between the clarifier mask and the non-clarifier area. The dam mask boundary is the set of boundary pixels at the boundary between the dam mask and the non-dam area. The first distance is the distance from any pixel within the beach mask to the boundary of the clarifier mask. The second distance is the distance from any pixel within the beach mask to the boundary of the dam mask. The beach distance coordinate is a structural position quantity constructed using the first and second distances, used to represent the relative position of the pixel between the boundary of the clarifier mask and the boundary of the dam mask. The constant to prevent the denominator from being zero represents an extremely small positive number to avoid calculation anomalies caused by a zero denominator. The preferred value is 0.000001 to 0.001, and further preferred to be 0.000001. Using this order of magnitude avoids a zero denominator without changing the magnitude relationship between the first and second distances.

[0045] It should be noted that, since the clarifier mask corresponds to the free water concentration area in the tailings dam sub-scene, it serves as the structural starting point in this invention, and subsequent wetting response values ​​and beach length coordinates are referenced to it. Since the beach mask corresponds to the main area gradually exposed and dried after tailings slurry deposition, the pixel set within the beach mask is the main area where continuous wetting changes, surface smoothing changes, and hazard probability changes occur simultaneously. Since the dam mask corresponds to the outer bearing boundary of the tailings dam, the dam mask boundary can serve as the structural endpoint, jointly defining the direction of the beach length coordinates with the clarifier mask boundary. Because the beach length coordinates are not image row / column positions, but rather relative structural position quantities constructed by the first and second distances, they can unify different beach surface pixels to the same engineering direction for comparison. Since the category probability field preserves the response strength of the same pixel to multiple categories, it is more suitable than direct hard classification results for generating stable clarifier masks, beach masks, and dam masks.

[0046] It should be noted that the implementation of the structural analysis network is as follows: First, the multispectral image and the elevation model are mapped to the same grid size. Then, the joint input data is input into the structural analysis network. The network outputs response values ​​for three categories—clarifier mask, beach mask, and dam mask—for each pixel. Subsequently, a normalization function is used to convert the three response values ​​into a category probability field that sums to 1. Finally, the category with the highest category probability is selected for each pixel to generate a mask. The boundary extraction operation for the clarifier mask and dam mask is performed as follows: the clarifier mask and dam mask are scanned pixel by pixel. Effective mask pixels directly adjacent to pixels with different values ​​are determined as boundary pixels. All boundary pixels are then used to form the boundaries of the clarifier mask and the dam mask, respectively. The distance calculation functions for the first and second distances are as follows: for any pixel within the beach mask, the straight-line distance from it to all boundary pixels of the corresponding mask boundary is calculated, and the minimum distance is taken as the distance from that pixel to the corresponding mask boundary. The constant value to prevent the denominator from being zero is selected as follows: a positive number much smaller than the smallest resolvable unit of the effective distance is chosen based on the magnitude of the first distance and the second distance, and this value is kept consistent across all samples to ensure that the magnitude of the beach coordinates is not changed.

[0047] In one embodiment of the present invention, a wetting response field is generated using the normalized difference of reflectance values ​​in a specified band of a multispectral image, including:

[0048] Based on multispectral imagery For any pixel Extracting green light band reflectance values and near-infrared band reflectance ;in For pixels Reflectance value in the green light band, For pixels Reflectance value in the near-infrared band;

[0049] Based on the set of pixels within the beach mask For any pixel The normalized difference term is calculated using the following formula. : ;

[0050] in For pixels The normalized difference term, For pixels Reflectance value in the green light band, For pixels Reflectance value in the near-infrared band, To prevent constants with a denominator of zero;

[0051] For any pixel The normalized difference term is obtained using the following formula. Mapped to wetting response value Generate a humid response field: ;

[0052] in For pixels The humidity response value, It is the set of pixels within the beach surface mask.

[0053] It should be noted that the green band reflectance value is the pixel reflectance value of the multispectral image in the green band, which can be obtained by reading the green band pixel values ​​of the multispectral image. The near-infrared band reflectance value is the pixel reflectance value of the multispectral image in the near-infrared band, which can also be obtained by reading the near-infrared band pixel values ​​of the multispectral image. The normalized difference term is a ratioized intermediate quantity constructed using the green band reflectance value and the near-infrared band reflectance value, used to compress the influence of overall brightness variation. The wetting response value is an expression of the continuous wetting degree of a single pixel obtained by mapping the normalized difference term. The wetting response field is a continuous state field formed by all wetting response values ​​within the beach mask. Because the normalized difference term uses the difference and sum of the green band reflectance value and the near-infrared band reflectance value, it can, to some extent, weaken the direct influence of overall brightness fluctuations on wetting judgment. Since the wetness response field expresses the continuous degree of wetness, rather than just the two results of having water or not having water, the wetness response field can retain the transition information from wet to dry inside the beach surface.

[0054] It should be noted that the extraction method for green band reflectance and near-infrared band reflectance is as follows: From all bands of the multispectral image, the band corresponding to the center wavelength within the green light range is selected as the green band, and the band corresponding to the center wavelength within the near-infrared range is selected as the near-infrared band. Then, the reflectance values ​​of both bands are read from any pixel within the beach mask. The constant value to prevent the denominator from being zero is determined by selecting a positive number much smaller than the minimum effective change in normal reflectance, based on the range of green band reflectance and near-infrared band reflectance values, and maintaining consistency across all samples to avoid altering the relative magnitude of the normalized difference term.

[0055] In one embodiment of the present invention, illuminance normalization and local gradient energy calculation are performed on multispectral images to generate a smooth surface response field, including:

[0056] Based on multispectral imagery For any pixel Extracting red light band reflectance values Green light band reflectance and blue light band reflectance And construct the grayscale field using the following formula :

[0057] ;

[0058] in For pixels Reflectance value in the red light band, For pixels Reflectance value in the green light band, For pixels Reflectance value in the blue light band For pixels grayscale value;

[0059] Based on elevation model and the set of pixels within the beach mask For any pixel The cosine of the local incident angle is calculated using the following formula. :

[0060] ;

[0061] in For elevation model Calculated pixels The slope angle, For elevation model Calculated pixels The slope angle, The zenith angle of the sun. The azimuth of the sun. For pixels Angle of incidence at point, The cosine of the local incident angle;

[0062] Based on the cosine of the local incident angle The set of pixels within the beach mask Calculate the average cosine of the local incident angle. :

[0063] ;

[0064] in The total number of pixels within the beach mask. The average cosine of the local incident angle within the beach mask; then for any pixel The corrected grayscale field is constructed using the following formula. :

[0065] ;

[0066] in For pixels Corrected grayscale value, To prevent constants with a denominator of zero;

[0067] Based on the corrected gray field For any pixel Constructing a pixel Centered on, with radius neighborhood The local gradient energy is then solved using the following formula. :

[0068] ;

[0069] in For pixels Centered on, with radius The neighborhood, For the neighborhood The total number of pixels inside, For the neighborhood The pixels inside, To correct the grayscale field In pixels gradient at, For pixels Local gradient energy;

[0070] Based on local gradient energy The set of pixels within the beach mask Determine the local gradient energy minimum. With the maximum local gradient energy :

[0071] ;

[0072] For any pixel The surface smoothing response value is generated by the following formula. The surface smooth response field is obtained:

[0073] ;

[0074] in This represents the minimum local gradient energy within the beach cover. This represents the maximum value of the local gradient energy within the beach mask. For pixels Surface smoothing response value.

[0075] It should be noted that the red band reflectance value is the pixel reflectance value of the multispectral image in the red band, which can be obtained by reading the red band pixel values ​​of the multispectral image. The blue band reflectance value is the pixel reflectance value of the multispectral image in the blue band, which can be obtained by reading the blue band pixel values ​​of the multispectral image. The grayscale field is a single-channel field formed by combining the red, green, and blue band reflectance values ​​according to a fixed weight. The grayscale value is the value of the grayscale field at a single pixel location. The slope angle is the angle measure of the degree of surface tilt at the pixel location, which is reflected by the elevation model. The aspect angle is the angle measure of the slope orientation at the pixel location, which is reflected by the elevation model. The solar zenith angle is the angle measure of the sun relative to the zenith direction during imaging, which can be obtained by reading image metadata. The solar azimuth angle is the direction angle of the sun on the horizontal plane during imaging, which can be obtained by reading image metadata. The angle of incidence is the angle between the direction of solar illumination and the normal to the Earth's surface. The local incident angle cosine is the cosine value corresponding to the incident angle, used to represent the local illumination level. The local incident angle cosine average is the average result of all local incident angle cosines within the beach mask. The corrected grayscale field is the field formed after normalizing the grayscale field using the local incident angle cosine and the local incident angle cosine average. The corrected grayscale value is the value of the corrected grayscale field at a single pixel location. The specified radius is a preset spatial scale used to construct the neighborhood range, preferably 3 to 15 pixels, with a further preferred value of 5 pixels. This range can balance local texture stability and spatial resolution differences, covering local appearance undulations without crossing excessively large structural regions. The neighborhood is a set of local pixels centered on a pixel and defined by the specified radius. The total number of pixels in the neighborhood is the total number of pixels in the neighborhood. The gradient is the strength of the change in the corrected grayscale field at a local location. The local gradient energy is the local undulation obtained by aggregating the gradient strengths within the neighborhood. The local gradient energy minimum is the minimum of all local gradient energies within the beach mask. The local gradient energy maximum is the maximum of all local gradient energies within the beach mask. The surface smoothing response value is an expression of the smoothness of a single pixel obtained by mapping local gradient energy. The surface smoothing response field is a continuous state field formed by all surface smoothing response values ​​within the beach mask.

[0076] It should be noted that since the local incident angle cosine directly characterizes the degree of matching between the direction of solar illumination and the slope orientation, it can reflect the illuminance differences of the same material at different locations on the slope. Because the corrected grayscale field is obtained by normalizing the grayscale field using the average and cosine values ​​of the local incident angles, it can reduce the direct impact of shadows and slope aspect differences on grayscale intensity. Since the local gradient energy aggregates the intensity of grayscale changes within a neighborhood, a larger local gradient energy usually indicates more pronounced local surface undulations, while a smaller local gradient energy usually indicates a smoother local surface. Because the surface smoothness response field is a continuous state field obtained by inverse mapping of the local gradient energy, it can convert local roughness into local smoothness.

[0077] It should be noted that the slope angle and aspect angle are calculated as follows: in the elevation model, the elevation difference between the current pixel and its neighboring pixels is read, and then the maximum descent direction and slope degree are obtained based on the elevation changes in two orthogonal directions. The slope degree corresponds to the slope angle, and the maximum descent direction corresponds to the aspect angle. The solar zenith angle and solar azimuth angle are obtained by directly reading the solar geometric information from the multispectral image recording and assigning this information to all pixels within the same scene. The specified radius is determined by first determining the minimum resolvable texture range based on the spatial resolution of the multispectral image, and then selecting the pixel radius that can cover local texture but does not cross large-scale structural changes as the specified radius. The gradient calculation is performed by calculating the gray-level change in the adjacent directions for each pixel in the neighborhood in the corrected gray-level field, converting the change into the gradient strength of a single pixel, and finally aggregating all gradient strengths in the neighborhood to obtain the local gradient energy. The constant value to prevent the denominator from being zero is selected as follows: a positive number much smaller than the minimum effective change is chosen for the normal numerical range of the local incident angle cosine and local gradient energy, and this number is fixed throughout the entire process.

[0078] In one embodiment of the present invention, the local rate of change of the wetting response field and the surface smoothing response field along the beach length coordinate direction is extracted and used as the wetting attenuation term and the smoothing return term, respectively. The beach surface mid-section enhancement term, the wetting attenuation term, and the smoothing return term based on the beach length coordinate are multiplied together to output the beach length wet texture return weakening amount, including:

[0079] Based on the set of pixels within the beach mask and beach coordinates For any central pixel Constructing a central pixel Centered on, with radius neighborhood For any neighboring pixel The neighborhood weighting coefficients are constructed using the following formula. :

[0080] ;

[0081] in For neighboring pixels For the central pixel neighborhood weighting coefficients, Center pixel With neighboring pixels The Euclidean distance between them The neighborhood radius, It is an exponential function;

[0082] For the field to be calculated Construct the local rate of change along the beach length coordinate direction. :

[0083] ;

[0084] in Center pixel field value at that location, For neighboring pixels field value at that location, Center pixel Beach coordinates, For neighboring pixels Beach coordinates, To prevent constants with a denominator of zero;

[0085] The field to be calculated Take as the humid response field The wetting attenuation term is extracted using the following formula. :

[0086] ;

[0087] in Center pixel The wetting attenuation term, The local rate of change of the wet response field along the beach length coordinate direction;

[0088] The field to be calculated Taken as the surface smooth response field The smoothed return term is extracted using the following formula. :

[0089] ;

[0090] in Center pixel The smoothed return term, The local rate of change of the surface smooth response field along the beach length coordinate direction;

[0091] Based on beach journey coordinates The reinforcement term in the middle of the beach is constructed by the following formula. :

[0092] ;

[0093] in Center pixel The enhancement in the middle of the beach surface;

[0094] The reinforcement term in the middle of the beach is calculated using the following formula. , Wetting attenuation and smooth turnaround term Couple the output to reduce the amount of wet ripples and weaken the output. :

[0095] ;

[0096] in Center pixel The amount of wet ripples on the beach weakens. It is a linear rectified function, calculated as follows: .

[0097] It should be noted that the neighborhood weighting coefficients are constructed based on the distance between neighboring pixels and the center pixel; the closer the distance, the greater the contribution. The field to be calculated is the input field used to calculate the local rate of change along the beach length coordinate direction, taking the wetting response field and the surface smoothing response field respectively. The Euclidean distance is the straight-line distance between the center pixel and neighboring pixels in a unified grid. The local rate of change along the beach length coordinate direction is the intensity of local change in the field to be calculated along the beach length coordinate direction. The wetting attenuation term is the negative of the local rate of change of the wetting response field along the beach length coordinate direction, used to represent the attenuation intensity from wet to dry. The smoothing return term is the local rate of change of the surface smoothing response field along the beach length coordinate direction, used to represent the enhancement intensity of surface smoothness in the beach length direction. The beach mid-area enhancement term is a structural modulation quantity constructed from the beach length coordinates, used to emphasize the central region of the beach. The beach length wet texture return weakening quantity is a core feature quantity obtained by coupling the beach mid-area enhancement term, the wetting attenuation term, and the smoothing return term.

[0098] It should be noted that since the neighborhood weighting coefficients attenuate according to the distance between the central pixel and neighboring pixels, they can enhance the influence of nearest neighbor pixels on the local rate of change along the beach length coordinate direction. Because the local rate of change along the beach length coordinate direction is not calculated along the horizontal or vertical direction of the image, but rather along the structural direction of the beach length coordinate, this rate of change can express the local evolution trend from the clarifier mask boundary to the dam mask boundary. Since the wetting attenuation term comes from the negative local rate of change of the wetting response field along the beach length coordinate direction, a larger wetting attenuation term indicates a more pronounced change from wet to dry. Since the smoothing return term comes from the local rate of change of the surface smoothing response field along the beach length coordinate direction, a larger smoothing return term indicates a more pronounced enhancement of surface smoothness along the beach length direction. Because the enhancement term in the middle of the beach surface is larger in the middle of the beach length coordinate and smaller at the ends, the enhancement term in the middle of the beach surface can concentrate the response in areas with more pronounced structural transitions. Since the beach wetting ripple return weakening quantity is coupled with the mid-shoal enhancement term, the wetting attenuation term, and the smoothing return term, the beach wetting ripple return weakening quantity can express the strength of the anomalous co-occurrence relationship within the beach surface.

[0099] It should be noted that the neighborhood weighting coefficients are constructed as follows: first, the Euclidean distance between the center pixel and neighboring pixels is calculated; then, this Euclidean distance is used in conjunction with a specified radius to construct the distance attenuation result, and neighboring pixels with closer distances are assigned larger neighborhood weighting coefficients. The local rate of change along the beach path coordinate direction is calculated as follows: using the center pixel as a reference, the field difference and beach path coordinate difference are calculated for all neighboring pixels within the neighborhood; then, the values ​​are weighted using the neighborhood weighting coefficients; finally, the ratio of the weighted field difference to the beach path coordinate difference is used to form the local rate of change of the center pixel along the beach path coordinate direction. The numerical stabilization method for the beach path wet ripple return weakening is as follows: first, the wetting attenuation term and the smoothing return term are positively preserved separately; then, they are multiplied with the beach surface mid-area enhancement term, thereby avoiding the negative change from causing a reverse cancellation of the beach path wet ripple return weakening.

[0100] In one embodiment of the present invention, multispectral images, wetting response fields, surface smoothness response fields, beach length coordinates, and beach length wet ripple return weakening amounts are input into a guiding network to output pixel-level hazard probabilities, including:

[0101] Based on multispectral imagery Humidity response field Surface smooth response field Beach Coordinates and the amount of weakening of wet ripples on the beach. The joint input tensor of the guiding network is constructed using the following formula. :

[0102] ;

[0103] in For multispectral images, For a humid response field, For a smooth surface response field, For the beach journey coordinates, The amount of wet ripples on the beach is weakened by the return of the wet ripples. This is for channel splicing operations. To guide the joint input tensor of the network;

[0104] Joint input tensor based on a guiding network Through convolution kernel size of The convolution operation constructs the initial features of the guiding network. :

[0105] ;

[0106] in To guide the initial characteristics of the network, The kernel size is Convolution operation;

[0107] For guiding the network The hierarchy weakens the wet texture of the beach and reduces the magnitude of the field. Downsampling to the Hierarchical resolution, obtained And construct the guiding network using the following formula. Hierarchical Guide Map :

[0108] ;

[0109] in To weaken the wet ripple return field of the beach process Downsampling to the The result after hierarchical resolution, The kernel size is Convolution operation, For the Sigmoid function, To guide the network Hierarchical guide map To guide network hierarchy labeling;

[0110] Based on the bootstrap network Hierarchical input features And guiding the network Hierarchical Guide Map The first step of the guiding network is obtained through the following formula. Hierarchical features :

[0111] ;

[0112] in To guide the network Hierarchical input features To guide the network Hierarchical features It is a non-linear activation function. For element-wise multiplication, To guide the network Hierarchical Guide Map The result after adding one;

[0113] Based on the final layer features of the guiding network Through upsampling operation The kernel size is Convolution operation and the Sigmoid function Constructing a pixel-level probability field of hidden dangers :

[0114] ;

[0115] in To guide the final layer features of the network, For the final level features The result after upsampling For a pixel-level probability field of potential hazards, for any pixel The probability of a hidden danger at the pixel level is denoted as .

[0116] It should be noted that the guiding network is a deep recognition network that receives multi-source inputs and outputs pixel-level hazard probabilities. The joint input tensor is the network input formed by channel merging of multispectral imagery, wet response field, surface smoothness response field, beach path coordinates, and beach path wet ripple return weakening. The initial features of the guiding network are the basic feature representations obtained by performing convolution operations on the joint input tensor in the first layer. The guiding network layers are the processing levels with different resolutions and depths within the guiding network. The downsampling result is the result of mapping the beach path wet ripple return weakening to the resolution of the corresponding guiding network layer. The guiding map of the guiding network layer is the guiding quantity obtained by performing convolution operations and normalization mapping on the downsampling result. The input features of the guiding network layer are the feature representations of a certain layer of the guiding network before entering the update. The features of the next layer of the guiding network are the feature representations of the next layer formed by performing convolution operations, nonlinear activation processing, and updating the guiding map after the input features of the current layer. The final layer features of the guiding network are the feature representations output by the deepest layer of the guiding network. The upsampling result is the result of restoring the final layer features of the guiding network to the target resolution. The pixel-level hazard probability field is a continuous probability field output by the network for all pixels. The pixel-level hazard probability is the probability value of the pixel-level hazard probability field at a single pixel location.

[0117] It should be noted that, since the hierarchical guidance map of the guidance network maps the beach wetting ripple weakening amount to the guidance amount after each level of resolution, it can constrain the network to prioritize locations with higher beach wetting ripple weakening amounts at multiple scales. Because the pixel-level hazard probability field retains the continuous probability that each pixel belongs to a hazard, it can be used for subsequent pixel-by-pixel analysis as well as scene-level convergence. Since the joint input tensor simultaneously includes multispectral imagery, wetting response field, surface smoothing response field, beach coordinates, and beach wetting ripple weakening amount, the guidance network can simultaneously utilize raw observation information, intermediate response information, and core feature information.

[0118] It should be noted that the guided network is implemented as follows: First, the joint input tensor undergoes a first-layer convolution operation to form the initial features of the guided network. Then, it is processed layer by layer according to the guided network hierarchy. In each layer, the wet ripple weakening amount is first downsampled to match the resolution of the current layer. Then, convolution operations and nonlinear activation processing are performed on the input features of the guided network layer. Finally, the guided network layer guidance map is combined with the processed features to obtain the features of the next layer of the guided network. The downsampling and upsampling results are formed by performing resolution transformation on the wet ripple weakening amount and the final layer features of the guided network according to the required resolution of the current layer, ensuring that the spatial correspondence remains consistent before and after the resolution transformation. The nonlinear activation processing is performed by applying monotonically nonlinear constraints to the features after convolution operation pixel by pixel, so that strong responses are preserved while weak and invalid responses are compressed, thereby enhancing the feature representation ability. The training method for the pixel-level hazard probability field is as follows: the annotation results of known hazard areas and non-hazard areas are used as supervision information to update the parameters of the guiding network. After the training converges, the guiding network is fixed, and then the pixel-level hazard probability field is output from the new joint input tensor.

[0119] In one embodiment of the present invention, the sum of the product of the pixel-level hazard probability and the beach wetting texture return weakening amount, calculated using the beach mask area, is normalized to generate a scene-level hazard energy density, including:

[0120] Based on the set of pixels within the beach mask Pixel-level risk probability and the amount of weakening of wet ripples on the beach. For any pixel Constructing product terms : ;

[0121] in For pixels The product term, For pixels The probability of pixel-level hidden dangers For pixels The amount of wet ripples on the beach weakens. The set of pixels within the beach surface mask;

[0122] Based on product terms The set of pixels within the beach mask Perform a summation operation within the product to obtain the sum of the products. : ;

[0123] in This is the result of summing the products;

[0124] Based on pixels Values ​​taken under the beach cover The set of pixels within the beach mask Internal calculation of beach surface cover area : ;

[0125] in For the area of ​​the beach surface covered by the membrane, For pixels Values ​​taken under the beach surface mask;

[0126] Based on the product summation result With the area of ​​the beach cover The following formula is used to construct the scene-level hazard energy density. : ;

[0127] in For scenario-level potential energy density, To prevent constants with a denominator of zero.

[0128] It should be noted that the product term is the result obtained by multiplying the pixel-level hazard probability by the beach ripple return weakening amount pixel by pixel. The sum of the products is the result of accumulating all product terms within the beach mask. The beach mask area is the area corresponding to the number of effective pixels within the beach mask. The scene-level hazard energy density is the scene-level quantity formed by normalizing the sum of the products to the beach mask area. Since the scene-level hazard energy density combines both the pixel-level hazard probability and the beach ripple return weakening amount, it is not a simple probability mean, but a scene-level response that takes into account both the network recognition result and the core feature strength. The quantification method for the beach mask area is as follows: under a uniform spatial resolution, first count the total number of effective pixels within the beach mask, and then use this number directly as the area normalization base. The constant value to prevent the denominator from being zero is selected as follows: a positive number much smaller than the minimum effective area base is chosen based on the minimum possible value of the beach cover area, so as to ensure that the area normalization process is stable and does not change the relative order between different scenarios.

[0129] In one embodiment of the present invention, the scene-level hazard energy density is mapped to a specified numerical range, and the hazard identification result is output, including:

[0130] Based on scenario-level hazard energy density Using the following formula through the proportionality coefficient and offset coefficient Constructing the mapping input : ;

[0131] in For scenario-level potential energy density, This is the proportionality coefficient. This is the offset coefficient. For mapping input items;

[0132] Based on mapping input items The interval-normalized response is constructed using the Sigmoid function according to the following formula. :

[0133] ;

[0134] in For the Sigmoid function, It is a natural constant. For interval normalized response;

[0135] Determine the lower bound of a specified range of values. and the upper bound of the specified numerical range Using interval normalized response The hazard identification results are output using the following formula. :

[0136] ;

[0137] Will as well as After substituting, we get:

[0138] ;

[0139] in For the results of hazard identification, To specify the lower bound of a numerical range, To specify the upper bound of the numerical range, Energy density of potential hazards at the scene level.

[0140] It should be noted that the scaling factor is a preset coefficient for linearly scaling the scene-level hazard energy density, with a preferred value of 8 to 12, and a further preferred value of 10. After the scene-level hazard energy density has been normalized by the previous steps, using this range can ensure sufficient resolution in the intermediate risk range. The offset factor is a preset coefficient for linearly shifting the scene-level hazard energy density, with a preferred value of -6 to -4, and a further preferred value of -5. This range can align the central sensitive area of ​​the interval normalized response with the main distribution range of the scene-level hazard energy density. The mapping input is the intermediate quantity formed after the scaling factor and offset factor are applied to the scene-level hazard energy density. The interval normalized response is the intermediate response after the mapping input is compressed to a fixed interval by the normalization mapping function. The lower bound of the specified numerical interval is the minimum value of the hazard identification result output interval, with a preferred value of 0, to facilitate uniformly mapping scenarios with no hazards or extremely low hazards to the starting point of the result interval. The upper bound of the specified numerical interval is the maximum value of the hazard identification result output interval, with a preferred value of 100, to facilitate uniformly mapping high-hazard scenarios to the end point of the percentage-based result interval. The hazard identification result is the final single-value result obtained by linear transformation, normalization mapping and interval mapping of the scene-level hazard energy density.

[0141] It should be noted that, because the interval normalized response first compresses the mapping input to a fixed range, and then completes the numerical mapping using the lower and upper bounds of the specified numerical interval, the interval normalized response can uniformly transform the energy density of hazards at different scenario levels to the same result scale. Since the hazard identification result is the final output after single-valued analysis of the scenario-level hazard energy density, it is suitable for direct use in scenario ranking, result comparison, and regulatory interpretation. The proportionality coefficient and offset coefficient are determined as follows: first, the distribution of scenario-level hazard energy density is statistically analyzed on samples with known result levels; then, proportionality coefficients and offset coefficients that differentiate low-level samples from high-level samples in the interval normalized response are selected and kept consistent across all tested scenarios. The lower and upper bounds of the specified numerical interval are determined as follows: a unified result range is pre-set according to regulatory display needs, and the lowest result is mapped to the lower bound of the specified numerical interval, while the highest result is mapped to the upper bound, allowing direct comparison of hazard identification results from different scenarios.

[0142] It should be noted that the data acquisition of this invention consists of three parallel and interconnected stages: remote sensing image data acquisition, elevation model data acquisition, and auxiliary parameter and preprocessing. Remote sensing image data acquisition utilizes satellite multispectral imaging equipment, airborne multispectral imaging equipment, or UAV multispectral imaging equipment to image the target tailings dam area. The imaging operation fully covers the entire target tailings dam area, including the clarifier, beach surface, and all structural areas of the dam. The imaging process synchronously records imaging time, solar zenith angle, and solar azimuth angle information. The acquired multispectral images include at least the red, green, blue, and near-infrared bands, with each band maintaining the same spatial resolution and georegistration accuracy. Elevation model data acquisition employs satellite stereo mapping, airborne photogrammetry, or ground-based lidar measurement to obtain elevation model data covering the same geographical area as the multispectral images. The spatial resolution of the elevation model data is not lower than that of the multispectral images, and the elevation model data and multispectral images are georegistered to ensure a one-to-one correspondence between image data and elevation data at the same pixel location. In the auxiliary parameter and preprocessing stage, the radiometric calibration parameters and geometric correction parameters of the multispectral image are recorded simultaneously during the acquisition process. The multispectral image is then radiometrically calibrated and geometrically corrected to eliminate radiometric and geometric distortions during the imaging process. The corrected multispectral image and elevation model data are unified under the same geographic coordinate system to form the basic dataset to be processed.

[0143] It should be noted that the output results of this invention are divided into two categories: intermediate results and final quantification results. All results have clear physical meaning and engineering application value. Intermediate results include clarifier mask, beach mask, dam mask, beach coordinate field, wetting response field, surface smoothness response field, beach wetting texture weakening field, and pixel-level hazard probability field. All intermediate results are raster data of the same size and geographic coordinates as the input image, which can be directly imported into geographic information system software for visualization and spatial analysis. The final quantification result is the hazard identification result, which is a single numerical value ranging from 0 to 100. The value is positively correlated with the degree of hazard in the tailings dam.

[0144] For example, for a tailings dam that is operating normally and has a uniformly dried surface, the system inputs the corresponding multispectral image and elevation model data. After the entire process is completed, the hazard identification result is 12. This result is in the low range of the numerical range, indicating that the hazard signs of the tailings dam conforming to the wet ripple return weakening mechanism are weak, and the overall state is stable. In the pixel-level hazard probability field output by the system, the proportion of high-probability pixels in the beach area is less than one percent, and there are no continuous high-value areas in the wet ripple return weakening field. The corresponding control requirement for this result is routine inspection, and remote sensing verification can be carried out according to the preset cycle.

[0145] For example, for a tailings dam experiencing frequent wet-dry cycles during the transition between rainy and dry seasons, the system inputs corresponding multispectral imagery and elevation model data. After the entire process, the hazard identification result is 58. This result falls within the median range, indicating that the tailings dam's beach area exhibits characteristic responses consistent with the wet ripple return weakening mechanism, requiring inclusion in the key observation area. In the pixel-level hazard probability field output by the system, there are continuously distributed medium-to-high probability pixels in the transition zone of the middle beach area, and the beach wet ripple return weakening field shows a continuous high-value distribution in the corresponding area. The corresponding control requirements are to increase the frequency of observations, shorten the remote sensing verification cycle, and conduct simultaneous on-site manual inspections.

[0146] For example, for a tailings dam with extensive development of the surface cementation layer and signs of localized structural weakening on the beach surface, the system inputs corresponding multispectral imagery and elevation model data. After the entire process, the hazard identification result is 91. This result is in the high range of the numerical range, indicating that the tailings dam beach area has significant hazard characteristics consistent with the wet ripple return weakening mechanism, requiring on-site verification and engineering treatment. In the pixel-level hazard probability field output by the system, there are large-scale continuous high-probability pixels in the transition zone in the middle of the beach surface, and the wet ripple return weakening field of the beach surface shows a high value distribution across the entire region. The corresponding control requirement for this result is special treatment: immediately stop tailings discharge operations, organize professional institutions to conduct on-site stability testing, and formulate a targeted engineering reinforcement plan.

[0147] The final quantitative results output by the system can be directly entered into the natural resources supervision platform as a quantitative basis for the graded management and control of tailings dam hazards. Different ranges of results correspond to different control levels, realizing standardized and quantitative supervision of tailings dam hazards, which will not be elaborated here.

[0148] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0149] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A multi-scene hazard identification system for natural resources based on deep learning of remote sensing images, characterized in that, include: The first module inputs multispectral images and elevation models into the structural analysis network, outputs the clarifier mask, the beach mask, and the dam mask, and uses the distance ratio between the boundary of the clarifier mask and the boundary of the dam mask to construct the beach distance coordinates within the beach mask; The second module generates a wet response field by using the normalized difference of reflectance values ​​in a specified band in the multispectral image. The third module performs illumination normalization and local gradient energy calculation on the multispectral image to generate a smooth surface response field. The fourth module extracts the local rate of change of the wetting response field and the surface smooth response field along the beach length coordinate direction, which are used as the wetting attenuation term and the smooth return term, respectively. The beach surface mid-section enhancement term, wetting attenuation term and smooth return term based on the beach length coordinate are multiplied together to output the beach length wet texture return weakening amount. The fifth module inputs multispectral images, wetting response field, surface smoothing response field, beach path coordinates, and beach path wet texture return weakening amount into the guiding network, and outputs pixel-level hazard probability. The sixth module normalizes the sum of the product of the pixel-level hazard probability and the beach wetting texture return weakening amount by using the beach mask area, and aggregates the results to generate scene-level hazard energy density. The seventh module maps the scene-level hazard energy density to a specified numerical range and outputs the hazard identification results; The fourth module specifically includes: For any central pixel in the set of pixels within the beach mask, a neighborhood is constructed with the central pixel as the center and a specified radius as the range. The Euclidean distance between the neighboring pixels and the central pixel is calculated, and the neighborhood weighting coefficient is constructed by combining the neighborhood radius with an exponential function. A formula for calculating the local rate of change of the field to be calculated along the beach length coordinate direction is constructed. The difference in field value between the neighboring pixels and the central pixel and the difference in beach length coordinate are weighted and summed using neighborhood weighting coefficients. The field to be calculated is set as the wetting response field. The local rate of change of the wetting response field along the beach length coordinate direction is extracted and negative values ​​are taken to obtain the wetting attenuation term. The field to be calculated is set as the surface smooth response field, and the local rate of change of the surface smooth response field along the beach length coordinate direction is extracted to obtain the smooth return term; Calculate the mid-shoal reinforcement term using beach length coordinates; The wetting attenuation term and the smoothing return term are extracted in the forward direction using a linear rectified function. The enhancement term in the middle of the beach, the wetting attenuation term extracted in the forward direction, and the smoothing return term extracted in the forward direction are multiplied together to output the beach wetting ripple return weakening amount.

2. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 1, characterized in that, Multispectral imagery and elevation model are stitched together to obtain joint input data. The joint input data is then fed into a structural analytical network. The output of the structural analytical network is processed using a normalization function to obtain the category probability field. In the category probability field, select the category with the highest probability for any pixel to generate the clarification pool mask, beach mask, and dam mask; Determine the set of pixels within the beach mask, and extract the boundaries of the clarifier mask and the dam mask; For any pixel in the set of pixels within the beach cover, calculate the first distance from the pixel to the boundary of the clarifier cover, and calculate the second distance from the pixel to the boundary of the dam cover. The beach distance coordinates are constructed within the beach mask by using the ratio between the sum of the first distance and the first distance, the second distance, and a constant to prevent the denominator from being zero.

3. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 1, characterized in that, Extract the green band reflectance value and near-infrared band reflectance value of any pixel based on multispectral imagery; For any pixel in the set of pixels within the beach mask, a normalized difference term is constructed using the green light band reflectance value, the near-infrared band reflectance value, and a constant to prevent the denominator from being zero. Based on the normalized difference term, the wetting response value of any pixel is generated in the pixel set within the beach mask, thus obtaining the wetting response field.

4. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 1, characterized in that, Based on multispectral images, the red band reflectance value, green band reflectance value, and blue band reflectance value of any pixel are extracted to construct a grayscale field; Based on the elevation model and the set of pixels within the beach cover, the slope angle and aspect angle are calculated for any pixel. Combined with the solar zenith angle and solar azimuth angle, the cosine of the local incident angle is obtained. Calculate the average cosine of the local incident angle within the beach mask, and construct a corrected gray field by combining the gray field, the local incident angle cosine, and the average local incident angle cosine with a constant to prevent the denominator from being zero. For any pixel in the set of pixels within the beach mask, a neighborhood centered on the pixel and with a specified radius is constructed. The gradient of the corrected grayscale field is calculated and summed within the neighborhood to solve for the local gradient energy. The minimum and maximum values ​​of the local gradient energy within the beach mask are determined. The local gradient energy, the minimum and maximum values ​​of the local gradient energy, and a constant to prevent the denominator from being zero are used for mapping to generate the surface smooth response value, thus obtaining the surface smooth response field.

5. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 1, characterized in that, The multispectral image, wetting response field, surface smoothing response field, beach length coordinates, and beach length wet texture return weakening amount are processed by channel splicing to construct the joint input tensor of the guiding network; The initial features of the guiding network are obtained by inputting the joint input tensor of the guiding network into a convolution operation with a kernel size of 3×3. For the guiding network layer, the wet ripple weakening amount of the beach process is downsampled according to the corresponding layer resolution, and the downsampled result is processed by convolution operation with a kernel size of 1×1 and normalization mapping function to construct the guiding network layer guiding graph.

6. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 5, characterized in that, Based on the input features of the guiding network layer and the guiding graph of the guiding network layer, convolution operation with a kernel size of 3×3 and non-linear activation processing are performed on the input features of the guiding network layer. The next layer features of the guiding network are then updated by combining the element-wise multiplication operation with the processing result of adding one to the guiding graph of the guiding network layer. The final layer features of the guiding network are upsampled, and the mapping is processed by a 1×1 convolution operation and a normalized mapping function to obtain a pixel-level hazard probability field, outputting the pixel-level hazard probability of any pixel.

7. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 1, characterized in that, For any pixel within the set of pixels in the beach mask, the pixel-level hazard probability is multiplied by the beach wetting texture return weakening amount to obtain the product term; The product terms are summed within the set of pixels inside the beach mask to obtain the product summation result; The area of ​​the beach mask is obtained by summing the values ​​of the pixels under the beach mask within the set of pixels inside the beach mask. The energy density of a scenario-level hazard is generated by dividing the sum of the products by the sum of the beach surface cover area and the constant used to prevent the denominator from being zero.

8. The multi-scene hazard identification system for natural resources based on deep learning of remote sensing images according to claim 1, characterized in that, The scene-level hazard energy density is linearly transformed using a scaling factor and an offset factor to construct a mapping input term. The normalized mapping function is used to process the mapping input to obtain the interval normalized response; Determine the lower and upper bounds of a specified numerical interval, use the interval normalized response to perform numerical mapping on the lower and upper bounds of the specified numerical interval, and output the hazard identification results.

Images