Landslide hazard remote sensing intelligent identification and extraction method and system based on improved samples

By establishing a partitioned classification sample library in the alpine canyon area and using FCN models, combined with landslide disaster characteristics, the problem of landslide disaster identification in complex backgrounds is solved, and high-precision and efficient landslide disaster recognition is achieved.

CN118941968BActive Publication Date: 2024-12-06HUNAN UNIV OF SCI & TECH
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411422780.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-10-12
Publication Date
2024-12-06
Estimated Expiration
2044-10-12

AI Technical Summary

Technical Problem

In the complex context of alpine canyon areas, it is difficult for the prior art to effectively identify and determine potential landslide disasters caused by regional environmental background differences, especially in areas with ravines, dense vegetation and large spatial heterogeneity.

Method used

By establishing a zoning classification sample library, considering the flowing water landform and vegetation distribution characteristics, a secondary zoning plan is proposed, including sparse forest areas, dense forest areas, canyon areas, river valley areas and valley areas, and combining the development and evolution characteristics of landslide disasters, seasonal replacement characteristics and Spatial perspective attitude characteristics, using FCN deep learning network to establish the best multi-parameter model for intelligent landslide disaster judgment.

Benefits of technology

It improves the scientificity and accuracy of landslide disaster recognition, and achieves landslide disaster recognition with high time efficiency and better than 80% accuracy in judgment results, meeting the needs of emergency application scenarios for large-scale landslide disasters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941968B_ABST
    Figure CN118941968B_ABST
Patent Text Reader

Abstract

The present invention discloses a landslide disaster remote sensing intelligent identification and extraction method and system based on improved samples, which belongs to the field of geological disaster remote sensing identification, and includes the steps of: taking into account the surface vegetation and water flow geomorphic features under the complex background of high mountain canyon areas, proposing a secondary zoning scheme, establishing and forming a zoning classification sample library; establishing a difference enhancement sample that takes into account the development and evolution characteristics, seasonal replacement characteristics and spatial perspective posture characteristics of landslide disasters, and improving the singleness of traditional landslide samples; based on the FCN deep learning network framework and the training index parameter value change characteristic curve, automatically matching the number of iterations to achieve the strongest convergence, and establishing the best multi-parameter FCN model for intelligent identification and extraction of landslide disasters. Aiming at the differences in the spatiotemporal distribution of landslide disasters in high mountain canyon areas, the present invention establishes a difference enhancement sample library and an optimal multi-parameter FCN landslide identification model, improves the accuracy of landslide disaster identification, and provides a scientific and reliable reference basis for disaster prevention and mitigation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of geological disaster remote sensing identification, and in particular to a landslide disaster remote sensing intelligent identification extraction method and system based on improved samples. Background Art

[0002] Landslide refers to the phenomenon that the soil and rock layers on the slope slide down the slope under the action of gravity. It is a common geological disaster. Landslides may be caused by many factors, including but not limited to:

[0003] 1. Precipitation: A large amount of rainwater seeps into the soil, reducing the friction between the soil and is a common cause of landslides;

[0004] 2. Earthquakes: The vibrations caused by earthquakes can destroy the stability of slopes and trigger landslides;

[0005] 3. Human factors: such as excessive excavation, quarrying at the foot of the slope, irrigation activities, etc., can change the stress balance of the slope and cause landslides;

[0006] 4. Geological structure: Unstable rock formations, faults and easily eroded soil types are also prone to landslides.

[0007] Measures to prevent and control landslides usually include establishing a monitoring system, implementing slope protection projects, improving drainage conditions and restoring vegetation.

[0008] At present, there are many mature models and methods for remote sensing identification of regional landslide hazards, and they are widely used. However, there are few related studies on the complex background of high mountain canyon areas, which are characterized by criss-crossing gullies, dense vegetation, and large spatial heterogeneity within the area. Summary of the invention

[0009] The purpose of the present invention is to provide a landslide hazard remote sensing intelligent identification and extraction method and system based on improved samples, in order to solve the problem of potential landslides caused by regional environmental background differences under the complex background of high mountain canyon areas, taking into account the characteristics of water flow landforms and vegetation distribution to establish a zoning classification sample library, and fully considering the internal differences in the region, which not only provides a reliable reference basis for landslide sample improvement, but also improves the scientificity and accuracy of landslide hazard identification, and provides an important new idea for landslide hazard identification under the complex background of high mountain canyon areas with large spatial heterogeneity. .

[0010] To achieve the above object, the present invention provides a landslide disaster remote sensing intelligent identification and extraction method based on improved samples, comprising the following steps:

[0011] Step S1: Taking into account the surface vegetation and water geomorphic features of the alpine canyon area, a secondary zoning scheme is proposed. The secondary zoning includes sparse forest area and dense forest area. The sparse forest area includes canyon area, gully area and river valley area. A zoning classification sample library is established and formed;

[0012] Step S2: Establishing a differentially enhanced sample that takes into account the development and evolution characteristics of landslide hazards, seasonal changes, and spatial perspective characteristics;

[0013] Step S3: Based on the FCN deep learning network framework and the characteristic curve of the training index parameter value change, the appropriate number of iterations is automatically matched to achieve the strongest convergence, and the optimal multi-parameter FCN model is established for intelligent identification and extraction of landslide hazards.

[0014] Preferably, step S1 comprises the following steps:

[0015] S101, given the vector data range of the study area;

[0016] S102, according to the surface vegetation coverage in the image, divide the first-level zoning types into sparse forest area and dense forest area;

[0017] S103. Define canyon area types;

[0018] S104. Define the river valley area type;

[0019] S105. Define the type of gully area;

[0020] S106. Define the type of dense forest area;

[0021] S107, performing adaptive partitioning according to the partition types of S102-S106;

[0022] S108. Establish a landslide disaster sample library based on the zoning results.

[0023] Preferably, the canyon area type is deep valley, steep slope, and narrow and long terrain;

[0024] The types of river valley areas are alluvial mudflats and tributary confluence;

[0025] The type of gully area is that the channels are developed and connected;

[0026] The dense forest area type has dense vegetation and flat terrain;

[0027] When dividing the four secondary zones of canyon area, river valley area, river valley area, and gully area, the image was segmented into several slices using the slicing tool in ArcGIS software. Visual interpretation was used to obtain the features of the objects in each slice image, and the slice rectangles were matched one by one to belong to one of the four types: dense forest area, canyon area, gully area, and canyon area.

[0028] Preferably, step S2 comprises the following steps:

[0029] S201, create a new surface vector data, draw a closed curve along the landslide boundary on the image, form a landslide sample, and store it as a record in the shapefile file, the record contains identification number, name, first-level partition type, second-level partition type, area, and remarks attribute information;

[0030] S202, establish a landslide development and evolution difference sample set;

[0031] S203, establishing a landslide spatiotemporal attitude difference sample set;

[0032] S204, establishing a landslide seasonal difference sample set;

[0033] S205. Based on S202-S204, a comprehensive difference enhanced sample library is established to complete the establishment of the improved sample library.

[0034] Preferably, step S3 comprises the following steps:

[0035] S301, input image data to form an input layer;

[0036] S302, performing a first feature extraction on the input image to form a convolution layer;

[0037] S303, performing a second feature extraction on the image to form a pooling layer;

[0038] S304, sampling the features extracted by the convolution layer to restore them to the same size as the input image, forming an upsampling layer;

[0039] S305. The established FCN model consists of 16 convolutional layers, 5 pooling layers, 3 convolutional layers and 1 upsampling layer in sequence. The activation function after each convolutional layer is the ReLU function, and the number of iterations is 6000.

[0040] S306, Take into account , , , , Indicator parameters, multiple adaptive matching;

[0041] S307, establishing an optimal multi-parameter FCN model for intelligent identification and extraction of landslide hazards;

[0042] The features extracted in steps S302 and S303 are both pixel similarity features of the input image and the sample image.

[0043] Preferably, the ReLU function calculation formula is as follows:

[0044] ;

[0045] In the formula, represents the input vector, Represents a vector Transpose, , The initial values ​​are , , , , Indicates Second convolution, represents the intercept, , The value of The initial value of is inferred according to the rule of matrix convolution multiplication.

[0046] Preferably, the loss function value The calculation formula is as follows:

[0047] ;

[0048] In the formula, For a row of data List, represents the true value of the data, Indicates the predicted value obtained by model operation of the data;

[0049] Learning Rate The calculation formula is as follows:

[0050] ;

[0051] In the formula, Represents the parameter set of the updated model, Both represent the parameter set of the model before updating. is the objective function Relative to The gradient of the model parameters.

[0052] Preferably, average intersection-over-union ratio The calculation formula is as follows:

[0053] ;

[0054] In the formula, represents the true value, represents the predicted value, is the number of types, Indicates that Prediction The number of , called false negatives, Indicates that Prediction The number of , called false positives, Indicates that Prediction The quantity is called true, the plus sign indicates addition, and the minus sign indicates subtraction.

[0055] Preferably, average F score The calculation formula is as follows:

[0056] ;

[0057] In the formula, is the number of types, , For accuracy, , is the recall rate;

[0058] Average accuracy The calculation formula is as follows:

[0059] ;

[0060] In the formula, is the number of types, express The number of pixels for which the class prediction is correct, express The number of pixels with true class labels.

[0061] The present invention provides a system for intelligent identification and extraction of landslide disaster remote sensing based on improved samples, comprising:

[0062] Adaptive zoning type establishment module: Based on the vector data range of the given study area, adaptive zoning types are established according to the water landform and vegetation distribution characteristics of the area;

[0063] Difference Enhanced Sample Establishment Module: Combine the differences in landslide development and evolution, spatiotemporal attitude differences, and seasonal changes to establish enhanced sample annotations;

[0064] Multi-parameter model multiple iteration identification and extraction module: Based on the FCN model and the change characteristics of training indicator parameter values, a multi-parameter FCN landslide hazard identification and extraction model with adaptive iteration numbers is established.

[0065] Therefore, the present invention adopts the above-mentioned landslide disaster remote sensing intelligent identification and extraction method and system based on improved samples, which has the following beneficial effects:

[0066] (1) In view of the complex background of high mountain canyon areas, large areas of surface vegetation cover and long-term erosion of flowing landforms have created an environment with gullies, lush vegetation and fragile geology, thus providing potential energy and kinetic energy for the formation and development of landslide disasters. In order to solve the potential landslides caused by regional environmental background differences, the present invention takes into account the characteristics of flowing landforms and vegetation distribution to establish a zoning classification sample library, namely, canyon area, river valley area, gully area and dense forest area. Traditional sample annotation is based on the entire area, while the present invention refines a region into different types of areas, fully considering the intrinsic differences in the region. It not only provides a reliable reference basis for the improvement of landslide samples, but also improves the scientificity and accuracy of landslide disaster identification.

[0067] (2) The present invention takes into account the changing characteristics of the landslide disaster "formation-development-evolution-disaster" process, seasonal changes and spatial perspective posture characteristics, and establishes a difference-enhanced sample, that is, an improved sample; compared with the traditional static landslide sample, it encompasses the dynamic characteristics of the landslide time series.

[0068] (3) The present invention forms a fast convergence state based on the FCN model and multi-index parameter adaptive iteration, and establishes the optimal multi-parameter FCN landslide disaster intelligent identification and extraction model. Its time efficiency can reach the hour level, and the accuracy of the identification result is better than 80%. The model of the present invention meets the current emergency application scenarios of most large-scale landslide disasters. Different from other deep learning models, it has better automatic matching of optimal parameters and avoids the problems of long-term training, multiple iterations and low convergence.

[0069] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0071] Figure 2 A technical roadmap for an embodiment of the present invention;

[0072] Figure 3 This is an example diagram of a canyon area according to an embodiment of the present invention;

[0073] Figure 4 This is an example diagram of a river valley area according to an embodiment of the present invention;

[0074] Figure 5 This is an example diagram of a valley area according to an embodiment of the present invention;

[0075] Figure 6 This is an example diagram of a dense forest area according to an embodiment of the present invention;

[0076] Figure 7 A sample diagram showing the difference in landslide development and evolution according to an embodiment of the present invention;

[0077] Figure 8 A sample (horizontal view) diagram of the time-space attitude difference of a landslide according to an embodiment of the present invention;

[0078] Fig. 9 A sample (side view) diagram of the time-space attitude difference of a landslide according to an embodiment of the present invention;

[0079] Fig.10 A sample map of landslide seasonal changes according to an embodiment of the present invention;

[0080] Fig.11 This is a schematic diagram for explaining the mean intersection-over-union (mIoU) parameter of an embodiment of the present invention;

[0081] Fig.12 Schematic diagram of an optimal multi-parameter FCN model according to an embodiment of the present invention;

[0082] Fig.13 A schematic diagram of a system module according to an embodiment of the present invention;

[0083] Fig.14 A schematic diagram of a system interface according to an embodiment of the present invention; DETAILED DESCRIPTION

[0084] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0085] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0086] Like reference numerals and letters denote similar items in the following drawings, and thus, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0087] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, or are directions or positional relationships in which the product of the invention is usually placed when in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as a limitation on the present invention.

[0088] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "setting", "installation" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0089] like Figure 1 , Figure 2 As shown, the present invention provides a landslide disaster remote sensing intelligent identification and extraction method based on improved samples, comprising the following steps:

[0090] Step S1: Considering the surface vegetation and water features under the complex background of the alpine canyon area, a secondary zoning scheme is proposed, namely, sparse forest area (canyon area, gully area, river valley area) and dense forest area, and a zoning classification sample library is established and formed. The specific steps are as follows:

[0091] S101, given the vector data range of the study area ( .shp); .shp is a storage format of shapefile in the computer, which is used to store vector geographic information system data. Indicates a custom prefix name, and shp is the suffix.

[0092] S102, according to the surface vegetation coverage in the image, divide the first-level zoning types into sparse forest area and dense forest area;

[0093] S103. Define the canyon area type, i.e. deep valley, steep slope, narrow and long terrain, see Table 1 for details;

[0094] S104, define the types of river valley areas, i.e. alluvial mudflats and tributary confluence, see Table 1 for details;

[0095] S105. Define the types of gully areas, i.e., the development of gullies, connections and intersections, as shown in Table 1;

[0096] S106. Define the type of dense forest area, i.e. dense vegetation and flat terrain, see Table 1 for details;

[0097] When performing the above-mentioned secondary zoning, taking the Wenchuan County experimental area as an example, the image was first segmented into rectangles of 512 pixels × 512 pixels through the slicing tool in the ArcGIS software, resulting in 230 slices. Visual interpretation was then used to obtain the features of the landforms in each slice image, such as vegetation, ditches, valleys, rivers, mudflats, terrain, slopes, etc. Finally, the slice rectangles were matched one by one to belong to one of the four types: dense forest area, canyon area, gully area, and canyon area, such as 21 dense forest areas, 106 canyon areas, 54 river valley areas, and 49 gully areas.

[0098] Feature extraction is performed by manually visually observing the quantitative features of the following four areas:

[0099] Canyon area: In the image, the canyon area is mainly distributed in areas with steep terrain, large slopes, and less vegetation. The color is grayish yellow and the texture is uneven, such as the narrow valley of the mainstream of the Minjiang River;

[0100] Valley area: In the image, valleys are mainly distributed on the slopes. They develop along the ridges to the valleys on both sides and appear as stripes, such as dry river gullies.

[0101] River valley area: In the image, river valleys are mainly distributed on both sides of the river. Affected by water scouring and erosion, they form a radial distribution, such as the wide valley of the Minjiang River tributary.

[0102] Dense forest area: Most areas in the image are green (covered by vegetation), with relatively smooth texture and generally flat terrain, such as Chenjiashan Village.

[0103] S107. Perform adaptive zoning according to the zoning types of S102-S106. Adaptive zoning refers to manual matching based on the characteristics of the input image and the zoning type to ensure that the zoning result conforms to the natural geographical status of the experimental area.

[0104] S108. Establish a landslide disaster sample library based on the zoning results.

[0105] Table 1 Partition Types

[0106] .

[0107] Step S2: Establish a difference-enhanced sample that takes into account the development and evolution characteristics of landslide hazards, seasonal changes, and spatial perspective characteristics, thereby improving the singleness of traditional landslide samples. The specific steps are as follows:

[0108] S201. Create a new (surface) vector data, draw a closed curve along the landslide boundary on the image, form a landslide sample, and store it as a shapefile ( shp) file, which contains attribute information such as identification number, name, primary zone type, secondary zone type, area, and remarks;

[0109] S202, establish a landslide development and evolution difference sample set: ① create a new polygon (surface) element; ② edit the landslide boundary, that is, the closed curve. The landslide boundary of the same landslide is different at different stages. Figure 7 As shown, in stage 1, the landslide was initially formed, in stage 2, the outer contour of the landslide became obvious, in stage 3 to stage 4, the landslide gradually moved inward, in stage 5, it evolved into a strip-shaped landslide, and in stage 6, the landslide shrank inward due to the recovery of the surrounding vegetation;

[0110] S203, establish a sample set of landslide spatiotemporal attitude differences: ① create a new polygon (surface) element; ② edit the landslide boundary, that is, the closed curve, and perform integrity editing on the same landslide boundary from the perspective of the 3D image head-on and side views. Figure 8 , Fig. 9 As shown, the landslide contour and boundary are vectorized in detail from the horizontal and side views;

[0111] S204. Establish a sample set of landslide seasonal variation differences: ① Create a new polygon (surface) feature; ② Edit the landslide boundary, i.e., the closed curve. The landslide boundary of the same landslide in different seasons (spring, summer, autumn, and winter) is different. Fig.10 As shown in the figure, affected by the change of time series seasons (rainy season and non-rainy season), the contour of the same landslide expands outward first and then contracts inward from change 1 to change 3, and a similar phenomenon occurs from change 4 to change 6;

[0112] S205. Based on S202-S204, a comprehensive difference enhanced sample library is established, that is, an improved sample library.

[0113] Step S3: Based on the FCN deep learning network framework and the characteristic curve of the change of training index parameter values, the appropriate number of iterations is automatically matched to achieve the strongest convergence, and the optimal multi-parameter FCN model is established for intelligent identification and extraction of landslide hazards. Its partitioning accuracy and overall accuracy are better than 80% and 90% respectively.

[0114] S301, input image data, i.e., input layer;

[0115] S302, performing the first feature extraction on the input image, i.e., the convolution layer;

[0116] S303, in order to reduce the amount of calculation and prevent overfitting, perform a second feature extraction on the image, i.e., a pooling layer;

[0117] The features extracted in steps S302 and S303 are both pixel similarity features of the input image and the sample image.

[0118] S304, sampling the features extracted by the convolution layer to restore them to the same size as the input image, i.e., upsampling layer;

[0119] S305,FCN model consists of 16 convolutional layers, 5 pooling layers, 3 convolutional layers and 1 upsampling layer. The activation function after each convolutional layer is ReLU function, and the number of iterations is 6000;

[0120] S306, Take into account , , , , Indicator parameters, multiple adaptive matching;

[0121] S307, establish the best multi-parameter FCN model for intelligent identification and extraction of landslide hazards, such as Fig.11 shown.

[0122] The ReLU function calculation formula is as follows:

[0123] ;

[0124] In the formula, represents the input vector, Represents a vector Transpose, , The initial values ​​are , , , , Indicates Second convolution, represents the intercept, , The value of The initial value of is inferred according to the rule of matrix convolution multiplication. The specific method of analogy is as follows:

[0125] The matrix is ​​composed of two The matrix is ​​merged into The matrix of , and then multiply it by a coefficient, The matrix is ​​composed of two The matrix is ​​merged into , then multiply it by a coefficient, and so on.

[0126] Similarly, and It is based on the law of matrix convolution multiplication. , and By analogy with the rule of matrix convolution multiplication , each time you get Substituting into the formula .

[0127] Table 2 Training index parameters

[0128] .

[0129] like Fig.13 , Fig.14 The system of the landslide disaster remote sensing intelligent identification and extraction method based on improved samples described in the present invention includes:

[0130] Adaptive zoning type establishment module: Based on the vector data range of the given study area, adaptive zoning types are established according to the water landform and vegetation distribution characteristics of the area;

[0131] Difference Enhanced Sample Establishment Module: Combine the differences in landslide development and evolution, spatiotemporal attitude differences, and seasonal changes to establish enhanced sample annotations;

[0132] Multi-parameter model multiple iteration identification and extraction module: Based on the FCN model and the change characteristics of training indicator parameter values, a multi-parameter FCN landslide hazard identification and extraction model with adaptive iteration numbers is established.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A landslide disaster remote sensing intelligent identification and extraction method based on improved samples, characterized by: The following steps are involved: Step S1: Taking into account the surface vegetation and water geomorphic features of the alpine canyon area, a secondary zoning scheme is proposed. The primary zoning includes sparse forest area and dense forest area, wherein the sparse forest area includes canyon area, gully area, and river valley area, and the secondary zoning includes canyon area, gully area, river valley area, and dense forest area. A zoning classification sample library is established and formed; Step S2: Establishing a differentially enhanced sample that takes into account the development and evolution characteristics of landslide hazards, seasonal changes, and spatial perspective characteristics; Step S3: Based on the FCN deep learning network framework and the characteristic curve of the training index parameter value change, the appropriate number of iterations is automatically matched to achieve the strongest convergence, and the optimal multi-parameter FCN model is established for intelligent identification and extraction of landslide hazards; Step S3 includes the following steps: S301, input image data to form an input layer; S302, performing a first feature extraction on the input image to form a convolution layer; S303, performing a second feature extraction on the image to form a pooling layer; S304, sampling the features extracted by the convolution layer to restore them to the same size as the input image, forming an upsampling layer; S305. The established FCN model consists of 16 convolutional layers, 5 pooling layers, 3 convolutional layers and 1 upsampling layer in sequence. The activation function after each convolutional layer is the ReLU function, and the number of iterations is 6000. S306, taking into account the loss, Lr, mIoU, mFscore, mAcc indicator parameters, multiple adaptive matching is performed; S307, establishing an optimal multi-parameter FCN model for intelligent identification and extraction of landslide hazards; The features extracted in steps S302 and S303 are both pixel similarity features of the input image and the sample image; The ReLU function calculation formula is as follows: In the formula, x represents the input vector, T represents the transposition of the vector W, and W (i) 、b (i) The initial values ​​of (i=0,i=1) are i represents the i-th convolution, b represents the intercept, W (i=2…) 、b (i=2…) The value of is inferred by the initial values ​​of i=0 and i=1 according to the rule of matrix convolution multiplication; The loss function value loss calculation formula is as follows: In the formula, m is the mth column in a row of data, y t Represents the true value of the data, y p Indicates the predicted value obtained by model operation of the data; The learning rate Lr is calculated as follows: Where M new Represents the parameter set of the updated model, M old Both represent the parameter set of the model before updating. is the objective function J relative to M old The gradient of the model parameters; The calculation formula for the average intersection over union (mIoU) is as follows: In the formula, i represents the true value, j represents the predicted value, k is the number of types, V ij represents the number of predictions of i as k, called false negatives, V ji V represents the number of predictions of k as i, called false positives. ii It indicates the number of predicted values ​​of i as i, which is called true. The plus sign indicates addition, and the minus sign indicates subtraction. The mean F score mFscore is calculated as follows: In the formula, k is the number of types, V p For accuracy, V r is the recall rate; The calculation formula of average accuracy mAcc is as follows: In the formula, k is the number of types, V(i) pre_ture Represents the number of pixels predicted correctly for class i, V(i) real_ture Represents the number of pixels with true labels for class i.

2. The landslide disaster remote sensing intelligent identification and extraction method based on improved samples according to claim 1 is characterized by: Step S1 includes the following steps: S101, given the vector data range of the study area; S102, according to the surface vegetation coverage in the image, divide the first-level zoning types into sparse forest area and dense forest area; S103. Define canyon area types; S104. Define the river valley area type; S105. Define the type of gully area; S106. Define the type of dense forest area; S107, performing adaptive partitioning according to the partition types of S102-S106; S108. Establish a landslide disaster sample library based on the zoning results.

3. The landslide disaster remote sensing intelligent identification and extraction method based on improved samples according to claim 2 is characterized by: The canyon area type is deep valleys, steep slopes, and narrow terrain; The types of river valley areas are alluvial mudflats and tributary confluence; The type of gully area is that the channels are developed and connected; The dense forest area type has dense vegetation and flat terrain; When dividing the four secondary zones into canyon area, river valley area, gully area and dense forest area, the image is segmented into several slices through the slicing tool in ArcGIS software, and the features of the objects in each slice image are obtained by visual interpretation. The slice rectangles are matched one by one to belong to one of the four types: canyon area, river valley area, gully area and dense forest area.

4. The landslide disaster remote sensing intelligent identification and extraction method based on improved samples according to claim 3 is characterized by: Step S2 includes the following steps: S201, creating a new surface vector data, drawing a closed curve along the landslide boundary on the image, forming a landslide sample, and storing it as a record in the shapefile file, the record including identification number, name, first-level partition type, second-level partition type, area, and remark attribute information; S202, establish a landslide development and evolution difference sample set; S203, establishing a landslide spatiotemporal attitude difference sample set; S204, establishing a landslide seasonal difference sample set; S205. Based on S202-S204, a comprehensive difference enhanced sample library is established to complete the establishment of the improved sample library.

5. A system for intelligent identification and extraction of landslide disasters based on remote sensing of improved samples according to any one of claims 1 to 4, characterized in that: include Adaptive zoning type establishment module: Based on the vector data range of a given study area, adaptive zoning types are established according to the water landform and vegetation distribution characteristics of the study area; Difference Enhanced Sample Establishment Module: Combine the differences in landslide development and evolution, spatiotemporal attitude differences, and seasonal changes to establish enhanced sample annotations; Multi-parameter model multiple iteration identification and extraction module: Based on the FCN model and the change characteristics of training indicator parameter values, a multi-parameter FCN landslide hazard identification and extraction model with adaptive iteration numbers is established.

Citation Information

Patent Citations

  • Method for extracting landslide disaster information in alpine and valley areas based on unmanned aerial vehicle image data

    CN115638772A

  • Landslide disaster multi-scale space-time research and judgment method, system and equipment and storage medium

    CN118410194A

  • Ground pressure measurement and danger warning system using air lidar

    KR102488400B1

  • Landslide recognition method based on laplacian pyramid remote sensing image fusion

    US11521377B1

  • System and Method for Extremely Efficient Image and Pattern Recognition and Artificial Intelligence Platform

    US20180204111A1