Method for automatically extracting geological disaster information from unmanned aerial vehicle image based on deep learning

Abstract

A deep learning based method for automatic geological disaster extraction from unmanned aerial vehicle image, belonging to the technical field of unmanned aerial vehicle image processing. The method c

Description

[01] This application claims priority to Chinese Patent Application publication No. CN114821376A, filed with the China National Intellectual Property Administration on June 27, 2022, and entitled "METHOD FOR AUTOMATICALLY EXTRACTING GEOLOGICAL DISASTER INFORMATION FROM UNMANNED AERIAL VEHICLE IMAGE BASED ON DEEP LEARNING", which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[02] The present disclosure relates to a technical field of unmanned aerial vehicle (UAV) image processing, and in particular, to a method for automatically extracting geological disaster information from a UAV image based on deep learning. BACKGROUND

[03] Along with the increasingly severe geological disasters in China, it is necessary to apply high-tech means to assist in investigation, which is also an inevitable development trend of future geological disaster investigation. UAV remote sensing technology is of great help to improve accuracy and effectiveness of geological disaster investigation. High-resolution remote sensing data and geological images may be obtained by using the UAV remote sensing technology, which provides important data support for geological disaster evaluation and monitoring and post-disaster reconstruction.

[04] In the conventional art, after images returned by a UAV are obtained, geological disaster information is usually manually determined and extracted based on the returned UAV images. However, the limitation of professional knowledge and experience of technicians and the difficulty in unifying standards of different technicians would cause wrong determination and missed determination, etc. in geological disaster information extraction. In addition, observing the returned images by human also requires high-intensity labor, which incurs the problems of high labor costs and low efficiency, and is difficult to meet a modern requirement of acquisition of geological disaster information. SUMMARY

[05] In view of the foregoing problems, the present disclosure aims to provide a method for automatically extracting geological disaster information from a UAV image based on deep learning.

[06] The objective of the present disclosure is implemented by using the following technical solutions:

[07] According to a first aspect, the present disclosure provides a method for automatically extracting geological disaster information from the UAV image based on deep learning, including:

[08] S1: acquiring the UAV image;

[09] S2: preprocessing the acquired UAV image to obtain a preprocessed UAV image;

[010] S3: extracting features based on the preprocessed UAV image to obtain feature information of the UAV image; and

[011] S4: inputting the feature information of the UAV image into a trained neural network model to obtain a geological disaster information extraction result.

[012] In an implementation, S1 includes:

[013] acquiring the UAV image collected and returned by a UAV from a target region in real time.

[014] In an implementation, S2 includes:

[015] S21: performing projection transformation, radiation correction, image registration and image clipping on the acquired UAV image to obtain a standard UAV image; and

[016] S22: performing enhancement processing based on the standard UAV image to obtain the preprocessed UAV image.

[017] In an implementation, S3 includes:

[018] S31: performing red green blue (RGB) channel separation based on the preprocessed UAV image to obtain red channel component features, green channel component features, and blue channel component features of the UAV image;

[019] S32: extracting infrared reflectivity features based on the preprocessed UAV image, to obtain infrared reflectivity features of the UAV image;

[020] S33: extracting texture features based on the preprocessed UAV image to obtain texture features of the UAV image;

[021] S34: extracting vegetation index features based on the preprocessed UAV image to obtain vegetation index features of the UAV image; and

[022] S35: performing fusion based on the obtained red channel component features, green channel component features, blue channel component features, infrared reflectivity features, and vegetation index features, to obtain a feature matrix of the UAV image.

[023] In an implementation, S4 includes:

[024] S41: obtaining the feature matrix of the UAV image;

[025] S42: obtaining a reference image feature matrix corresponding to the UAV image; and

[026] S43: inputting the feature matrix of the UAV image and the reference image feature matrix as input set into the trained neural network model to obtain the geological disaster information extraction result output by the neural network model.

[027] In an implementation, the trained neural network model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a first fully connected layer, a second fully connected layer, and a softmax layer which are sequentially connected;

[028] where input of the input layer is the feature matrix of the UAV image and the corresponding reference image feature matrix; the first convolutional layer and the second convolutional layer each include 32 convolution kernels with sizes of 3x3 and 5x5 respectively, the pooling layer is maxpooling, with a size of 3x3, the first fully connected layer includes 128 neurons, and the second fully connected layer includes 16 neurons, with an output of a feature vector that reflects whether the UAV image contains geological disaster information and types of geological disasters; and the softmax layer performs classification based on the feature vector output by the second fully connected layer, and outputs the geological disaster information extraction result.

[029] In an implementation, the method further includes:

[030] SB1: training the neural network model, including:

[031] obtaining two groups of UAV images obtained in different time periods at a same position, where the UAV image captured in an earlier time period is taken as a first image, and the UAV image captured in a later time period is taken as a second image; the first image and the second image may be UAV images before and after a geological disaster at the same position, and the geological disasters include cracks and landslides;

[032] extracting features based on the first image and the second image respectively to obtain a feature matrix corresponding to the first image and a feature matrix corresponding to the second image, and constructing a training set by using the feature matrix corresponding to the first image, the feature matrix corresponding to the second image, and correspondingly-calibrated geological disaster information extraction marks; and

[033] training the neural network model based on the constructed training set, testing the neural network model by using a test set, and obtaining the trained neural network model when a pass rate of the neural network model reaches a predetermined standard.

[034] According to a second aspect, the present disclosure provides a system for automatically extracting geological disaster information from the UAV image based on deep learning, where the system is used to implement the method for automatically extracting geological disaster information from the UAV image based on deep learning according to any one of the implementations of the first aspect.

[035] The present disclosure has the following beneficial effects: the present disclosure provides the method for automatically extracting geological disaster information from the UAV image based on deep learning, such that preprocessing and feature extraction may be performed based on the acquired UAV image, and self-adaptive geological disaster identification and extraction of corresponding geological disaster inspection information are performed based on the trained neural network model. This helps to improve efficiency and accuracy of geological disaster detection, effectively reduces labor costs, and improves an intelligent level of geological disaster detection.

[036] The invention is set out in the appended set of claims. BRIEF DESCRIPTION OF THE DRAWINGS

[037] The present disclosure is further described by using the accompanying drawing, but the embodiments in the accompanying drawing do not constitute any limitation to the present disclosure. For a person of ordinary skill in the art, other accompanying drawings may be further obtained based on following accompanying drawing without creative efforts.

[038] FIG. 1 is a schematic flowchart of a method for automatically extracting geological disaster information from a UAV image based on deep learning according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[039] The technical solutions of the embodiments of the present disclosure are clearly and completely described below with reference to the accompanying drawing. Apparently, the described embodiments are merely some rather than all of the embodiments of the present disclosure. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[040] A method for automatically extracting geological disaster information from a UAV image based on deep learning shown in the embodiment in FIG. 1 includes S1, S2, S3 and S4.

[041] S1: A UAV image is acquired.

[042] In an implementation, S1 includes:

[043] acquiring the UAV image collected and returned by a UAV from a target region in real time.

[044] The method proposed above may be completed based on a server, a smart device, etc. The server receives in real time a remote sensing image collected from the target region (area) and returned by a UAV in real time, and extracts and processes geological disaster information based on the acquired remote sensing image, so that geological disaster information existing in the UAV image may be timely and accurately extracted.

[045] S2: The acquired UAV image is preprocessed to obtain a preprocessed UAV image.

[046] In an implementation, S2 includes S21 and S22.

[047] S21: Projection transformation, radiation correction, image registration, image clipping, etc. are performed on the acquired UAV image to obtain a standard UAV image.

[048] Due to the influence of different factors such as atmospheric refraction and land-surface radiation resolution, the acquired UAV image is easily distorted, resulting in an error in extracting geological disaster information from the UAV image. Therefore, preprocessing such as projection transformation, radiation correction, image registration and image clipping, is first performed on the acquired UAV image, to obtain the standard UAV image.

[049] Through the projection transformation, the UAV image may correspond to a standard coordinate system, which helps to project the UAV image to an appropriate size level. In addition, through the radiation correction, an influence of a ground angle, a velocity, air refraction, ground curvature, etc. during collection of the UAV image may be eliminated, thereby eliminating the distortion of the UAV image. In addition, based on the image registration and the image clipping, confirmation with the target region may be performed based on the acquired UAV image, so as to obtain the standard UAV image corresponding to the target region.

[050] S22: Enhancement processing is performed based on the standard UAV image to obtain the preprocessed UAV image.

[051] The UAV image is prone to be affected by high-altitude illumination, air conditions, etc. during collection, such that the UAV image is easily subjected to brightness deviation or unsharpness, etc., which affects accuracy of geological disaster information extraction. Therefore, after the standard UAV image is obtained, further enhancing the UAV image may help to improve sharpness of the UAV image, highlight feature information therein, and indirectly improve accuracy of subsequent geological disaster feature extraction based on the UAV image.

[052] In an implementation, in S22, performing enhancement processing based on the standard UAV image includes:

[053] converting the standard UAV image from RGB space to gray space based on the standard UAV image, to obtain a first grayscale image, where a gray space conversion function used is: h1(x,y) = 0.299 x r(x,y) + 0.587 x g(x,y) + 0.114 x b(x,y)

[054] where h1(x,y) represents a gray level of a pixel (x,y) in the first grayscale image; r(x,y) represents an R component level of the pixel (x,y) in the standard UAV image, g(x,y) represents a G component level of the pixel (x,y) in the standard UAV image, and b(x,y) represents a B component level of the pixel (x,y) in the standard UAV image;

[055] performing local grayscale adjustment processing based on the obtained first grayscale image to obtain a second grayscale image, where a local grayscale adjustment processing function used is: h2(x,y) hi(x,y) a x [max(r(x,y),g(x,y),b(x,y)) - hi(x, / hi(x,y) -hiT2o' \ 255 —hlT2O . hi(x,y) min 4 I hiT80,hi(x,y) + P x- (max(r(x,y), g(x,y),b(x,y)) - min(r(x,y), g(x,y),b(x, y)))) hi(x,y) >hiT2o hiT80 <hi(x,y) <hiT20 hi(x,y) <hiTOo

[056] where h2(x,y) represents a gray level of the pixel (x,y) in the second grayscale image; h1T20 and h1T80 represent gray levels corresponding to a 20%th pixel and an 80%th pixel in a sequence obtained by ranking gray levels of all pixels in the first grayscale image from large to small, respectively; and max(r(x,y), g(x,y), b(x,y)) represent a maximum value among the R component level, the G component level, and the B component level of the pixel (x,y);

[057] performing global grayscale adjustment processing based on the obtained second grayscale image to obtain a third grayscale image, where a global grayscale adjustment processing function used is: , r x , r X hT — h2med , , f A , hT — h2med h3(x,y) = WiX h25(x,y) x---—---+ w2 x h2(x,y) + w3 x---—----. 255                       max(h2) — mm(h2) x 255

[058] where h3(x,y) represents a gray level of the pixel (x,y) in the third grayscale image, h2g(x,y) represents an average gray level of all pixels in a neighborhood range of the second grayscale image with the pixel (x, y) as a center, h2med represents a median gray level of the second grayscale image, max(h2) and min(h2) represent a maximum gray level and a minimum gray level of the second grayscale image respectively, and hT represents a predetermined standard gray level, where hT has a value range of [150,170]; roi, «2 and «3 represent predetermined weight coefficients respectively, where roi has a value range of [0.2,0.4], ov has a value range of [0.3,0.5], ©3 has a value range of [0.2,0.4], and ro1+ro2+ro3 has a value range of [1,1.1]; and

[059] converting the obtained third grayscale image from the gray space to the RGB space to obtain the preprocessed UAV image.

[060] In view of the problem that the UAV image is prone to be affected by high-altitude illumination, air conditions, etc. during collection, such that the UAV image is easily subjected to brightness deviation or unsharpness, etc., in the foregoing implementation, a technical solution for specifically enhancing the UAV image is proposed. The standard UAV image is first converted from RGB space to gray space. Based on the grayscale image obtained through conversion, local grayscale adjustment for local sharpness adjustment is first performed, an improved local grayscale adjustment processing function is proposed, and self-adaptive equalization adjustment is performed particularly for extra-bright and extra-dark (brightness deviation) pixels in the grayscale image. During local grayscale adjustment, RGB spatial features of pixels are specially considered to control a degree of local grayscale adjustment, which avoids the pixel distortion caused in the local grayscale adjustment process. Global grayscale adjustment is further performed on the grayscale image based on the local grayscale adjustment, which may perform overall stretching and global brightness adjustment on the grayscale image, and effectively improve a contrast of the UAV image, and finally the obtained grayscale image is converted back to the RGB space to obtain the preprocessed UAV image. The enhanced UAV image may effectively improve display of details in the image, highlight detail features, and ensure overall sharpness of the image, avoiding the distortion caused by conventional image processing, and laying a foundation for further geological disaster feature extraction based on the UAV image.

[061] S3: Features is extracted based on the preprocessed UAV image to obtain feature information of the UAV image.

[062] In an implementation, in S3, extracting features based on the preprocessed UAV image includes S31, S32, S33, S34 and S35.

[063] S31: RGB channel separation is performed based on the preprocessed UAV image to obtain red channel component features, green channel component features, and blue channel component features of the UAV image.

[064] S32: Infrared reflectivity features are extracted based on the preprocessed UAV image, to obtain infrared reflectivity features of the UAV image.

[065] S33: Texture features are extracted based on the preprocessed UAV image to obtain texture features of the UAV image.

[066] S34: Vegetation index features are extracted based on the preprocessed UAV image to obtain vegetation index features of the UAV image.

[067] S35: Fusion is performed based on the obtained red channel component features, green channel component features, blue channel component features, infrared reflectivity features, and vegetation index features, to obtain a feature matrix of the UAV image.

[068] In the foregoing implementation, a technical solution for UAV image feature extraction is provided, which may construct a multi-dimensional feature matrix based on RGB channel information, infrared reflection feature information, texture information and normalized vegetation index (NDVI) feature information of the UAV image, so that geological feature information in the UAV image is fed back by using multi-dimensional features, thereby helping to improve the diversity of information extraction and accuracy of geological disaster information extraction.

[069] S4: The feature information of the UAV image is inputted into a trained neural network model to obtain a geological disaster information extraction result.

[070] In an implementation, in S4, inputting the feature information of the UAV image into the trained neural network model to obtain the geological disaster information extraction result includes S41, S42 and S43.

[071] S41: The feature matrix of the UAV image is obtained.

[072] S42: A reference image feature matrix corresponding to the UAV image is obtained.

[073] S43: The feature matrix of the UAV image and the reference image feature matrix as input set are inputted into the trained neural network model to obtain the geological disaster information extraction result output by the neural network model.

[074] In a scenario, for extraction of geological disaster information, it is required to compare data at a current moment with previous data, so as to accurately find the occurrence of geological disasters and extract information. Therefore, when the input set is constructed, the corresponding feature matrix is extracted based on the image information returned by the UAV in real time. In addition, based on a corresponding position of the image information, an image collected at the previous moment (last time) at this position is acquired, and a corresponding feature matrix is extracted in the same mode based on the images collected at the previous moment. The feature matrix of the image returned at the current moment and the feature matrix of the image at the same position at the previous moment are taken as an input set, and the input set is input into the trained neural network model.

[075] In an implementation, the trained neural network model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a first fully connected layer, a second fully connected layer, and a softmax layer which are sequentially connected;

[076] where input of the input layer is the feature matrix of the UAV image and the corresponding reference image feature matrix; the first convolutional layer and the second convolutional layer include 32 convolution kernels with sizes of 3X3 and 5x5, respectively, the pooling layer is maxpooling, with a size of 3x3, the first fully connected layer includes 128 neurons, and the second fully connected layer includes 16 neurons, with an output of a feature vector that reflects whether the UAV image contains geological disaster information and types of geological disasters; and the softmax layer performs classification based on the feature vector output by the second fully connected layer, and outputs the geological disaster information extraction result.

[077] In an implementation, an activation function used by the neural network model is Relu.

[078] Based on the neural network model constructed in the present disclosure, self-adaptive feature extraction may be performed based on the data of the two feature matrices in the input set, and whether a geological disaster occurs may be determined based on the comparison of previous and later feature information. In addition, a type of the geological disaster may be further identified, and the extraction of geological disaster information may be accurately completed.

[079] The geological disasters include cracks, landslides, etc.

[080] In an implementation, the method further includes SB1.

[081] SB1: The neural network model is trained, which includes:

[082] obtaining two groups of UAV images obtained in different time periods at a same position, where the UAV image captured in an earlier time period is taken as a first image, and the UAV image captured in a later time period is taken as a second image; the first image and the second image may be UAV images before and after a geological disaster at the same position, and the geological disasters include cracks and landslides;

[083] extracting features based on the first image and the second image respectively to obtain a feature matrix corresponding to the first image and a feature matrix corresponding to the second image, and constructing a training set by using the feature matrix corresponding to the first image, the feature matrix corresponding to the second image, and correspondingly calibrated geological disaster information extraction marks; and

[084] training the neural network model based on the constructed training set, testing the neural network model by using a test set, and obtaining the trained neural network model when a pass rate of the neural network model reaches a predetermined standard.

[085] Training the neural network model based on the foregoing method may ensure effects and accuracy of the neural network model, and improve efficiency and reliability of geological disaster information extraction based on the UAV image.

[086] The foregoing implementation of the present disclosure provides the method for automatically extracting geological disaster information from the UAV image based on deep learning, so that reprocessing and feature extraction may be performed based on the acquired UAV image, and self-adaptive geological disaster identification and extraction of corresponding geological disaster inspection information are performed based on the trained neural network model. This helps to improve efficiency and accuracy of geological disaster detection, effectively reduces labor costs, and improves an intelligent level of geological disaster detection.

[087] Based on the method for automatically extracting geological disaster information from the UAV image based on deep learning shown in FIG. 1, the present disclosure further provides a system for automatically extracting geological disaster information from the UAV image based on deep learning, where the system is used to implement the method for automatically extracting geological disaster information from the UAV image based on deep learning shown in FIG. 1 and specific embodiments corresponding to the steps of the method, and description thereof is not provided herein again in the present application.

[088] It should be noted that functional units / modules in the embodiments of the present disclosure may be integrated into one processing unit / module, or each of the units / modules may exist alone physically, or two or more units / modules may be integrated into one unit / module. The foregoing integrated unit / module may be implemented either in a form of hardware or in a form of a software functional unit / module.

[089] From the foregoing description of the implementations, a person skilled in the art may clearly understand that the embodiments described herein may be implemented in hardware, software, firmware, middleware, code or any suitable combination thereof. For hardware implementation, a processor may be implemented in one or more of the following units: an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, another electronic unit designed to implement the functions described herein, or a combination thereof. For software implementation, some or all of processes of the embodiments may be completed by a computer program instructing related hardware. During implementation, the foregoing program may be stored in a computer-readable medium or transmitted as one or more instructions or code on the computer-readable medium. The computer-readable medium includes a computer storage medium and a communication medium, where the communication medium includes any medium that facilitates transfer of a computer program from one place to another. The storage medium may be any available medium accessible by a computer. The computer-readable medium may include, but is not limited to, a random access memory (RAM), a read only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or another optical storage, a magnetic disk storage medium or another magnetic storage device, or any other medium that may be used to carry or store desired program code in the form of an instruction or a data structure and that may be accessed by a computer.

[090] Finally, it should be noted that the foregoing embodiments are provided merely to describe the technical solutions of the present disclosure, rather than to limit the protection scope of the present disclosure. Although the present disclosure is described in detail with reference to preferred embodiments, a person of ordinary skill in the art should understand that modifications or equivalent replacements may be made to the technical solutions of the present disclosure without departing from the essence and scope of the technical solutions of the present disclosure.

Claims

1. A method for automatically extracting geological disaster information from an unmanned aerial vehicle (UAV) image based on deep learning, comprising:S1: acquiring the UAV image;S2: preprocessing the acquired UAV image to obtain a preprocessed UAV image;S3: extracting features based on the preprocessed UAV image to obtain feature information of the UAV image; andS4: inputting the feature information of the UAV image into a trained neural network model to obtain a geological disaster information extraction result;wherein S1 comprises: acquiring, the UAV image collected and returned by a UAV from a target region in real time;wherein S2 comprises:S21: performing projection transformation, radiation correction, image registration and image clipping on the acquired UAV image to obtain a standard UAV image; andS22: performing enhancement processing based on the standard UAV image to obtain the preprocessed UAV image; andwherein in S22, performing enhancement processing based on the standard UAV image comprises:converting the standard UAV image from RGB space to gray space based on the standard UAV image, to obtain a first grayscale image, wherein a gray space conversion function used is:hi(x,y) = 0.299 x r(x,y) + 0.587 x g(x,y) + 0.114 x b(x,y)wherein h1(x,y) represents a gray level of a pixel (x,y) in the first grayscale image; r(x,y) represents an R component level of the pixel (x,y) in the standard UAV image, g(x,y) represents a G component level of the pixel (x,y) in the standard UAV image, and b(x,y) represents a B component level of the pixel (x,y) in the standard UAV image;performing local grayscale adjustment processing on the obtained first grayscale image to obtain a second grayscale image, wherein a local grayscale adjustment processing function used is:h2(x,y) / hi(x,y) -ax [max(r(x,y),g(x,y),b(x,y)) - hi(x,y)] x I'hi(x,y) -hiT20255 -hiT20hi(x,y)min|hiT8o,hi(x,y) + 0 x |(max(r(x,y),g(x,y),b(x,y)) - mm(r(x,y), gfoy^bfoy))})hi(x,y) >hiT2ohiT80 <hi(x,y) <hlT20hi(x,y) <hiTsowherein h2(x,y) represents a gray level of the pixel (x,y) in the second grayscale image;h1T20 and h1T80 represent gray levels corresponding to a 20%th pixel and an 80%th pixel in a sequence obtained by ranking gray levels of all pixels in the first grayscale image from large to small, respectively; and max(r(x,y), g(x,y), b(x,y)) represents a maximum value among the R component level, the G component level, and the B component level of the pixel (x,y);performing global grayscale adjustment processing based on the obtained second grayscale image to obtain a third grayscale image, wherein a global grayscale adjustment processing function used is:i / a .,1 r a ., hT - h2med , .,, / a . ., hT-h2med ., nrrh3(x,y) = «i X h25(x,y) x---—----+ w2 x h2(x,y) + w3 x---—----.     x 255255                        max(h2) — min(h2)wherein h3(x,y) represents a gray level of the pixel (x,y) in the third grayscale image, h2s(x,y) represents an average gray level of all pixels in a neighborhood range of the second grayscale image with the pixel (x, y) as a center, h2med represents a median gray level of the second grayscale image, max(h2) and min(h2) represent a maximum gray level and a minimum gray level of the second grayscale image respectively, and hT represents a predetermined standard gray level, wherein hT has a value range of [150,170]; ®i, ®2 and ®3 represent predetermined weight coefficients respectively, wherein roi has a value range of [0.2,0.4], «2 has a value range of [0.3,0.5], «3 has a value range of [0.2,0.4], and ®1+®2+®3 has a value range of [1,1.1]; andconverting the obtained third grayscale image from the gray space to the RGB space to obtain the preprocessed UAV image.

2. The method for automatically extracting geological disaster information from the UAV image based on deep learning according to claim 1, wherein S3 comprises:S31: performing red green blue (RGB) channel separation based on the preprocessed UAV image to obtain red channel component features, green channel component features, and blue channel component features of the UAV image;S32: extracting infrared reflectivity features based on the preprocessed UAV image, to obtain infrared reflectivity features of the UAV image;S33: extracting texture features based on the preprocessed UAV image to obtain texture features of the UAV image;S34: extracting vegetation index features based on the preprocessed UAV image to obtain vegetation index features of the UAV image; andS35: performing fusion based on the obtained red channel component features, green channel component features, blue channel component features, infrared reflectivity features, and vegetation index features, to obtain a feature matrix of the UAV image.

3. The method for automatically extracting geological disaster information from the UAV imagebased on deep learning according to claim 2, wherein S4 comprises:S41: obtaining the feature matrix of the UAV image;S42: obtaining a reference image feature matrix corresponding to the UAV image; andS43: inputting the feature matrix of the UAV image and the reference image feature matrix as input set into the trained neural network model to obtain the geological disaster information extraction result output by the neural network model.

4. The method for automatically extracting geological disaster information from the UAV image based on deep learning according to claim 3, wherein the trained neural network model comprises an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a first fully connected layer, a second fully connected layer, and a softmax layer which are sequentially connected;wherein input of the input layer is the feature matrix of the UAV image and the corresponding reference image feature matrix; the first convolutional layer and the second convolutional layer each comprise 32 convolution kernels with sizes of 3X3 and 5X5 respectively, the pooling layer is maxpooling, with a size of 3X3, the first fully connected layer comprises 128 neurons, and the second fully connected layer comprises 16 neurons, with an output of a feature vector that reflects whether the UAV image contains geological disaster information and types of geological disasters; and the softmax layer performs classification based on the feature vector output by the second fully connected layer, and outputs the geological disaster information extraction result.

5. The method for automatically extracting geological disaster information from the UAV image based on deep learning according to claim 4, wherein the method further comprises:SB1: training the neural network model, comprising:obtaining two groups of UAV images obtained in different time periods at a same position, wherein the UAV image captured in an earlier time period is taken as a first image, and the UAV image captured in a later time period is taken as a second image; the first image and the second image are UAV images before and after a geological disaster at the same position, and the geological disasters comprise cracks and landslides;extracting features based on the first image and the second image respectively to obtain a feature matrix corresponding to the first image and a feature matrix corresponding to the second image, and constructing a training set by using the feature matrix corresponding to the first image, the feature matrix corresponding to the second image, and correspondingly-calibrated geological disaster information extraction marks; andtraining the neural network model based on the constructed training set, testing the neuralnetwork model by using a test set, and obtaining the trained neural network model when a pass rate of the neural network model reaches a predetermined standard.

6. A system for automatically extracting geological disaster information from a UAV image based on deep learning, wherein the system is used to implement the method for automatically extracting geological disaster information from the UAV image based on deep learning according to any one of the claims 1 to 5.

Images