A riverbed bedrock identification method, system, device and application based on unmanned aerial vehicle multispectral data

Abstract

The present invention belongs to the technical field of remote sensing identification of geological resources, and discloses a riverbed bedrock identification method, system, equipment and application based on multispectral data of unmanned aerial vehicles, including obtaining high-resolution multispectral remote sensing images of unmanned aerial vehicles and performing image preprocessing; obtaining typical blocks of riverbed bedrock as training samples, performing U‑Net deep learning, and determining the distribution area of ​​riverbed bedrock and bedrock debris deposits; performing image masking on the distribution range of riverbed bedrock and bedrock debris deposits; performing MNF transformation on the masked parts of the prominent riverbed bedrock and bedrock debris deposits, and generating false color images; selecting riverbed bedrock and bedrock debris deposits as training samples for improved MNF‑U‑Net deep learning; and distinguishing riverbed bedrock from bedrock debris deposits. The data accuracy of the present invention is significantly improved, and at the same time, it can reduce the survey cost and can efficiently, quickly and accurately identify the distribution range of riverbed bedrock in a relatively short time.

Description

Technical Field

[0001] The present invention belongs to the technical field of geological body remote sensing identification, and specifically relates to a riverbed bedrock identification method, system, equipment and application based on unmanned aerial vehicle multispectral data. Background Art

[0002] The comprehensive management of small watersheds is to establish a comprehensive management model of small watersheds in semi-mountainous areas with soil and water conservation and efficient ecological and economic functions based on the natural and socio-economic conditions of small watersheds and the requirements of regional national economic development, with the management of soil and water loss in small watersheds as the center, the improvement of ecological and economic benefits and sustainable socio-economic development as the goal, and the optimization of basic farmland structure and efficient utilization and vegetation construction as the focus. In natural river channels, according to the different surface characteristics of the riverbed, it can be divided into bedrock exposed riverbed and sand and pebble covered riverbed. In the past, it was generally believed that in river channels with steep slopes, narrow and deep, the sediment transport capacity is strong, and it is easier to form bedrock exposed riverbeds. For river channels with gentle slopes, wide and shallow, the sediment transport capacity is weak, and continuous sand from the upstream will cause siltation of the riverbed, which is more likely to form sand and pebble covered riverbeds. The bedrock exposure and sand and pebble coverage in the riverbed of mountain rivers interact and influence each other. When the boundary conditions of the riverbed, the upstream sand and water flow conditions change, the surface morphology of the riverbed also adjusts accordingly. In terms of flood evolution, riverbed evolution, aquatic ecology, water conservancy project construction and landform evolution, the two types of riverbeds show significantly different characteristics.

[0003] Basin geological survey is the premise and basis for basin regulation. In recent years, with the continuous development and maturity of UAV technology, a number of small-scale UAV aerial survey equipment have emerged, which can efficiently and quickly obtain multi-spectral remote sensing image data within a small basin. It provides a means of quickly surveying strata and lithology for basin geological survey work in the early stage of basin regulation. At present, the highest resolution satellite remote sensing image resolution is 20cm, with only 4 bands in the range of 400-900nm, and the data acquisition cost is high, the accuracy is low, and the accuracy of image survey is seriously affected.

[0004] Through the above analysis, the problems and defects of the existing satellite-borne lithology identification technology are as follows:

[0005] (1) The highest spatial resolution of currently used satellite-borne high-resolution remote sensing data is 20 cm, which is lower than that of drone remote sensing data. The signal-to-noise ratio is poor, and useful information is mixed with a large amount of useless information, which seriously affects the accuracy of image surveys.

[0006] (2) Since the riverbed bedrock and bedrock debris deposits have the same material composition and are similar in image, it is difficult to distinguish the two using traditional image processing methods. However, these two types of landforms belong to different lithologies in geology, and current survey methods do not provide an effective processing method specifically for identifying riverbed bedrock in small watersheds.

[0007] The significance of solving the above problems and defects is:

[0008] (1) The use of high-precision UAV remote sensing data can obtain accurate information on the morphology of the riverbed bedrock, which is conducive to a comprehensive and accurate investigation and research of the riverbed bedrock.

[0009] (2) Through the training of deep learning models, the identification of riverbed bedrock and bedrock debris deposits can be automatically realized, greatly improving the efficiency and accuracy of geological exploration work in the basin.

[0010] (3) In the study of riverbed bedrock identification, a special processing method is proposed for riverbed bedrock identification in small watersheds to meet the survey needs. Summary of the invention

[0011] The present invention provides a riverbed bedrock identification method, system, equipment and application based on unmanned aerial vehicle multispectral data, which solves the problems of low data accuracy, difficulty in distinguishing riverbed bedrock and bedrock debris deposits with similar images, high survey cost and low efficiency in the prior art.

[0012] The technical solution adopted by the present invention is a riverbed bedrock identification method based on multispectral data of unmanned aerial vehicles, comprising the following steps:

[0013] Acquire high-resolution multispectral remote sensing images from drones;

[0014] Image processing was performed on high-resolution multispectral remote sensing images to obtain typical blocks of riverbed bedrock as training samples, and U-Net deep learning was performed to determine the distribution area of ​​riverbed bedrock and bedrock debris deposits;

[0015] According to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, the processed high-resolution image is masked to obtain an image highlighting the riverbed bedrock and bedrock debris accumulation;

[0016] The mask part of the prominent riverbed bedrock and bedrock debris accumulation is transformed by MNF, and a false color image is generated;

[0017] The riverbed bedrock and bedrock debris accumulation were selected as training samples on the MNF transformed false color images for improved MNF-U-Net deep learning;

[0018] The model obtained by deep learning of the improved MNF-U-Net is used to classify the MNF transformed images and distinguish the riverbed bedrock and bedrock debris deposits.

[0019] Furthermore, the determination of the distribution area of ​​the riverbed bedrock and bedrock debris deposits specifically includes:

[0020] The B5, B4, B3, B2, and B1 bands of the high-resolution multispectral images collected by the UAV are combined to obtain a combined image of the B5, B4, B3, B2, and B1 bands of the images;

[0021] The five-band combined image is processed for decorrelation analysis to obtain a processed high-resolution image;

[0022] Based on the fact that riverbed bedrock and bedrock debris deposits have similar tones in high-resolution images, for example, the bedrock outcrops and rubble deposits of Tertiary sandstone are both yellow-brown, some bedrock outcrops and rubble deposits are circled as training samples;

[0023] Based on the training samples, the image is subjected to improved U-Net deep learning to identify the distribution range of bedrock outcrops and rubble deposits in the entire image.

[0024] Furthermore, obtaining typical blocks of riverbed bedrock as training samples specifically includes:

[0025] Perform natural color synthesis of RGB bands on high-resolution drone images to obtain natural true color images;

[0026] The exposed riverbed bedrock has a high light reflectivity, so it appears as a bright yellow-brown hue in the natural true color image, and the hue is different from the grayish white of the concrete floor and road surface. The bedrock debris deposits appear in a similar but lower hue to the riverbed bedrock in the natural true color image.

[0027] According to the image characteristics of the riverbed bedrock and bedrock debris deposits, the region of interest (ROI) tool was used to define the sample range.

[0028] Furthermore, the improved MNF-U-Net deep learning is performed by selecting riverbed bedrock and bedrock debris deposits as training samples on the MNF transformed false color image, specifically including:

[0029] First, the first U-Net algorithm learning was performed on samples of riverbed bedrock and bedrock debris accumulation, and the distribution range of riverbed bedrock and bedrock debris accumulation was delineated on high-resolution multispectral images of drones.

[0030] Then, the distribution range of the riverbed bedrock and bedrock debris deposits is masked, and the masked area is subjected to MNF transformation to enhance the difference between the riverbed bedrock outcrop and the bedrock debris deposits in the image.

[0031] Samples of riverbed bedrock and bedrock debris deposits are input, and a second U-Net algorithm learning is performed in the mask area to distinguish the riverbed bedrock outcrops and bedrock debris deposits in the mask area.

[0032] Further, the mask portion of the protruding riverbed bedrock and bedrock debris deposits is subjected to MNF transformation to generate a false color image, specifically including:

[0033] According to the distribution area of ​​the identified riverbed bedrock and bedrock debris accumulation, mask processing is performed on the image of the identified area;

[0034] The masked area is processed by MNF transformation, and the first three MNF transformation bands are synthesized to output a false color image that enhances the difference between the riverbed bedrock and the bedrock debris deposits.

[0035] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0036] Acquire high-resolution multispectral remote sensing images from drones;

[0037] Perform image processing on high-resolution multispectral remote sensing images to obtain typical blocks of riverbed bedrock as training samples, conduct deep learning, and determine the distribution area of ​​riverbed bedrock and bedrock debris accumulation;

[0038] According to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, the processed high-resolution image is masked to obtain an image highlighting the riverbed bedrock and bedrock debris accumulation;

[0039] The mask part of the prominent riverbed bedrock and bedrock debris accumulation is transformed by MNF, and a false color image is generated;

[0040] The riverbed bedrock and bedrock debris deposits were selected as training samples on the MNF transformed false color images for improved MNF-U-Net deep learning;

[0041] The model obtained by deep learning of the improved MNF-U-Net is used to classify the MNF transformed images and distinguish the riverbed bedrock and bedrock debris deposits.

[0042] Another object of the present invention is to provide a riverbed bedrock and bedrock debris accumulation identification system based on drone multispectral data for implementing the riverbed bedrock identification method based on drone multispectral data, the riverbed bedrock identification system based on drone multispectral data comprising:

[0043] Image acquisition module, a UAV remote sensing data acquisition system used to acquire high-resolution multispectral UAV remote sensing images. The acquired multispectral remote sensing data is 5-band remote sensing image data;

[0044] The MNF-U-Net riverbed bedrock identification module is used to distinguish riverbed bedrock from bedrock debris deposits and requires three submodules;

[0045] The riverbed bedrock and bedrock debris accumulation identification submodule uses the U-Net algorithm to combine and identify the riverbed bedrock and bedrock debris accumulation;

[0046] The riverbed bedrock and bedrock debris accumulation mask enhancement submodule is used to perform masking according to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and perform MNF transformation processing on the image of the flooded area to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation;

[0047] The riverbed bedrock and bedrock debris deposits differentiation submodule is used to distinguish the riverbed bedrock and bedrock debris deposits in the masked area using the U-Net algorithm.

[0048] Furthermore, the three submodules specifically include:

[0049] The riverbed bedrock and bedrock debris accumulation identification submodule uses the U-Net algorithm to combine and identify the riverbed bedrock and bedrock debris accumulation;

[0050] The riverbed bedrock and bedrock debris accumulation mask enhancement submodule is used to perform masking according to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and perform MNF transformation processing on the image of the flooded area to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation;

[0051] The riverbed bedrock and bedrock debris deposits differentiation submodule is used to distinguish the riverbed bedrock and bedrock debris deposits in the masked area using the U-Net algorithm.

[0052] Furthermore, the image acquisition module specifically includes:

[0053] Multi-rotor drone platform and 5-band multispectral camera sensor;

[0054] The high-resolution multispectral images of UAVs are stitched and corrected to obtain a 5-band multispectral image with pixel values ​​as surface reflectance values.

[0055] Another object of the present invention is to provide a method for identifying riverbed bedrock in a small watershed, wherein the method for identifying riverbed bedrock in a small watershed uses the method for identifying riverbed bedrock based on drone multispectral data.

[0056] The beneficial effects of the present invention are:

[0057] 1. Obtain high-resolution multispectral images from drones. High-resolution multispectral images from drones have high geometric resolution, and the surface objects are clearer, so the geometric morphological analysis is more accurate. Compared with high-resolution CCD images from drones, they have reflectance data in 5 bands, which can reflect the physical and chemical properties of surface materials more accurately. The combination of high-resolution and multispectral technologies can give full play to their respective advantages in geometric resolution and spectral performance.

[0058] 2. Perform MNF transformation operation on the distribution area of ​​riverbed bedrock and bedrock debris deposits identified by the high-resolution multispectral image of the UAV, and output the MNF123 band false color image. The difference between riverbed bedrock and bedrock debris deposits can be enhanced and displayed on the MNF false color image, which is the key mark for the present invention to finally distinguish riverbed bedrock and bedrock debris deposits.

[0059] 3. The present invention is a remote sensing identification method specifically proposed for bedrock and bedrock debris deposits in small watersheds, which can be applied to geological survey and mapping of small watersheds in river regulation work in small watersheds. The heavy exploration technology cost in these areas is very high. Therefore, the present invention is extremely important to reduce the cost of geological survey and mapping of watersheds in such areas.

[0060] 4. The method provided by the present invention obtains the combination of color tone and texture characteristics of riverbed bedrock and bedrock debris deposits, and finds that the combination of color tone and texture has the characteristics of distinguishing riverbed bedrock and bedrock debris deposits. It solves the difficult problem that riverbed bedrock and bedrock debris deposits have the same material composition and similar spectral characteristics, and traditional image processing methods are difficult to effectively distinguish.

[0061] 5. The high-resolution multispectral data obtained by the UAV of the present invention is pre-processed by data collection, splicing, reflectance image output, etc., so as to facilitate the delineation of the riverbed bedrock and bedrock debris deposit area, and according to the distribution area of ​​the riverbed bedrock and bedrock debris deposit, the masked UAV high-resolution multispectral image is subjected to MNF transformation to obtain an MNF false color image of the enhanced riverbed bedrock and bedrock debris deposit area; the riverbed bedrock and bedrock debris deposit samples are screened for the obtained MNF false color image of the enhanced riverbed bedrock and bedrock debris deposit area; collecting the characteristic samples of the riverbed bedrock and bedrock debris deposit in the MNF false color image is an important link, and the key is to select the riverbed bedrock outcrop with relatively smooth surface and the surface texture Rich samples of bedrock debris deposits, two samples are input for U-Net deep learning on a small range of images, and the recognition model of riverbed bedrock and bedrock debris deposits is obtained. Finally, the riverbed bedrock and bedrock debris deposits are divided out through the model; based on the first U-Net deep learning, the distribution area of ​​riverbed bedrock and bedrock debris deposits is delineated, and the distribution area of ​​riverbed bedrock and bedrock debris deposits is masked, and then MNF change processing is performed in the masked area to enhance the difference between riverbed bedrock and bedrock debris deposits in MNF change images, and finally U-Net deep learning classification is performed for the second time, and the boundary of riverbed bedrock and bedrock debris deposits is efficiently, quickly and accurately determined in a short time. The present invention can efficiently, quickly and accurately identify the distribution range of riverbed bedrock in small watersheds in the lithology mapping of small watershed river channel regulation work.

[0062] 6. The present invention provides a method and system for identifying riverbed bedrock in a small watershed based on UAV remote sensing technology, which can efficiently, quickly and accurately identify the distribution range of riverbed bedrock in a relatively short time. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A flow chart of a riverbed rock identification method based on drone multispectral data provided in an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the ordinary U-Net image recognition results in the Weihe River Basin in Chang'an District, Xi'an City.

[0065] Figure 3 It is a schematic diagram of the MNF-U-Net image recognition results of the Weihe River Basin in Chang'an District, Xi'an City provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention clearer and more understandable, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0067] like Figure 1As shown, the present invention provides a riverbed bedrock identification method based on drone multispectral data, which includes the following steps:

[0068] S101: Acquire high-resolution multispectral images from drones, with 5 bands;

[0069] S102: performing preprocessing such as orthorectification, stitching, and color uniformity on the high-resolution multispectral image of the UAV;

[0070] S103: Select local typical riverbed bedrock and bedrock debris accumulation as training samples on the preprocessed UAV multispectral image, use U-Net algorithm to perform deep learning, and identify the distribution range of riverbed bedrock and bedrock debris accumulation;

[0071] S104: masking the identified riverbed bedrock and the distribution range of bedrock debris deposits, performing MNF transformation on the masked area, and generating a false color image;

[0072] S105: Select part of the riverbed bedrock and bedrock debris deposits as training samples on the MNF transformed false color image, use the improved MNF-U-Net algorithm for deep learning, and finally distinguish the riverbed bedrock and bedrock debris deposits.

[0073] In step S103, the U-Net algorithm model specifically includes:

[0074] (1) Open the training area image containing the sample and create a deep learning raster;

[0075] (2) Conduct deep learning model training and output recognition model;

[0076] (3) Open the image of the identification area, input the identification model, perform image classification, and output the distribution area of ​​the riverbed bedrock and bedrock debris deposits. At this time, the riverbed bedrock and bedrock debris deposits are classified into one category.

[0077] The U-Net network is a CNN-based image segmentation network proposed in 2015. The network was originally proposed for cell wall segmentation, and later it has performed well in lung nodule monitoring and retinal blood vessel extraction. Due to its good results, it has been widely used in various segmentation tasks, and many U-net-based segmentation models have been derived. It is based on an encoder-decoder structure and uses long-distance skip connections combined with details from the bottom layer, which effectively compensates for the loss of spatial information during the downsampling operation and helps the network restore more accurate positioning, which is crucial for intensive segmentation tasks such as medical image segmentation, remote sensing image segmentation, and cutouts that attach great importance to details.

[0078] In step S103, local typical riverbed bedrock and bedrock debris deposits are selected as training samples, specifically including:

[0079] (1) Perform natural color synthesis of RGB bands on high-resolution drone images to obtain natural true color images;

[0080] (2) The exposed riverbed bedrock has a high light reflectivity, so it appears as a bright yellow-brown hue in natural true color images, and the hue is different from the grayish white of concrete floors and pavement. The bedrock debris deposits appear in a similar but lower hue to the riverbed bedrock in natural true color images;

[0081] (3) Based on the image characteristics of the riverbed bedrock and bedrock debris deposits, the region of interest (ROI) tool was used to define the sample range.

[0082] In step S103, the distribution range of the riverbed bedrock and bedrock debris deposits is identified, which specifically includes:

[0083] The B5, B4, B3, B2, and B1 bands of the high-resolution multispectral images collected by the UAV are combined to obtain a combined image of the B5, B4, B3, B2, and B1 bands of the images;

[0084] The five-band combined image is processed for decorrelation analysis to obtain a processed high-resolution image;

[0085] Based on the fact that riverbed bedrock and bedrock debris deposits have similar tones in high-resolution images, for example, the bedrock outcrops and rubble deposits of Tertiary sandstone are both yellow-brown, some bedrock outcrops and rubble deposits are circled as training samples;

[0086] Based on the training samples, the image is subjected to improved U-Net deep learning to identify the distribution range of bedrock outcrops and rubble deposits in the entire image.

[0087] In step S104, MNF transformation specifically includes:

[0088] (1) Based on the distribution area of ​​the identified riverbed bedrock and bedrock debris deposits, mask processing is performed on the image of the identified area;

[0089] (2) Perform MNF transformation on the masked area, synthesize the first three MNF transformation bands, and output a false color image that enhances the difference between the riverbed bedrock and the bedrock debris deposits.

[0090] The Minimum Noise Fraction Rotation (MNF Rotation) is used to determine the intrinsic dimensionality (i.e., number of bands) of the image data, to separate the noise in the data, and to reduce the computational requirements for subsequent processing. MNF is essentially a two-step stacked principal component transformation. The first transformation (based on an estimated noise covariance matrix) is used to separate and rescale the noise in the data, which results in the transformed noisy data having minimal variance and no inter-band correlation. The second step is a standard principal component transformation of the noise-whitened data. For further spectral processing, the intrinsic dimensionality of the data is determined by examining the final eigenvalues ​​and correlation images. The data space can be divided into two parts: one part is associated with larger eigenvalues ​​and corresponding eigenimages, and the rest is associated with images with approximately the same eigenvalues ​​and noise dominating.

[0091] In step S104, mask processing is performed on the identified distribution range of the riverbed bedrock and bedrock debris deposits, MNF transformation is performed on the mask area, and a false color image is generated, which specifically includes:

[0092] (1) Based on the distribution area of ​​the identified riverbed bedrock and bedrock debris deposits, mask processing is performed on the image of the identified area;

[0093] (2) Perform MNF transformation on the masked area, synthesize the first three MNF transformation bands, and output a false color image that enhances the difference between the riverbed bedrock and the bedrock debris deposits.

[0094] In step S105, part of the riverbed bedrock and bedrock debris deposits are selected as training samples on the MNF transformed false color image, and the improved MNF-U-Net algorithm is used for deep learning to finally distinguish the riverbed bedrock and bedrock debris deposits, specifically including:

[0095] (1) First, the first U-Net algorithm learning was performed on the samples of riverbed bedrock and bedrock debris accumulation, and the distribution range of riverbed bedrock and bedrock debris accumulation was delineated on the high-resolution multispectral images of UAVs;

[0096] (2) Then, the distribution range of the riverbed bedrock and bedrock debris deposits is masked, and the masked area is subjected to MNF transformation to enhance the difference between the riverbed bedrock outcrop and the bedrock debris deposits in the image;

[0097] (3) Samples of riverbed bedrock and bedrock debris deposits are input, and a second U-Net algorithm learning is performed in the mask area to distinguish the riverbed bedrock outcrops and bedrock debris deposits in the mask area.

[0098] The method for identifying riverbed rock based on multispectral data of unmanned aerial vehicle provided by the present invention can also be implemented by ordinary technicians in the industry using other steps. Figure 1 The riverbed bedrock identification method based on multispectral data of unmanned aerial vehicles provided by the present invention is only a specific embodiment. The remote sensing data of the present invention adopts the Pegasus D2000 system, and other unmanned aerial vehicle multispectral remote sensing systems can also be adopted.

[0099] The present invention uses remote sensing image processing software ENVI, ERDAS, PCI and the like as a digital image processing platform, adopts an image processing scheme that is conducive to riverbed bedrock identification, and processes UAV high-resolution multispectral images.

[0100] According to the characteristics of riverbed bedrock in high-resolution multispectral images of drones, typical riverbed bedrock and bedrock debris deposits were selected as training samples to facilitate riverbed bedrock identification. The riverbed bedrock and bedrock debris deposits initially identified in the high-resolution multispectral images of drones were masked, and the masked areas were transformed by MNF to highlight the contrast between the riverbed bedrock and bedrock debris deposits.

[0101] The UAV high-resolution multispectral remote sensing data acquisition platform consists of a UAV flight system and a multispectral data acquisition system. The imaging bands of the multispectral acquisition system are located in the visible light and near-infrared bands. The first to third bands of the UAV high-resolution multispectral remote sensing images are located in the visible light part; the fourth band is located in the red edge part; the fifth band is located in the near-infrared part; the geometric resolution of the five bands can be adjusted in the range of 5-10cm according to the flight altitude. See Table 1 for details.

[0102] Table 1 Parameters of UAV high-resolution multispectral remote sensing images

[0103] Band Name Central wavelength (nm) Band width (nm) blue 475 20 green 560 20 red 668 10 Red Edge 717 40 Near Infrared 840 10

[0104] The present invention provides a riverbed rock identification system based on multispectral data of unmanned aerial vehicles, comprising:

[0105] Image acquisition module, a UAV remote sensing data acquisition system used to acquire high-resolution multispectral UAV remote sensing images. The acquired multispectral remote sensing data is 5-band remote sensing image data;

[0106] The MNF-U-Net riverbed bedrock identification module is used to distinguish riverbed bedrock from bedrock debris deposits. It requires three submodules:

[0107] The riverbed bedrock and bedrock debris accumulation identification submodule uses the U-Net algorithm to combine and identify the riverbed bedrock and bedrock debris accumulation;

[0108] The riverbed bedrock and bedrock debris accumulation mask enhancement submodule is used to perform masking according to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and perform MNF transformation processing on the image of the flooded area to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation;

[0109] The riverbed bedrock and bedrock debris deposits differentiation submodule is used to distinguish the riverbed bedrock and bedrock debris deposits in the masked area using the U-Net algorithm.

[0110] The image acquisition module specifically includes:

[0111] (1) Multi-rotor UAV platform and 5-band multispectral camera sensor;

[0112] (2) The high-resolution multispectral images of UAVs are stitched and corrected to obtain a 5-band multispectral image with pixel values ​​as surface reflectance values.

[0113] Among them, the three sub-modules of the MNF-U-Net riverbed rock recognition module include:

[0114] (1) The riverbed bedrock and bedrock debris accumulation identification submodule uses the U-Net algorithm to combine and identify the riverbed bedrock and bedrock debris accumulation;

[0115] (2) The riverbed bedrock and bedrock debris accumulation mask enhancement submodule is used to perform masking according to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and perform MNF transformation processing on the image of the flooded area to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation;

[0116] (3) The riverbed bedrock and bedrock debris deposits differentiation submodule is used to distinguish the riverbed bedrock and bedrock debris deposits in the masked area using the U-Net algorithm.

[0117] The technical effects of the present invention are described in detail below in conjunction with experiments.

[0118] This experiment uses the middle and upper reaches of the Fuhe River in Chang'an District, Xi'an as the experimental area. It originates from the northern foot of the Qinling Mountains and its source is the Dayu River. After leaving the valley, it flows through Yinzhen, Wangmang Street, Duqu, Weiqu, Guodu Town, Xinglong Street, Luanzhen, Taiyi Palace Street, Ziwu Town, Wutai, Huangliang Street, Wangqu Street and other street offices and towns, and merges into the Fenghe River at the head of the Fenghui Canal in Qinzhen. From the source of the river, Dayu River, to the entrance of Fengkou, the river is 64.2 kilometers long, with a basin area of ​​687 square kilometers, an average gradient of I = 0.0063, and an average annual runoff of 210.5 million cubic meters. The region exposes yellow-brown to brown-red siltstone and mudstone of the Upper Tertiary Miocene, plastic, saturated and Quaternary clastic sediments. It is widely exposed on both sides of the Fuhe River and the river section from Xiahongmiao Village to Duqu Town.

[0119] The collected high-resolution multispectral images were preprocessed and displayed according to the B321 band. The riverbed bedrock and bedrock debris deposits were yellow-brown to brown-red in color on the high-resolution multispectral images of the drone and were distributed on both sides of the river. Based on the above interpretation signs, the local typical riverbed bedrock and bedrock debris deposits on both sides of the river were interpreted as training samples. Figure 3 The gray area is circled by the black line in the middle. U-Net deep learning is performed on the small-scale images of the training samples and the coverage samples to establish a recognition model, and the recognition model is used to classify the images of the experimental area to obtain the distribution area of ​​the riverbed bedrock and bedrock debris accumulation; the high-resolution multispectral image of the drone is masked according to the distribution area of ​​the riverbed bedrock and bedrock debris accumulation, and the mask area is transformed by MNF to enhance the difference between the riverbed bedrock and bedrock debris accumulation in the image; some typical riverbed bedrock and bedrock debris accumulation distribution areas are selected on the MNF transformed image, and U-Net deep learning is performed again to establish a recognition model to distinguish the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and model recognition is performed to distinguish the riverbed bedrock and bedrock debris accumulation distribution areas in the image.

[0120] The spatial resolution of high-resolution multispectral images from drones can reach 5cm, while the exposed width of the riverbed bedrock ranges from several meters to tens of meters. After processing, the multispectral data can more clearly reflect the texture difference between the riverbed bedrock and the bedrock debris deposits. Therefore, the technical solution provided by the present invention is to identify the riverbed bedrock information and detect the distribution area of ​​the riverbed bedrock and bedrock debris deposits based on the high-resolution multispectral images from drones. It not only avoids the problems of low spatial resolution, inaccurate abnormal positions, mixed noise, and serious image extraction accuracy caused by satellite remote sensing data alone, but also realizes the rapid mapping of small watershed geological surveys, improves work efficiency, and saves funds invested by the state and society. Therefore, the method or system provided by the present invention can efficiently, quickly, and accurately identify the riverbed bedrock of a small watershed in a relatively short period of time, and has not been proposed by predecessors. The technology of the present invention can fill the gaps in related fields and has certain significance for guiding the geological mapping of small watersheds. The specific effects of the present invention also include Table 2, Table 3, Figure 2 , Figure 3 Content shown.

[0121] Table 2 Confusion matrix of object category extraction accuracy using single U-Net learning

[0122] category vegetation architecture Sediment the way water surface Riverbed rock Debris accumulation total Uncategorized 0 0 0 0.03 0.08 0 0 0.01 vegetation 95.05 0.23 0.23 0.05 0.06 0.15 0.77 76.91 architecture 0.62 26.8 0.02 3.05 0.11 0 0.62 1.06 Sedimentary layer 0.01 0.77 80.76 1.5 0.04 7.39 15.23 1.48 the way 0.22 67.05 3.79 90 3.1 0.13 7.38 3.73 water surface 1.12 0.08 0.04 0.13 96.57 0.02 0 7.91 Riverbed rock 2.02 1.46 7.21 0.48 0.02 64.48 12.55 3.06 Debris accumulation 0.95 3.61 7.94 4.76 0.01 27.83 63.45 5.83 total 100 100 100 100 100 100 100 100

[0123] Overall accuracy of riverbed bedrock and debris accumulation: 63.97%.

[0124] Table 3 Confusion matrix of the accuracy of extracting riverbed bedrock and debris deposits by the MNF-U-Net method of the present invention

[0125] category Riverbed rock Debris accumulation total Riverbed rock 76.85 18.21 25.61 Debris accumulation 23.15 81.79 74.39 total 100 100 100

[0126] Overall accuracy of riverbed bedrock and debris accumulation: 81.16%.

[0127] In order to illustrate the effect of the present invention on the extraction of the distribution range of riverbed bedrock in a small watershed, the extraction of riverbed bedrock and bedrock debris deposits in a small watershed area was also carried out using the multispectral data of the Pegasus D2000 UAV, and the results are shown in Tables 1 and 2. Table 1 shows the extraction results of the traditional U-Net deep learning feature classification method, and the extraction accuracies of riverbed bedrock and bedrock debris deposits are 64.48% and 63.45% respectively, and Table 2 shows the extraction results of the present invention, and the extraction accuracies of riverbed bedrock and bedrock debris deposits are 76.85% and 81.79% respectively. It can be seen that the high-resolution multispectral data obtained by the UAV in the present invention are pre-processed by data acquisition, stitching, reflectance image output, etc., which is convenient for delineating the riverbed bedrock and bedrock debris area, and according to the distribution area of ​​the riverbed bedrock and bedrock debris, the masked high-resolution multispectral image of the UAV is transformed into MNF to obtain the MNF false color image of the enhanced riverbed bedrock and bedrock debris area; the acquired MNF false color image of the enhanced riverbed bedrock and bedrock debris area is subjected to riverbed bedrock and bedrock debris sample screening; collecting the characteristic samples of riverbed bedrock and bedrock debris in the MNF false color image is an important link, and the key is to select the riverbed bedrock outcrops with relatively smooth surface and the bedrock debris samples with rich surface texture, and input the two samples on the small range image. U-Net deep learning is carried out to obtain the identification model of riverbed bedrock and bedrock debris, and finally the riverbed bedrock and bedrock debris are separated through the model; based on the first U-Net deep learning, the distribution area of ​​riverbed bedrock and bedrock debris deposits is delineated, and the distribution area of ​​riverbed bedrock and bedrock debris deposits is masked, and then MNF change processing is performed in the masked area to enhance the difference between riverbed bedrock and bedrock debris deposits in MNF change images, and finally U-Net deep learning classification is performed for the second time, which can efficiently, quickly and accurately determine the boundary of riverbed bedrock and bedrock debris deposits in a short time, and can efficiently, quickly and accurately identify the distribution range of riverbed bedrock in small watersheds in lithology mapping for small watershed river regulation work.

[0128] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

Claims

1. A riverbed rock identification method based on UAV multispectral data, characterized in that: The following steps are involved: Acquire high-resolution multispectral remote sensing images from drones; Image processing was performed on high-resolution multispectral remote sensing images to obtain typical blocks of riverbed bedrock as training samples, and U-Net deep learning was performed to determine the distribution area of ​​riverbed bedrock and bedrock debris deposits; The method for determining the distribution area of ​​riverbed bedrock and bedrock debris deposits specifically includes: Combine the B5, B4, B3, B2, and B1 bands of the high-resolution multispectral images collected by the UAV to obtain a combined image of the B5, B4, B3, B2, and B1 bands; The five-band combined image is processed for decorrelation analysis to obtain a processed high-resolution image; Based on the similar tones of riverbed bedrock and bedrock debris deposits in high-resolution images, some bedrock outcrops and rubble deposits were circled as training samples. Based on the training samples, the image is subjected to improved U-Net deep learning to identify the distribution areas of bedrock outcrops and rubble deposits in the entire image; According to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, the processed high-resolution image is masked to obtain an image highlighting the riverbed bedrock and bedrock debris accumulation; The mask part of the prominent riverbed bedrock and bedrock debris accumulation is transformed by MNF, and a false color image is generated; The riverbed bedrock and bedrock debris accumulation were selected as training samples on the MNF transformed false color images for improved MNF-U-Net deep learning; The model obtained by deep learning of the improved MNF-U-Net is used to classify the MNF transformed images and distinguish the riverbed bedrock and bedrock debris deposits.

2. The riverbed rock identification method based on UAV multispectral data according to claim 1 is characterized in that: The typical blocks of riverbed bedrock are obtained as training samples, specifically including: Perform natural color synthesis of RGB bands on high-resolution drone images to obtain natural true color images; The exposed riverbed bedrock has a high light reflectivity, so it appears as a bright yellow-brown hue in natural true color images, and the hue is different from the grayish white of concrete floors and pavement. The bedrock debris deposits appear in a hue similar to that of the riverbed bedrock in natural true color images, but with a lower reflectivity. According to the image characteristics of the riverbed bedrock and bedrock debris deposits, the region of interest (ROI) tool was used to define the sample range.

3. The riverbed rock identification method based on UAV multispectral data according to claim 1 is characterized in that: The method of selecting riverbed bedrock and bedrock debris deposits as training samples on the MNF transformed false color image to perform improved MNF-U-Net deep learning specifically includes: First, the first U-Net algorithm learning was performed on samples of riverbed bedrock and bedrock debris accumulation, and the distribution areas of riverbed bedrock and bedrock debris accumulation were delineated on high-resolution multispectral images of drones. Then, the distribution area of ​​the riverbed bedrock and bedrock debris accumulation is masked, and the masked area is subjected to MNF transformation to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation in the image; Samples of riverbed bedrock and bedrock debris deposits are input, and a second U-Net algorithm learning is performed in the mask area to distinguish the riverbed bedrock outcrops and bedrock debris deposits in the mask area.

4. The riverbed rock identification method based on UAV multispectral data according to claim 1 is characterized in that: The mask part of the protruding riverbed bedrock and bedrock debris deposits is subjected to MNF transformation and a false color image is generated, specifically including: According to the distribution area of ​​the identified riverbed bedrock and bedrock debris accumulation, mask processing is performed on the image of the identified area; The masked area is processed by MNF transformation, and the first three MNF transformation bands are synthesized to output a false color image that enhances the difference between the riverbed bedrock and the bedrock debris deposits.

5. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps: Acquire high-resolution multispectral remote sensing images from drones; Perform image processing on high-resolution multispectral remote sensing images to obtain typical blocks of riverbed bedrock as training samples, conduct deep learning, and determine the distribution area of ​​riverbed bedrock and bedrock debris accumulation; The method for determining the distribution area of ​​riverbed bedrock and bedrock debris deposits specifically includes: Combine the B5, B4, B3, B2, and B1 bands of the high-resolution multispectral images collected by the UAV to obtain a combined image of the B5, B4, B3, B2, and B1 bands; The five-band combined image is processed for decorrelation analysis to obtain a processed high-resolution image; Based on the similar tones of riverbed bedrock and bedrock debris deposits in high-resolution images, some bedrock outcrops and rubble deposits were circled as training samples. Based on the training samples, the image is subjected to improved U-Net deep learning to identify the distribution areas of bedrock outcrops and rubble deposits in the entire image; According to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, the processed high-resolution image is masked to obtain an image highlighting the riverbed bedrock and bedrock debris accumulation; The mask part of the prominent riverbed bedrock and bedrock debris accumulation is transformed by MNF, and a false color image is generated; The riverbed bedrock and bedrock debris accumulation were selected as training samples on the MNF transformed false color images for improved MNF-U-Net deep learning; The model obtained by deep learning of the improved MNF-U-Net is used to classify the MNF transformed images and distinguish the riverbed bedrock and bedrock debris deposits.

6. A riverbed rock and bedrock debris accumulation identification system based on drone multispectral data that implements the riverbed rock identification method based on drone multispectral data as described in any one of claims 1 to 4, characterized in that: The riverbed bedrock identification system based on UAV multispectral data includes: Image acquisition module, a UAV remote sensing data acquisition system used to acquire high-resolution multispectral UAV remote sensing images. The acquired multispectral remote sensing data is 5-band remote sensing image data; The MNF-U-Net riverbed bedrock identification module is used to distinguish riverbed bedrock from bedrock debris deposits and requires three submodules; The riverbed bedrock and bedrock debris accumulation identification submodule uses the U-Net algorithm to combine and identify the riverbed bedrock and bedrock debris accumulation; The riverbed bedrock and bedrock debris accumulation mask enhancement submodule is used to perform masking according to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and perform MNF transformation processing on the image of the flooded area to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation; The riverbed bedrock and bedrock debris deposits differentiation submodule is used to distinguish the riverbed bedrock and bedrock debris deposits in the masked area using the U-Net algorithm.

7. The riverbed bedrock and bedrock debris accumulation identification system based on drone multispectral data according to claim 6 is characterized in that: The three submodules specifically include: The riverbed bedrock and bedrock debris accumulation identification submodule uses the U-Net algorithm to combine and identify the riverbed bedrock and bedrock debris accumulation; The riverbed bedrock and bedrock debris accumulation mask enhancement submodule is used to perform masking according to the distribution area of ​​riverbed bedrock and bedrock debris accumulation, and perform MNF transformation processing on the image of the flooded area to enhance the difference between the riverbed bedrock outcrop and the bedrock debris accumulation; The riverbed bedrock and bedrock debris deposits differentiation submodule is used to distinguish the riverbed bedrock and bedrock debris deposits in the masked area using the U-Net algorithm.

8. The riverbed bedrock and bedrock debris accumulation identification system based on drone multispectral data according to claim 6 is characterized in that: The image acquisition module specifically includes: Multi-rotor drone platform and 5-band multispectral camera sensor; The high-resolution multispectral images of UAVs are stitched and corrected to obtain a 5-band multispectral image with pixel values ​​as surface reflectance values.

9. A method for identifying bedrock in a small watershed, characterized in that: The method for identifying riverbed bedrock in a small watershed uses the method for identifying riverbed bedrock based on drone multispectral data as described in any one of claims 1 to 4.

Images