River water ecological space global monitoring method and system

Abstract

The invention relates to a river water ecological space global monitoring method and system, and the method comprises the steps: recognizing and segmenting a water body image from an RGB image photographed by an unmanned plane based on a U-Net deep learning model, and forming a water body recognition image; converting and reconstructing the RGB image aerially photographed by the unmanned aerial vehicle into a water body hyperspectral image based on a hyperspectral information continuation reconstruction model; the U-Net deep learning model extracts a multispectral segment sequence of pixel points of a water body in the water body hyperspectral image and sends the multispectral segment sequence to the classification regression model; and calculating the multispectral segment sequence based on a classification regression model so as to label a concentration classification result of chlorophyll a in a water body hyperspectral image, thereby realizing visualization of chlorophyll concentration. The method gets rid of limitation of technical means and time acquisition of hyperspectral data acquisition, and realizes the effect of acquiring river global images at any time and at any moment.

Description

Technical Field

[0001] The present invention relates to the technical field of water environment monitoring, and in particular to a method and system for monitoring the entire water ecological space of a river. Background Art

[0002] Currently, remote sensing data is generally used to infer the spatial distribution of water quality parameters, thereby achieving global water quality monitoring. Spaceborne hyperspectral data is acquired using hyperspectral imagers aboard satellites in Earth orbit. This data has a wide coverage area and provides continuous spectral information, making it suitable for large-scale surface monitoring and analysis. For example, the Earth Observation-1 (EO-1) satellite, launched by NASA in 2000, carried the Hyperion imaging spectrometer. Its data contains 220 spectral channels, covering a range of 0.4 to 2.5 μm. Its orbital altitude is 705 km, with a spatial resolution of 30 m and a swath width of 7.7 km. The Gaofen-5 (GF-5) satellite is the world's first hyperspectral satellite to cover the entire spectral range, enabling simultaneous observation of both land and atmosphere. Its Advanced Hyperspectral Imager (AHSI) covers 330 bands, covering a range of 0.4 to 2.5 μm, with a spatial resolution of 30 m and a swath width of 60 km.

[0003] Airborne hyperspectral data is collected at a lower altitude by carrying a hyperspectral imager on an aircraft. Compared with satellite-borne data, airborne data has higher spatial resolution and more flexible flight arrangements, making it suitable for detailed monitoring and analysis of specific areas. Airborne hyperspectral remote sensing technology can obtain continuous and complete spectral data of ground objects, with a spectral resolution of 10 -2λ Above the magnitude.

[0004] The acquisition of satellite-borne and airborne hyperspectral data each has its advantages and disadvantages. Although satellite-borne hyperspectral data has a wide coverage area and can provide sufficient support for water quality inversion of large rivers, it is limited by the satellite's transit time and spatial resolution, making it impossible to pay attention to the tributaries of large rivers. At the same time, satellite photography is also limited in terms of temporal flexibility. The target can only be photographed when the satellite is passing, and it is impossible to detect and analyze the current water quality at any time and quickly. Airborne hyperspectral data can simultaneously guarantee high spatial resolution and spectral resolution, making it the best choice for remote sensing water quality inversion of tributaries. However, the acquisition of airborne hyperspectral data has difficulties such as high acquisition cost, high experimental difficulty and complex operating personnel skills. These problems restrict the practical application of airborne water quality remote sensing.

[0005] In addition, remote sensing inversion of chlorophyll concentration in the Beibu Gulf is currently being conducted based on the OCLI water color sensor from the Sentinel-3A satellite. OCLI water color sensor data relies on satellite sensors and may be affected by natural factors such as weather and cloud cover, resulting in some data being unavailable or unstable. Furthermore, when remotely sensing large areas, the spatial resolution of OCLI water color sensor data from the Sentinel-3A satellite can be relatively low. In contrast, RGB imagery from drones can have higher spatial resolution, enabling more detailed capture of surface features.

[0006] For the synchronized inversion of small tributaries across the entire river basin, satellite aerial photography struggles to accurately capture information due to natural factors like weather and cloud cover, as well as limited acquisition cycles. Consequently, this inversion cannot be performed accurately. While existing drone aerial photography can capture full-basin images of small river tributaries without time constraints, the accuracy of calculating chlorophyll a concentration in these small tributaries is relatively low.

[0007] For another example, CN111274938A discloses a web-based method for dynamic monitoring of river water quality using high-resolution remote sensing. This method includes the following steps: S1. Obtaining two high-resolution remote sensing river water quality inversion result images to be compared; S2. Performing difference calculations on the two high-resolution remote sensing river water quality inversion result images to obtain a difference raster image of the two high-resolution remote sensing river water quality inversion results; S3. Publishing the difference raster image of the two high-resolution remote sensing river water quality inversion results on a web backend; S4. Slicing the published difference raster image of the two high-resolution remote sensing river water quality inversion results. This technical solution relies on the acquisition and preprocessing of remote sensing image data and water body image data, and the calculation model cannot accurately invert chlorophyll a concentration.

[0008] CN117392564A discloses a river water quality inversion method based on deep learning. The method collects remote sensing image data and water body image data of the target area where the river is located; preprocesses the collected remote sensing image data and water body image data; constructs a river water quality inversion model; inputs the preprocessed remote sensing image data and water body image data into a neural network model to identify the river water body, and then extracts reflectivity data to obtain a data set for river water quality inversion; the obtained data set for river water quality inversion is divided into a training set and a test set; the constructed river water quality inversion model is trained using the obtained training set, and then the trained river water quality inversion model is tested to obtain river water quality inversion results based on deep learning. This technical solution relies on the acquisition and preprocessing of remote sensing image data and water body image data. If the data quality is low or the acquisition is incomplete, it may directly affect the accuracy and robustness of the model.

[0009] CN117422987A discloses a river water quality monitoring method based on satellite remote sensing imagery and machine learning. This method combines satellite remote sensing imagery with machine learning algorithms to construct a surface water quality parameter inversion model. The method includes: locating and acquiring Landsat 8 imagery at monitoring points and preprocessing it; band-combining the preprocessed image dataset; extracting the river imagery to be analyzed in ArcGIS; selecting the corresponding imagery and extracting the reflectivity of each band to form a water quality inversion dataset; constructing a water quality parameter inversion model based on the water quality inversion dataset using a machine learning algorithm; simulating river water quality parameters using the water quality parameter inversion model, and visualizing the simulation results. This technical solution uses Landsat 8 satellite imagery, which may have issues with temporal and spatial resolution, hindering real-time monitoring of water bodies and detailed observation of small-scale water bodies.

[0010] Although the above-mentioned patent documents use machine learning models to process remote sensing image data of large rivers, they can only invert the water quality of rivers. Currently, there is no computational model that can effectively invert the chlorophyll content of rivers.

[0011] For another example, CN116067897A discloses a chlorophyll a concentration inversion method based on an integrated factor system. By comprehensively selecting a chlorophyll a concentration inversion factor construction method, a large number of inversion factors are constructed, and the important inversion factors are preliminarily screened out using factor analysis. Then, a chlorophyll a concentration inversion model is constructed based on measured data and important inversion factors, and the chlorophyll a concentration inversion model is applied to remote sensing images to obtain the chlorophyll a inverted concentration. Although this technical solution improves the inversion accuracy, it depends on the measured data and the selected inversion factors, which limits the universality and adaptability of the method. If the measured data is insufficient or the inversion factors are improperly selected, the accuracy and generalization ability of the model may be affected.

[0012] The present invention hopes to make new improvements to artificial intelligence models and hyperspectral information processing to achieve effective inversion and monitoring of chlorophyll concentration in small tributaries of large rivers.

[0013] In addition, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making the present invention, but due to space limitations, not all details and contents are listed in detail. However, this does not mean that the present invention does not have the characteristics of these prior arts. On the contrary, the present invention already has all the characteristics of the prior art, and the applicant reserves the right to add relevant prior art to the background technology. Summary of the Invention

[0014] The defects of the existing technology include:

[0015] First, it is highly dependent on the reception and processing of hyperspectral data. Although satellite-borne hyperspectral data has a wide coverage area and can provide sufficient support for water quality inversion of large rivers, its spatial resolution is low and it cannot focus on the tributaries of large rivers. At the same time, satellite photography is also limited in its temporal flexibility. The target can only be photographed when the satellite is passing, and it is impossible to detect and analyze the current water quality at any time and quickly. The acquisition of airborne hyperspectral data is difficult due to relatively high acquisition costs, high experimental difficulty, and complex operator skills. These problems restrict the practical application of airborne water quality remote sensing. How to predict and monitor chlorophyll concentration in rivers without relying on hyperspectral data is a technical problem that needs to be solved.

[0016] Second, existing technologies use the OCLI water color sensor to remotely invert chlorophyll concentrations in the Beibu Gulf. This data relies on satellite sensors and may be affected by natural factors such as weather and cloud cover, resulting in some data being unavailable or unstable. Especially in remote sensing of large areas, the spatial resolution of Sentinel-3A satellite data can be relatively low. Furthermore, during the data processing of the OCLI water color sensor, the data model for inverting chlorophyll a concentration is complex and cannot be directly applied to the needs of specific scenarios. More samples and parameter adjustments are required to meet the needs of specific scenarios. Simplifying the calculation model is also a technical issue that needs to be addressed.

[0017] In response to the shortcomings of the existing technology, the present invention provides a method for full-area monitoring of river water ecological space from a first aspect, the method comprising: identifying and segmenting water body images from RGB images taken by drones based on a U-Net deep learning model to form a water body identification image; converting and reconstructing the RGB images taken by drones into water body hyperspectral images based on a hyperspectral information extension reconstruction model and sending the images to the U-Net deep learning model; extracting a multispectral segment sequence of pixel points of the water body in the water body hyperspectral image by the U-Net deep learning model and sending the sequence to a classification and regression model; calculating the multispectral segment sequence based on the classification and regression model to mark the chlorophyll a concentration classification result in the water body hyperspectral image, thereby realizing visualization of chlorophyll concentration.

[0018] The present invention uses RGB images taken by drones as image samples to conduct full-area monitoring of rivers, breaking away from the technical means and time limitations of collecting hyperspectral data, and achieving the effect of collecting full-area images of rivers at any time. The present invention converts aerial RGB images into hyperspectral images through the construction and use of U-Net deep learning models, hyperspectral information extension reconstruction models, and classification regression models, thereby enhancing the ability to capture details of the water environment and providing richer spectral information for water quality analysis. The present invention also predicts and displays the high and low values ​​of chlorophyll concentration in the river in a visual manner, providing intuitive full-area water quality monitoring results, which is conducive to the rapid identification and response to water quality problems.

[0019] According to a preferred embodiment, a method for constructing a U-Net deep learning model includes: inputting RGB images with water body annotations into the U-Net deep learning model as a training set for training, and inputting RGB images without water body annotations into the U-Net deep learning model as a test set for testing, until the recognition rate of the output water body recognition images reaches a specified prediction success rate. Through iterative training of the training and test sets, the present invention optimizes the recognition capabilities of the U-Net deep learning model, ensures the accuracy and reliability of water body recognition, and improves the prediction success rate of the entire monitoring system.

[0020] According to a preferred embodiment, a method for constructing a hyperspectral information continuation reconstruction model includes: linearly rescaling an RGB image to form a dataset; scaling hyperspectral image samples from the dataset to a preset ratio; extracting R, G, and B channel values ​​from the hyperspectral image samples to construct a corresponding RGB image; and establishing HSI-RGB sample pairs. This process utilizes linear scaling and sample scaling techniques to improve the model's ability to process images under varying lighting and environmental conditions, ensuring the accuracy and integrity of spectral information. By establishing HSI-RGB sample pairs, the model's ability to identify and reconstruct water features is enhanced by combining color and spectral information.

[0021] According to a preferred embodiment, the method for constructing a classification and regression model includes: classifying hyperspectral images captured indoors according to chlorophyll concentration, where each pixel in the same hyperspectral image has the same category; dividing the hyperspectral images into training sets for concentration prediction; and visually annotating and displaying the concentration predictions in the water hyperspectral images. This classification and concentration prediction training for indoor hyperspectral images improves the model's ability to identify and predict different concentration levels, providing strong support for quantitative analysis of biological indicators in water. The application of visual annotation technology makes monitoring results more intuitive, facilitating rapid identification and assessment of water quality by monitoring personnel.

[0022] According to a preferred embodiment, the network structure of the hyperspectral information extension reconstruction model includes a minimum spanning tree and a cascade structure composed of several monopole spectrum converters, wherein the minimum spanning tree of the water body identification image is established based on the MST algorithm to retain the correlation relationship between pixels in the water body identification image; the monopole spectrum converter captures the overall context information of the water body identification image.

[0023] The present invention utilizes a minimum spanning tree algorithm to preserve pixel associations in water body recognition images, enhancing the model's ability to capture water body structural features. The present invention's cascaded structure of single-pole spectrum converters effectively captures and converts the spectral information of water body images through the collaborative work of an encoder, bottleneck layer, and decoder.

[0024] According to a preferred embodiment, the unipolar spectrum converter includes an encoder, a bottleneck layer and a decoder; the encoder includes a first downsampling module, N1 spectrum attention mechanism modules, a second downsampling module and N2 spectrum attention mechanism modules; the bottleneck layer includes N3 spectrum attention mechanism modules; and a jumper connection is provided between the encoder and the decoder.

[0025] The downsampling module and spectral attention mechanism module in the encoder improve the model's ability to extract image details and spectral features, enhancing its in-depth understanding of the aquatic environment. The spectral attention mechanism module in the bottleneck layer optimizes the model's spectral conversion efficiency by focusing on key information.

[0026] According to a preferred embodiment, the spectral attention mechanism module is composed of a feedforward network, a spectral multi-head self-attention mechanism module and a two-layer normalization module; the feedforward network captures the nonlinear characteristics of the input based on nonlinear operations; the multi-head mechanism of the spectral multi-head self-attention mechanism module provides attention to multiple spectral scales, and the two-layer normalization module is used to ensure that the output maintains an appropriate scale.

[0027] The nonlinear operations of the feedforward network enhance the model's ability to express input features nonlinearly, improving its adaptability to complex water environments. The multi-scale attention of the spectral multi-head self-attention mechanism enables the model to simultaneously consider information at different spectral scales, enhancing its ability to comprehensively analyze water characteristics. The application of a two-layer normalization module ensures the stability and accuracy of the model's output, providing reliable data support for water monitoring.

[0028] According to a second aspect, the present invention provides a system for monitoring the entire water ecological space of a river. The system includes an image acquisition device and a processor. The image acquisition device is provided on a drone and is used to take RGB images of the water body of the river from the air. The processor is in communication with the image acquisition device and is configured to: identify and segment the water body image from the RGB image taken by the drone based on a U-Net deep learning model to form a water body recognition image; convert and reconstruct the RGB image taken by the drone into a water body hyperspectral image based on a hyperspectral information extension reconstruction model and send it to the U-Net deep learning model; the U-Net deep learning model extracts a multispectral segment sequence of pixel points of the water body in the water body hyperspectral image and sends it to a classification regression model; and calculates the multispectral segment sequence based on the classification regression model to mark the chlorophyll a concentration classification result in the water body hyperspectral image, thereby realizing visualization of the chlorophyll concentration.

[0029] The system of the present invention uses RGB images taken by drones as image samples to conduct full-area monitoring of rivers, breaking away from the technical means and time limitations of collecting hyperspectral data, and achieving the effect of collecting full-area images of rivers at any time. The present invention converts aerial RGB images into hyperspectral images through the construction and use of U-Net deep learning models, hyperspectral information extension reconstruction models, and classification regression models, thereby enhancing the ability to capture details of the water environment and providing richer spectral information for water quality analysis. The present invention also predicts and displays the high and low values ​​of chlorophyll concentration in the river in a visual manner, providing intuitive full-area water quality monitoring results, which is conducive to the rapid identification and response to water quality problems.

[0030] Not only that, the system uses the image acquisition device carried by the drone to achieve real-time RGB image acquisition of river water bodies, greatly reducing the manpower and material resources required by traditional monitoring methods and improving the efficiency and coverage of data collection.

[0031] The processor's U-Net deep learning model leverages efficient parallel computing to rapidly identify and segment water bodies from massive amounts of aerial imagery, significantly improving the speed and accuracy of water body identification. The system's use of drone-based aerial photography enables continuous monitoring of the entire river, eliminating the time and space limitations of traditional monitoring methods and enhancing their flexibility and real-time performance.

[0032] The application of the hyperspectral information extension reconstruction model converts RGB images into hyperspectral images through advanced algorithms, improving the richness and resolution of spectral information and enhancing the ability to capture details of the water environment. This also reduces reliance on hyperspectral equipment and reduces costs. Through deep learning and hyperspectral reconstruction technology, the system can rapidly process and analyze large amounts of image data, significantly improving data processing speed, shortening monitoring cycles, and making water quality analysis more rapid and timely.

[0033] According to a preferred embodiment, the network structure of the hyperspectral information extension reconstruction model includes a minimum spanning tree and a cascade structure composed of several monopole spectrum converters, wherein the minimum spanning tree of the water body identification image is established based on the MST algorithm to retain the correlation relationship between pixels in the water body identification image; the monopole spectrum converter captures the overall context information of the water body identification image.

[0034] The network structure of the hyperspectral information extension reconstruction model is configured in this way. The minimum spanning tree constructed based on the MST algorithm optimizes the correlation between pixels in water body identification images, improving image processing efficiency and accuracy while reducing computing resource consumption. The cascade structure design of the single-pole spectrum converter, through the collaborative work of the encoder, bottleneck layer, and decoder, achieves efficient spectral conversion and information extraction for water body identification images, improving computing performance and processing speed.

[0035] According to a preferred embodiment, the unipolar spectrum converter includes an encoder, a bottleneck layer and a decoder; the encoder includes a first downsampling module, N1 spectrum attention mechanism modules, a second downsampling module and N2 spectrum attention mechanism modules; the bottleneck layer includes N3 spectrum attention mechanism modules; and a jumper connection is provided between the encoder and the decoder.

[0036] The downsampling module and spectral attention mechanism module in the encoder improve the computational efficiency of the model by reducing data dimensions and focusing on key spectral features, while maintaining sensitivity to water body characteristics.

[0037] The spectral attention mechanism module of the bottleneck layer optimizes the spectral conversion efficiency of the model by centrally processing key information, reduces unnecessary calculations, and improves the overall computing performance.

[0038] The jumper connection between the encoder and decoder enables effective transmission and fusion of information, reduces redundant calculations, and improves the computational efficiency and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a communication connection diagram of the river water ecological space full-area monitoring system provided by the present invention;

[0040] Figure 2 This is a simplified schematic diagram of the module connection relationship of the method for monitoring the entire river water ecological space provided by the present invention;

[0041] Figure 3 This is a schematic diagram of water body annotation in RGB images taken by drones using the U-Net deep learning model provided by the present invention;

[0042] Figure 4It is a descending line graph of the loss function of the present invention;

[0043] Figure 5 is a schematic diagram of the recognition results in the test set provided by the present invention;

[0044] Figure 6 It is a schematic diagram of the MST outer layer architecture provided by the present invention;

[0045] Figure 7 is a schematic diagram of the SST architecture provided by the present invention;

[0046] Figure 8 Schematic diagram of the structure of SAB, FFN and S-MSA provided by the present invention;

[0047] Figure 9 This is a schematic diagram of the result of the spectral information continuation reconstruction of the RGB image of a water body provided by the present invention;

[0048] Figure 10 This is a schematic diagram of the visualization results of the field sampling data of some rivers provided by the present invention;

[0049] Figure 11 It is a schematic diagram of the visualization results of the field sampling data of part of the river provided by the present invention.

[0050] Reference Signs List

[0051] 100: UAV; 110: Image acquisition device; 200: Processor; 210: U-Net deep learning model; 220: Hyperspectral information extension and reconstruction model; 230: Classification and regression model; 300: River. DETAILED DESCRIPTION

[0052] The following is a detailed description with reference to the accompanying drawings.

[0053] The defects of the existing technology include:

[0054] First, it relies heavily on the reception and processing of hyperspectral data. While satellite-borne hyperspectral data offers wide coverage and sufficient support for water quality inversion of major rivers, its low spatial resolution prevents the analysis of smaller tributaries. Furthermore, satellite imaging is limited in its temporal flexibility; images can only be taken when the satellite is in transit, preventing the ability to rapidly monitor and analyze current water quality at any given moment. The acquisition of airborne hyperspectral data presents challenges such as relatively high acquisition costs, experimental complexity, and complex operator skills. These issues hinder the practical application of airborne water quality remote sensing. Predicting and monitoring chlorophyll a concentrations in rivers without relying on hyperspectral data presents a technical challenge that needs to be addressed.

[0055] Second, existing technologies use the OCLI water color sensor to remotely invert chlorophyll concentrations in the Beibu Gulf. This data relies on satellite sensors and may be affected by natural factors such as weather and cloud cover, resulting in some data being unavailable or unstable. Especially in remote sensing of large areas, the spatial resolution of Sentinel-3A satellite data can be relatively low. Furthermore, during the data processing of the OCLI water color sensor, the data model for inverting chlorophyll a concentration is complex and cannot be directly applied to the needs of specific scenarios. More samples and parameter adjustments are required to meet the needs of specific scenarios. Simplifying the calculation model is also a technical issue that needs to be addressed.

[0056] For special scenarios requiring simultaneous inversion of the entire basin of small tributaries, current computational models for inversion of watershed chlorophyll a concentrations are unable to accurately calculate and invert chlorophyll a concentrations. The resulting chlorophyll a concentration results are often representative data for a single point in the river basin. This representative chlorophyll a concentration data for a specific watershed in a river only provides a general overview of the basin, but cannot capture the overall chlorophyll a concentration distribution across the entire basin of a river's small tributaries, particularly along the river's banks. Therefore, it is impossible to intuitively determine the true distribution of chlorophyll a concentration across the entire basin of each small tributary.

[0057] Chlorophyll a concentration inversion has important environmental and ecological implications for monitoring water bodies such as rivers. Chlorophyll a is a key indicator of aquatic ecological pollution. Quantitative inversion of chlorophyll a concentration helps monitor changes in the nutrient status of water bodies and is crucial for the management of eutrophic water bodies. By constructing a chlorophyll a concentration inversion model, we can explore the temporal and spatial variations in chlorophyll a concentration. Analyzing the temporal and spatial distribution of chlorophyll a can provide insights into the growth of phytoplankton and the health of the ecosystem within the water body. Monitoring chlorophyll a concentration helps assess and prevent eutrophication, a phenomenon in which excessive nutrients in water lead to excessive algal growth, potentially causing ecological problems such as algal blooms. Therefore, the more accurate the chlorophyll a concentration inversion results, the better.

[0058] To address the shortcomings of existing technologies, the present invention provides a system and method for monitoring the entire river water ecological space. The present invention also provides a system and method for inverting the entire river water ecological space. The present invention also provides a method and system for inverting chlorophyll a in the entire basin of a small tributary.

[0059] The present invention provides a river water ecological space full-area monitoring system, such as Figure 1 As shown, the system includes an image acquisition device 110 and a processor 200. The image acquisition device 110 is provided on the drone 100 and is used to take aerial photos of RGB images of the water body of the river 300. The processor 200 is in communication with the image acquisition device 110.

[0060] The image acquisition device 110 includes a camera module, particularly a high-resolution digital camera capable of capturing high-quality RGB images. The camera module must have sufficient optical zoom capability to accommodate varying shooting distances and viewing angles. Because the drone 100 may encounter airflow disturbances during flight, the image acquisition device 110 also includes a stabilization system. This stabilization system (e.g., a three-axis gimbal) ensures the camera remains stable during capture, reducing image jitter and improving image quality.

[0061] The image acquisition device 110 also includes a communication module. Alternatively, the image acquisition device 110 is connected to the communication module of the drone 100 for data transmission between the drone 100 and the ground processor 200, including image data and other flight parameters of the drone 100. This typically includes a wireless transmission device such as a Wi-Fi or radio frequency (RF) module.

[0062] In addition, the image acquisition device 110 also includes hardware components such as a power module, a storage module, a control module, a connection interface, and a protective housing.

[0063] There are many types of drones 100 equipped with image acquisition devices 110, which are not limited here. The types of drones 100 include fixed-wing drones, multi-rotor drones, unmanned helicopters, professional surveying and mapping drones, etc. For example, professional surveying and mapping drones such as the DJI Phantom 4 RTK are designed specifically for surveying and mapping. They have high-precision image acquisition capabilities and are suitable for the accuracy requirements of 1:500 topographic map aerial photogrammetry industry specifications. For another example, the DJI Mavic 3Pro has high-performance image acquisition capabilities, is equipped with a Hasselblad camera, supports multiple zoom, and is suitable for high-quality image acquisition. For another example, the DJI Matrice 350RTK can be equipped with a variety of sensors, including the Zenmuse P1 full-frame mapping camera and the Zenmuse L2 lidar, and is suitable for high-precision terrain measurement and 3D modeling.

[0064] Processor 200 is used to execute the global monitoring method for river water ecological space of the present invention. This global monitoring method for river water ecological space also includes a global inversion method for river water ecological space. The present invention also provides an unmanned aerial vehicle 100 for executing the global monitoring method for river water ecological space of the present invention, wherein processor 200 is carried on unmanned aerial vehicle 100.

[0065] The processor 200 of the present invention contains a computer model for monitoring the entire river water ecological space. The processor 200 of the present invention may also be connected to other hardware such as a memory, a communication module, and a power supply. The processor 200 may also be integrated into various terminal devices such as servers, cloud servers, ground control stations, and computers to execute the method for monitoring the entire river water ecological space of the present invention.

[0066] The processor 200 of the present invention is a general term, and when produced and implemented, it may include components such as a dedicated integrated chip, an arithmetic logic unit (ALU), a register group, a control unit, a cache, a memory controller, an input / output controller, and a bus interface.

[0067] The processor 200 of the present invention is provided with a U-Net deep learning model 210, a hyperspectral information extension reconstruction model 220 and a classification regression model 230. Figure 2 As shown, the processor 200 is used to execute the river water ecological space global monitoring method of the present invention, including:

[0068] S1: Based on the U-Net deep learning model 210, a water body image is identified and segmented from the RGB image taken by the drone 100 to form a water body recognition image.

[0069] S2: Based on the hyperspectral information extension reconstruction model 220 , the RGB image taken by the drone 100 is converted and reconstructed into a water body hyperspectral image and sent to the U-Net deep learning model 210 .

[0070] S3: The U-Net deep learning model 210 extracts a multispectral segment sequence of pixel points of the water body in the water body hyperspectral image and sends it to the classification regression model 230.

[0071] S4: The classification regression model 230 calculates the multispectral segment sequence to mark the chlorophyll a concentration classification result in the water body hyperspectral image to achieve visualization of the chlorophyll a concentration.

[0072] The system of the present invention uses RGB images captured by drone 100 as image samples for full-scale monitoring of river 300. This eliminates the technical and time constraints of hyperspectral data acquisition, enabling the capture of full-scale river images at any time. The resolution of RGB images captured by drone 100 is superior to that of satellite images, meaning that individual pixels in drone 100 aerial images contain more detailed information, resulting in a clearer and more discernible river channel. However, due to the resolution limitations of satellite remote sensing images, it is difficult to locate small tributaries, making chlorophyll a concentration inversion even more difficult.

[0073] By constructing and utilizing a U-Net deep learning model 210, a hyperspectral information extension and reconstruction model 220, and a classification and regression model 230, the present invention converts aerial RGB images into hyperspectral images, enhancing the ability to capture details of the water environment and providing richer spectral information for water quality analysis. The present invention also visually predicts and displays the high and low values ​​of chlorophyll a concentration in river 300, providing intuitive global water quality monitoring results and facilitating rapid identification and response to water quality issues.

[0074] Furthermore, the system, through the image acquisition device 110 carried by the drone 100, achieves real-time RGB image acquisition of 300 river water bodies, significantly reducing the manpower and material resources required by traditional monitoring methods and improving data collection efficiency and coverage. Compared to traditional satellite remote sensing, drone 100 aerial photography can capture high-resolution RGB imagery at a lower altitude, making it easier to capture details of small tributaries and complex water bodies, and can flexibly adjust aerial photography missions in different weather conditions. The drone's equipped hyperspectral sensor enables the precise acquisition of hyperspectral images of water bodies.

[0075] The U-Net deep learning model 210 configured in the processor 200 leverages efficient parallel computing capabilities to rapidly identify and segment water bodies from large amounts of aerial imagery, significantly improving the speed and accuracy of water body identification. The system's use of drone 100 aerial photography technology enables continuous monitoring of the entire river, eliminating the time and space limitations of traditional monitoring methods and improving monitoring flexibility and real-time performance.

[0076] The application of Hyperspectral Information Extension Reconstruction Model 220 converts RGB images into hyperspectral images through advanced algorithms, improving the richness and resolution of spectral information and the ability to capture details of the water environment while reducing reliance on hyperspectral equipment and lowering costs. Through deep learning and hyperspectral reconstruction technology, the system can rapidly process and analyze large amounts of image data, significantly improving data processing speed, shortening monitoring cycles, and making water quality analysis more rapid and timely.

[0077] The U-Net deep learning model 210 is mainly used for image segmentation tasks, especially pixel-level semantic segmentation. The structure of the U-Net deep learning model 210 includes two parts: an encoder and a decoder, which extract and restore image features through convolution, pooling, upsampling and deconvolution operations. The U-Net deep learning model 210 introduces skip connections, which enables the network to establish a direct, layer-spanning connection between the encoder and the decoder, thereby more effectively capturing the global and local information of the image. The U-Net deep learning model 210 includes two parts: an encoder (downsampling path) and a decoder (upsampling path) as well as an output layer. The encoder part mainly extracts image features through convolution layers, activation function layers, pooling layers and dropout layers; the decoder part restores image features through deconvolution layers, skip connections, splicing layers and upsampling convolution layers; the output layer generates the final segmentation result through convolution layers and activation function layers.

[0078] The method for constructing the U-Net deep learning model 210 includes:

[0079] S11: The RGB images with completed water body annotations are input as a training set into the U-Net deep learning model 210 for training.

[0080] S12: Inputting the RGB image without water body annotation into the U-Net deep learning model 210 as a test set for testing until the recognition rate of the output water body recognition image reaches a specified prediction success rate.

[0081] For example, Figure 3 As shown, 24 RGB images of a river 300 taken by a drone 100 are selected as a training set and manually annotated in ENVI. After the annotation is completed, a classification image is created using the ROI tool and saved in TIFF format for import into the U-Net deep learning model 210 for training.

[0082] After the U-Net deep learning model 210 code training is completed, the training code is trained on a V100 32GB GPU, the batch size is set to 12, and the training is carried out for 20 rounds. The loss function decline line graph is as follows Figure 4 shown.

[0083] After the training is completed, 6 water body extraction effect samples are randomly selected from the test set to test the U-Net deep learning model 210. The test results are as follows Figure 5 shown.

[0084] For example, 10 samples are randomly selected from the test set, and their water body prediction success rate (S) and water body misjudgment rate (F) are calculated. The results are shown in Table 1.

[0085] Table 1: Evaluation results of water body recognition in the test set

[0086]

[0087] Water body prediction success rate:

[0088] Average water body misjudgment rate:

[0089] The U-Net deep learning model 210 has a controllable success rate and error rate for water body identification, making it capable of performing water body identification tasks. Therefore, using the U-Net deep learning model 210 for water body identification can obtain accurate segmentation results while retaining the details and shape characteristics of the water body, helping to more accurately identify and extract water body areas and provide a good data foundation for concentration inversion. The present invention optimizes the recognition capabilities of the U-Net deep learning model 210 through iterative training of training sets and test sets, ensuring the accuracy and reliability of water body identification and improving the prediction success rate of the entire monitoring system.

[0090] According to a preferred embodiment, the network structure of the hyperspectral information extension reconstruction model 220 (MST++ algorithm model (Multi-stage Spectral-wise Transformer)) includes a minimum spanning tree and a cascade structure composed of several single-stage spectral-wise transformers (SST), which is used to achieve accurate mapping from RGB to HSI (hyperspectral image). The hyperspectral information extension reconstruction model 220 consists of N s It is composed of a cascade of single-stage spectral-wise transformers (SSTs), such as Figure 6 shown.

[0091] A minimum spanning tree (MST) of the water body identification image is constructed based on the MST (Mask-guided Spectral-wise Transformer) algorithm, preserving the associations between pixels in the water body identification image. This is achieved by constructing a graph adjacency matrix and applying the MST algorithm. In this process, spatial proximity or color similarity is often used as edge weights in the graph to ensure that the generated minimum spanning tree preserves the image's structural information.

[0092] The network structure of the hyperspectral information extension reconstruction model 220 is configured in this way. The minimum spanning tree constructed based on the MST algorithm optimizes the correlation between pixels in water body identification images, improving the efficiency and accuracy of image processing while reducing computing resource consumption. The cascade structure design of the single-pole spectrum converter, through the coordinated operation of the encoder, bottleneck layer, and decoder, achieves efficient spectrum conversion and information extraction for water body identification images, improving computing performance and processing speed.

[0093] The Single-Stage Spectral Transformer (SST) is used to capture the overall contextual information of the water body recognition image. The Single-Stage Spectral Transformer includes an encoder, a bottleneck layer, and a decoder. Specifically, the Single-Stage Spectral Transformer (SST) is a U-shaped network consisting of an encoder, a bottleneck layer, and a decoder, such as Figure 7 As shown. Figure 7 In

[15] , the embedding block and the mapping block are a single conv 3×3 layer.

[0094] The encoder consists of a downsampling module (a stride conv 4×4 layer), N1 spectral attention mechanism (SAB) modules, a downsampling module, and N2 spectral attention mechanism (SAB) modules. The bottleneck layer consists of N3 spectral attention mechanism (SAB) modules.

[0095] In the decoder, the upsampling module is a cross-row deconv 2×2 layer. To avoid information loss during the downsampling process, the encoder and decoder are connected with jumpers. Preferably, the decoder adopts a symmetrical structure.

[0096] The downsampling module and spectral attention mechanism (SAB) module in the encoder improve the computational efficiency of the model by reducing data dimensions and focusing on key spectral features, while maintaining sensitivity to water body characteristics.

[0097] The spectral attention mechanism module of the bottleneck layer optimizes the spectral conversion efficiency of the model by centrally processing key information, reduces unnecessary calculations, and improves the overall computing performance.

[0098] The jumper connection between the encoder and decoder enables effective transmission and fusion of information, reduces redundant calculations, and improves the computational efficiency and accuracy of the model.

[0099] The encoder is designed based on a U-shaped network structure, consisting of an embedding block and a mapping block. These two blocks use 3x3 convolutional layers to preserve the spatial information of the image. The embedding block introduces feature maps, while the mapping block converts these maps into a higher-dimensional representation.

[0100] The core operation of the encoder is downsampling of the feature map. Through a 4x4 cross-row convolution operation, the size of the feature map is effectively reduced, allowing it to better capture local and global spectral features in subsequent operations.

[0101] According to a preferred embodiment, Figure 8 As shown in Figure 3, the spectral attention mechanism module (SAB) consists of a feedforward network (FFN), a spectral multi-head self-attention mechanism module (S-MSA) and a two-layer normalization module (layer norm).

[0102] Feedforward Network (FFN): This introduces nonlinear mapping through nonlinear operations (such as ReLU), which helps better capture the nonlinear characteristics of the input. The nonlinear operations of the feedforward network enhance the model's ability to express nonlinear input features and improve its adaptability to complex water environments.

[0103] Spectral Multi-Head Self-Attention (S-MSA): This attention mechanism learns both local and global spectral information from the input data. This multi-head mechanism focuses on multiple spectral scales, enhancing the model's sensitivity to spectral features. The multi-scale attention of the Spectral Multi-Head Self-Attention module enables the model to simultaneously consider information at different spectral scales, improving its ability to comprehensively analyze water characteristics.

[0104] Normalization Module (Layer Norm): Two-layer normalization ensures that the output maintains the appropriate scale and contributes to the stability and convergence of the model. The application of the two-layer normalization module ensures the stability and accuracy of the model output, providing reliable data support for water body monitoring.

[0105] The decoder adopts a symmetrical structure and includes an upsampling operation (2x2 cross-row deconvolution layer). This process aims to gradually restore the details and structure of the image by restoring the spatial dimensions.

[0106] To prevent information from being lost during the downsampling process between the encoder and decoder, the hyperspectral information extension reconstruction model 220 introduces skip connections. This connection allows information to be passed through different layers, ensuring that the model can better capture the overall context of the image.

[0107] Finally, the RGB image is converted scientifically and rigorously to HSI using the feature maps processed by the hyperspectral information continuation reconstruction model 220. This process utilizes the rgb_to_hsi function in the colorsys library to ensure accuracy and repeatability of the conversion.

[0108] According to a preferred embodiment, the method for constructing the hyperspectral information continuation reconstruction model 220 includes:

[0109] S21: Linearly rescale the RGB image and form a dataset.

[0110] S22: Scale out hyperspectral image (HSI) samples from the dataset at a preset ratio.

[0111] S23: Extract the R, G, and B channel values ​​from the hyperspectral image sample to construct the corresponding RGB image; establish an HSI-RGB sample pair.

[0112] Specifically, during the training process, the RGB images taken by UAV 100 were linearly rescaled to [0, 1]. Then, HSI samples of size 2016×1512 were scaled by 50% from the dataset, and the specific R, G, and B channel values ​​were extracted from them to construct the corresponding RGB images, and finally the HSI-RGB sample pairs were obtained.

[0113] The batch size was set to 20, and the parameter optimization algorithm was modified using Adam with β1 = 0.9 and β2 = 0.999. The learning rate was initialized to 0.0004, and cosine annealing was used for 300 epochs. Random rotations and flips were used to augment the training data to improve the U-Net segmentation performance.

[0114] Preferably, three methods, stretching, cropping, and incremental dimming, can be used to achieve random rotation and flipping of the training data. Stretching and cropping randomly generate stretching ratios and cropping ranges, and stretch and crop the aerial images and GT images simultaneously. Brightness changes are 20% brighter and 20% darker.

[0115] The hyperspectral information extension reconstruction model 220 proposed in the present invention can be implemented on the Pytorch framework. It takes about 48 hours to train the network on a single RTX 3090 GPU. The MRAE loss function between the predicted HSI and the true HSI is used as the target. In the implementation of the hyperspectral information extension reconstruction model 220, N is set. s =3, N1=N2=N3=1 and C=31.

[0116] During the testing phase, the RGB image is also linearly scaled to [0, 1] and input into the network for spectral restoration. Scaling the RGB image linearly to [0, 1] increases computational accuracy, especially since using floating-point numbers provides more precise calculations and avoids precision loss in integer operations.

[0117] After testing, the hyperspectral information extension reconstruction model 220 takes 102.48 milliseconds to reconstruct each image (size 4032×3024×3) on the RTX 3090 GPU.

[0118] In this process, linear scaling and sample scaling techniques were used to improve the model's ability to process images under different lighting and environmental conditions, ensuring the accuracy and integrity of spectral information. By establishing HSI-RGB sample pairs and combining color and spectral information, the model's ability to identify and reconstruct water features was enhanced.

[0119] The hyperspectral information extension and reconstruction model 220 of the present invention has good effect evaluation.

[0120] Three evaluation metrics are used to assess the performance of the trained model. The first metric is the mean relative absolute error (MRAE), which calculates the pixel-level difference between the reconstructed HSI cube and the ground truth HSI at all wavelengths. Where Y represents the reconstructed HSI cube and N = H × W × Nλ represents the number of all values ​​on the image.

[0121]

[0122] The second metric is the root mean square error (RMSE), which is defined as follows:

[0123]

[0124] Directly set MRAE as the SR type training target. The last indicator is the peak signal-to-noise ratio (PSNR).

[0125] The quantitative analysis results of the hyperspectral information extension and reconstruction model 220 used in the present invention are shown in Table 2. It is obvious that the performance of the hyperspectral information extension and reconstruction model 220 has been greatly improved, and only very few parameters and FLOPS are required. For example, the hyperspectral information extension and reconstruction model 220 has achieved a significant optimization of 34.32db in PSNR, indicating that it has good performance in image reconstruction or other tasks related to image quality. At the same time, only 1.62M parameters are required, which is relatively small. The smaller number of parameters means that the model structure is relatively simple, which helps to reduce the risk of overfitting, and the storage space and computing resources required during deployment and application will also be relatively small. For MRAE, 0.1645 is a relatively low value, indicating that the relative error of the model prediction is small, indicating that the hyperspectral information extension and reconstruction model 220 has a high accuracy in the prediction task. In addition, the RMSE value is 0.0248, which is a low value, which further proves the prediction accuracy of the model.

[0126] Table 2: Quantitative results analysis of hyperspectral information extension reconstruction model 220

[0127]

[0128] From the perspective of qualitative analysis, the present invention also compares the reconstructed HSI with 4 of the 31 spectral channels of 9 SOTA methods and the MST++ model on the valid set.

[0129] Figure 9 This is the application result of the hyperspectral information extension and reconstruction model 220 (MST++ calculation model) to the spectral information extension and reconstruction of the RGB image of a water body. Figure 9 The color photo in depicts the input raw RGB image. Figure 9 The colored images in the figure show HSI patches of some reconstructed channels of an RGB image. Unlike traditional HSI detail recovery methods, which either achieve overly smoothed HSIs, sacrificing fine-grained content and structural details, or introduce unpleasant artifacts and speckled textures, the hyperspectral information extension reconstruction model 220 performs better in generating pleasing HSIs with sharp edges and maintaining spatial smoothness in uniform areas. This is primarily due to the MST++ computational model's expertise in modeling inter-spectral self-similarity and dependencies.

[0130] In this paper, K-Nearest Neighbors (KNN) is a simple and intuitive supervised learning algorithm that can be used for classification and regression tasks. The core idea of ​​this algorithm is to classify samples based on a distance metric. KNN can be used to solve a variety of problems, including pattern recognition, data mining, and recommendation systems.

[0131] KNN features include:

[0132] (1) Simple and intuitive: No assumptions about the data distribution are required, nor is a complex training process required.

[0133] (2) No data input bias: No assumptions are made about the distribution of data, so it is applicable to various data situations.

[0134] (3) Applicable to multi-classification problems: It can easily handle multi-classification problems. Water concentration classification is a typical multi-classification problem.

[0135] (4) Applicable to nonlinear data: It does not need to assume the linear separability of the data. This makes it more adaptable to complex and nonlinear data sets.

[0136] (5) Easy-to-understand results: Provides intuitive classification results and determines the category of the test sample by finding the category of the nearest neighbor sample and voting.

[0137] (6) No training process is required: It is an instance-based learning method. This means that once the training data is loaded, the algorithm can be immediately applied to new test data.

[0138] Therefore, the present invention uses the K-nearest neighbor algorithm to construct the classification regression model 230.

[0139] According to a preferred embodiment, the method for constructing the classification regression model 230 includes:

[0140] S31: Classifying the hyperspectral image taken indoors according to the concentration of chlorophyll a. Pixels in the same hyperspectral image have the same category.

[0141] Specifically, chlorophyll a solutions were imaged indoors, starting at 0 μg / L and increasing in concentration by 4 μg / L, all the way to 80 μg / L. Each two-step gradient was assigned to a class, for a total of 10 classes. Specifically, concentrations of 0-8 μg / L were assigned to class 0, 8-16 μg / L to class 1, and so on. All pixels in the same image were assigned to the same class. A random sample of 500 pixels from each class served as the training set. The entire image served as the test set.

[0142] The distance between two pixels in the feature space reflects the degree of similarity between the two pixels. The pixel feature space is a 3D real vector space, and the distance metric used is Euclidean distance.

[0143] S32: Divide the hyperspectral image into a training set to perform concentration prediction value training, and annotate and display the concentration prediction value in the water body hyperspectral image in a visual manner.

[0144] By training indoor hyperspectral imagery for classification and concentration prediction, the classification and regression model 230 has improved its ability to identify and predict different concentration levels, providing strong support for the quantitative analysis of biological indicators in water bodies. The application of visual annotation technology makes monitoring results more intuitive, allowing monitoring personnel to quickly identify and assess water quality.

[0145] After completing the water body extraction and obtaining the water body hyperspectral image, the U-Net deep learning model 210 extracts a multispectral segment sequence of the pixel points of the water body in the water body hyperspectral image. A multispectral segment sequence refers to extracting a set of reflectance values ​​of a single pixel in each band from a selected key band range (e.g., visible light to near-infrared region) in a hyperspectral image. The reflectance values ​​of each pixel in these bands form a sequence, such as {R1, R2, ..., R n}, where R i represents the reflectance of the i-th band. By analyzing multispectral sequences, water characteristics can be identified and classified for water quality monitoring and pollutant detection. This step simplifies the processing of hyperspectral data and improves the efficiency of feature recognition and data analysis.

[0146] The U-Net deep learning model 210 sends the multispectral segment sequence to the classification and regression model 230. The multispectral segment sequence of the pixel points identified as water bodies is input into the classification and regression model 230, and voting is performed based on the labels of the K nearest neighbor samples, and the category with the majority is selected as the predicted result of the chlorophyll a concentration classification. This operation is performed on all pixel points identified as water bodies, and the chlorophyll a concentration classification results are annotated in the water body identification image, so that the high and low concentrations of chlorophyll a are visualized, thereby enhancing the readability and understandability of the information. Based on the visualization of the chlorophyll a concentration, it is possible to support dynamic comparison of data from different periods, clearly display changes in the ecological environment of the water body, and facilitate ecological monitoring and management. The visualization results of the present invention can be presented in a graphical manner, which is not only convenient for professionals to analyze, but also helps to convey important water quality information to the public and decision makers, and enhance the public's awareness of environmental protection.

[0147] The classification indicators of the classification regression model 230 include:

[0148] (1) OA (Overall Accuracy) is an indicator that measures the overall accuracy of a classification model. It represents the ratio of the number of correctly classified samples to the total number of samples. The higher the OA, the better the overall classification accuracy of the model.

[0149] (2) AA (Average Accuracy) is a metric that measures the average accuracy of a classification model on each category. It calculates the accuracy of each category and then averages the accuracy of all categories. AA takes into account the weight of each category and can therefore better evaluate the performance of the model on different categories.

[0150] (3) Kappa (Cohen's Kappa) is an indicator that considers the consistency between the classifier's prediction and a random prediction. In addition to the accuracy of the classifier, it also considers the consistency between the classification result and the random prediction. The Kappa value ranges from -1 to 1, where 1 indicates perfect agreement, 0 indicates agreement with the random prediction, and -1 indicates complete disagreement.

[0151] After multiple rounds of testing, the classification regression model 230 of the present invention has an OA of 85%, an AA of 83%, and a Kappa of 0.81, indicating an excellent model effect.

[0152] The present invention also verifies the classification regression model 230 using KNN based on image data of wild rivers photographed by drones.

[0153] The results of applying the classification regression model 230 directly to the actual river are poor for the following reasons: (1) The water body is in a white water cup, while the bottom of the actual river is very complex, with some rocks, some water plants, some deep, and some shallow. (2) The indoor hyperspectral data was acquired under a halogen lamp environment, while the outdoor environment is normal sunlight; the lighting conditions of the two are also significantly different.

[0154] Therefore, the present invention makes the following improvement: the spectral sequence of the hyperspectral images taken at 21 field sampling points is subtracted from the spectral sequence of the corresponding indoor standard solution of chlorophyll a concentration. The result of the subtraction can be simply understood as compensation for the field background (river bank, algae) and lighting (the difference in intensity between sunlight and indoor halogen light).

[0155] After removing the background and illumination bias, the indoor model was used for prediction. After multiple rounds of testing, the OA of the classification and regression model 230 was 74%, the AA was 81%, and the Kappa was 0.77. The classification and regression model 230 was well adapted to the field sampling data.

[0156] The experimental visualization results are as follows Figure 10 and Figure 11 As shown: the measured concentration values ​​are [1.52, 10.7, 5.37, 19.0, 7.41, 9.44, 3.74, 4.13, 4.03] (unit: μg / L) and the label categories are [0, 1, 0, 2, 0, 1, 0, 0, 0].

[0157] As described above, through the U-Net deep learning model 210, the hyperspectral information extension and reconstruction model 220, and the classification and regression model 230, the present invention can achieve pixel-level chlorophyll a concentration inversion and global monitoring of the river 300 based on ordinary RGB images. The present invention can achieve pixel-level chlorophyll a concentration inversion based on ordinary RGB images, breaking away from the traditional methods' reliance on hyperspectral images, thereby making the measurement of chlorophyll a concentration in river bodies (e.g., chlorophyll a) more flexible and economical.

[0158] It should be noted that the above-mentioned specific embodiments are exemplary, and those skilled in the art can come up with various solutions inspired by the disclosure of the present invention, and these solutions also fall within the scope of the disclosure of the present invention and fall within the scope of protection of the present invention. Those skilled in the art should understand that the present invention specification and its drawings are illustrative and do not constitute a limitation on the claims. The scope of protection of the present invention is defined by the claims and their equivalents. The present invention specification contains multiple inventive concepts, such as "preferably" and "according to a preferred embodiment", which means that the corresponding paragraph discloses an independent concept, and the applicant reserves the right to file a divisional application based on each inventive concept.

Claims

1. A method for monitoring the entire water ecological space of a river, characterized in that: The method comprises: Based on the U-Net deep learning model (210), a water body image is identified and segmented from the RGB image taken by the drone (100) to form a water body identification image; Based on the hyperspectral information extension reconstruction model (220), the RGB image taken by the drone (100) is converted and reconstructed into a water body hyperspectral image; The U-Net deep learning model (210) extracts a multi-spectral segment sequence of pixel points of the water body in the water body hyperspectral image and sends the sequence to the classification regression model (230); The multi-spectral segment sequence is calculated based on the classification regression model (230) to mark the chlorophyll a concentration classification result in the water body hyperspectral image, thereby realizing the visualization of chlorophyll a concentration.

2. The method for monitoring the entire river water ecological space according to claim 1, characterized in that: The method for constructing the U-Net deep learning model (210) includes: The RGB image with the water body annotation is used as a training set and input into a U-Net deep learning model (210) for training. The RGB image without water body annotation is input as a test set into a U-Net deep learning model (210) for testing until the recognition rate of the output water body recognition image reaches a specified prediction success rate.

3. The method for monitoring the entire river water ecological space according to claim 1 or 2, characterized in that: The method for constructing the hyperspectral information extension reconstruction model (220) comprises: Linearly rescaling the RGB image to form a data set; Scaling a hyperspectral image sample from the data set at a preset ratio; Extracting the values ​​of the three channels of R, G, and B from the hyperspectral image sample to construct a corresponding RGB image; Create HSI-RGB sample pairs.

4. The method for monitoring the entire river water ecological space according to any one of claims 1 to 3, characterized in that: The method for constructing the classification regression model (230) includes: Hyperspectral images taken indoors are classified according to chlorophyll concentration, where each pixel in the same hyperspectral image has the same category; The hyperspectral image is divided into a training set for concentration prediction training, The concentration prediction value is marked and displayed in the water body hyperspectral image in a visual manner.

5. The method for monitoring the entire river water ecological space according to any one of claims 1 to 4, characterized in that: The network structure of the hyperspectral information extension reconstruction model (220) includes a minimum spanning tree and a cascade structure composed of a plurality of monopole spectrum converters. Wherein, a minimum spanning tree of the water body recognition image is established based on the MST algorithm, and the correlation relationship between pixels in the water body recognition image is retained; The monopole spectral converter captures the overall contextual information of the water body identification image.

6. The method for monitoring the entire river water ecological space according to claim 5, characterized in that: The unipolar spectrum converter includes an encoder, a bottleneck layer and a decoder; The encoder includes a first downsampling module, N1 spectral attention mechanism modules, a second downsampling module and N2 spectral attention mechanism modules; The bottleneck layer includes N3 spectral attention mechanism modules; The encoder and the decoder are connected by a jumper wire.

7. The method for monitoring the entire river water ecological space according to claim 6, characterized in that: The spectrum attention mechanism module is composed of a feedforward network, a spectrum multi-head self-attention mechanism module and a two-layer normalization module; The feedforward network captures the nonlinear characteristics of the input based on nonlinear operations; The multi-head mechanism of the spectral multi-head self-attention mechanism module provides attention to multiple spectral scales. The two-layer normalization module is used to ensure that the output maintains the appropriate scale.

8. A river water ecological space full-domain monitoring system, characterized in that: The system comprises: An image acquisition device (110), arranged on the drone (100), for taking an RGB image of a water body of a river (300) by aerial photography; The processor (200) is connected in communication with the image acquisition device (110), and is configured to: Based on the U-Net deep learning model (210), a water body image is identified and segmented from the RGB image taken by the drone (100) to form a water body identification image; Based on the hyperspectral information extension reconstruction model (220), the RGB image taken by the drone (100) is converted and reconstructed into a water body hyperspectral image and sent to the U-Net deep learning model (210); The U-Net deep learning model (210) extracts a multi-spectral segment sequence of pixel points of the water body in the water body hyperspectral image and sends the sequence to the classification regression model (230); The multi-spectral segment sequence is calculated based on a classification regression model (230) to mark the chlorophyll a concentration classification result in the water body hyperspectral image, thereby realizing visualization of chlorophyll a concentration.

9. The river water ecological space global monitoring system according to claim 8 is characterized in that: The network structure of the hyperspectral information extension reconstruction model (220) includes a minimum spanning tree and a cascade structure composed of a plurality of monopole spectrum converters. Wherein, a minimum spanning tree of the water body recognition image is established based on the MST algorithm, and the correlation relationship between pixels in the water body recognition image is retained; The monopole spectral converter captures the overall contextual information of the water body identification image.

10. The river water ecological space global monitoring system according to claim 9 is characterized in that: The unipolar spectrum converter includes an encoder, a bottleneck layer and a decoder; The encoder includes a first downsampling module, N1 spectral attention mechanism modules, a second downsampling module and N2 spectral attention mechanism modules; The bottleneck layer includes N3 spectral attention mechanism modules; The encoder and the decoder are connected by a jumper wire.

Images