Evaluation method and system for ecological system of water ecological space in reservoir area

Abstract

The invention provides a reservoir area water ecological space ecosystem evaluation method and system, and the method comprises the steps: obtaining a multiband spectral remote sensing image, and constructing a hyperspectral and vegetation remote sensing image; obtaining a ground feature category and spatial distribution through a spectrum detection network; modeling correlation between the vegetation index and the leaf area by using a Bayesian network, and constructing a leaf area network fusing a conditional probability structure and a deep neural network; screening a vegetation region based on ground feature distribution, and calculating a pixel-level leaf area index; and evaluating the structural integrity and health state of the ecological system by combining the leaf area index and the ground feature information. According to the method, the incidence relation between the vegetation indexes is quantified through the Bayesian network, the spatial features of the three-dimensional image are extracted in combination with the convolutional neural network, the vegetation growth state of the reservoir area and the spatial heterogeneity of the water ecosystem are effectively captured, and the method is suitable for the ecological space of the reservoir surface, the hydro-fluctuation area and the land area. And reliable technical support is provided for ecological environment assessment and dynamic monitoring and protection.

Description

Technical Field

[0001] This invention relates to the field of environmental protection in water conservancy projects, specifically a method and system for evaluating the water ecological space ecosystem in reservoir areas. Background Technology

[0002] Remote sensing interpretation refers to the process of analyzing remote sensing images or data to extract features of ground objects, identify target objects, and infer their attributes. It is one of the core applications of remote sensing technology and is widely used in resource surveys, environmental monitoring, disaster assessment, urban planning, agriculture, forestry, and other fields.

[0003] A spectrum is the optical spectrum of light, which is the pattern of light separated into monochromatic light by a dispersive system (such as a prism or grating) and then projected onto a detector by an imaging system. Since vegetation and water bodies have different reflectivities for different wavelengths of light, they can be distinguished using spectra.

[0004] Estimating leaf area index (LAI) through remote sensing is a crucial parameter for describing vegetation structure and function. LAI can reflect information such as photosynthesis, water use, material exchange, and energy balance in vegetation to a certain extent. Remote sensing image data, such as Normalized Difference Vegetation Index (NDVI) and Vertical Vegetation Index (PVI), can be used to establish models with measured LAI for detection.

[0005] However, due to the interrelationships between remote sensing data, it is a problem to find vegetation areas and build a more accurate model of remote sensing data and leaf area index to assess the ecosystem, considering multiple land cover categories. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for evaluating the aquatic ecosystem space in reservoir areas, in order to solve the above-mentioned problems existing in the prior art.

[0007] In a first aspect, embodiments of the present invention provide a method for evaluating the aquatic ecosystem space of a reservoir area, including:

[0008] Acquire multiple spectral remote sensing images corresponding to multiple bands covering the water ecological space of the reservoir area;

[0009] According to the spectral bands, all the spectral remote sensing images are constructed into hyperspectral remote sensing images; the spectral remote sensing images corresponding to the near-infrared band and the infrared band are constructed into vegetation remote sensing images.

[0010] The hyperspectral remote sensing image is input into a spectral detection network to detect land features and obtain land feature categories and spatial distribution information; the land feature categories include water bodies and vegetation.

[0011] A leaf area network is constructed using a Bayesian network; the leaf area network is used to detect the leaf area of ​​vegetation through spectral remote sensing images.

[0012] Based on the spatial distribution information of the ground features and the vegetation remote sensing images, multiple leaf area indices are obtained through a leaf area network; one pixel in the vegetation remote sensing image corresponds to one leaf area index.

[0013] Ecosystems are assessed based on leaf area index, land cover categories, and spatial distribution information of land cover.

[0014] Optionally, constructing the leaf area network using a Bayesian network includes:

[0015] Acquire training data sets and labeled data at multiple time points; the training data sets include training normalized vegetation index, training ratio vegetation index, and training vertical vegetation index; the labeled data are labeled leaf areas; the training data are data used to train the leaf area network; the leaf area network is a convolutional neural network.

[0016] Using a Bayesian network, based on the training and labeled data, correlations are detected, resulting in multiple conditional probability tables; these conditional probability tables represent the correlation between wavebands and vegetation.

[0017] Obtain multiple conditional probability structures; each conditional probability structure stores a conditional probability table.

[0018] Based on the conditional probability structure, the deep neural network is adjusted to obtain the leaf area network.

[0019] Optionally, the step of adjusting the deep neural network based on the conditional probability structure to obtain the leaf area network includes:

[0020] The conditional probability structure corresponds to two neurons in the input layer of a deep neural network; the outputs of the two neurons in the input layer are the inputs of a conditional probability structure; the output of the conditional probability structure is the input of one neuron in the hidden layer.

[0021] The values ​​corresponding to the two neurons in the input layer are multiplied by the corresponding probabilities in the conditional probability structure.

[0022] Optionally, the step of detecting associations using a Bayesian network based on the training and labeled data yields multiple conditional probability tables, including:

[0023] The training normalized vegetation index at multiple time points is classified to obtain the training normalized vegetation level and the corresponding training normalized vegetation level interval; the training normalized vegetation level is divided into two levels: low and high; the normalized vegetation level interval represents the range of normalized vegetation index corresponding to different levels.

[0024] The training ratio vegetation index corresponds to the training ratio vegetation level and the corresponding training ratio vegetation level range; the training vertical vegetation index corresponds to the training vertical vegetation level and the corresponding training vertical vegetation level range; and the labeled data corresponds to the labeled data level and the corresponding labeled data level range.

[0025] Based on the training normalized vegetation level and its corresponding training normalized vegetation level interval, the training ratio vegetation level and its corresponding training ratio vegetation level interval, the training vertical vegetation level and its corresponding training vertical vegetation level interval, and the labeled data level and its corresponding labeled data level interval, a conditional probability table is obtained.

[0026] Optionally, the step of obtaining a conditional probability table based on the training normalized vegetation level and its corresponding training normalized vegetation level interval, the training ratio vegetation level and its corresponding training ratio vegetation level interval, the training vertical vegetation level and its corresponding training vertical vegetation level interval, and the labeled data level and its corresponding labeled data level interval includes:

[0027] By using maximum likelihood estimation, we can find the degree to which training data influence each other and the degree to which training data influences labeled data, and obtain the conditional probability value.

[0028] Based on the Bayesian criterion and the conditional probability value, a first Bayesian network is obtained. The first Bayesian network includes multiple directed edges. The directed edges of the first Bayesian network represent the correlation between multiple training data and labeled data, and between multiple training data.

[0029] Based on the first Bayesian network, the conditional probability values ​​are recalculated to obtain the conditional probability table; one conditional probability table corresponds to one directed edge in the first Bayesian network.

[0030] Optionally, the training method for the leaf area network includes:

[0031] The trained normalized vegetation index, the trained ratio vegetation index, and the trained vertical vegetation index are input into the conditional probability structure of the leaf area network to obtain conditional probability features.

[0032] The conditional probability features are input into the neurons of the deep neural network in the leaf area network to obtain the training leaf area value.

[0033] The loss is calculated using the training leaf area values ​​and labeled data, and then the training leaf area network is backpropagated.

[0034] Optionally, based on the spatial distribution information of ground features and the vegetation remote sensing image, multiple leaf area indices are obtained through a leaf area network, including:

[0035] In the spatial distribution information of land cover category 'vegetation', obtain reflectance values ​​of multiple near-infrared bands, reflectance values ​​of multiple red bands, soil reflectance, and vegetation reflectance.

[0036] Based on the reflectance values ​​of the near-infrared band, the red band, the soil reflectance, and the vegetation reflectance, the normalized vegetation index, the specific vegetation index, and the vertical vegetation index are obtained; multiple pixels in the vegetation remote sensing image correspond to multiple normalized vegetation indices, multiple specific vegetation indices, and multiple vertical vegetation indices.

[0037] The normalized vegetation index, the specific vegetation index, and the vertical vegetation index are input into the leaf area network to obtain the first leaf area index feature.

[0038] Based on the normalized vegetation index, the ratio of vegetation index and the vertical vegetation index, the second leaf area index feature is obtained through the first convolutional neural network.

[0039] The first leaf area index feature and the second leaf area index feature are input into a fusion neural network to obtain the leaf area index; multiple normalized vegetation indices, multiple ratio vegetation indices and multiple vertical vegetation indices are used to obtain multiple leaf area indices.

[0040] Optionally, the step of obtaining the second leaf area index feature based on the normalized vegetation index, the ratio of vegetation index to vertical vegetation index, and through the first convolutional neural network includes:

[0041] Based on the position of the corresponding pixel, multiple normalized vegetation indices are used to construct a normalized vegetation image; the relative vegetation index is used to obtain a relative vegetation image; the vertical vegetation index is used to obtain a vertical vegetation image.

[0042] The vegetation remote sensing image, normalized vegetation image, vegetation image and vertical vegetation image are superimposed to form a three-dimensional image to obtain a fused vegetation image.

[0043] The fused vegetation image is input into a first convolutional neural network to obtain the second leaf area index feature.

[0044] Optionally, the training method for the first convolutional neural network includes:

[0045] Acquire training fused vegetation images; the training fused vegetation images represent a three-dimensional image composed of the training normalized vegetation index, the training ratio vegetation index, and the training vertical vegetation index used to train the first convolutional neural network.

[0046] The training fused vegetation image is input into the first convolutional neural network to obtain the training second leaf area index feature;

[0047] The second leaf area index features are trained and input into a fully connected neural network to obtain the second leaf area index.

[0048] The loss is calculated using the second leaf area index and labeled data, and the fully connected neural network and the first convolutional neural network are trained by backpropagation.

[0049] Secondly, embodiments of the present invention provide a reservoir aquatic ecological space ecosystem evaluation system, comprising:

[0050] The acquisition module is used to acquire multiple spectral remote sensing images corresponding to multiple bands covering the water ecological space of the reservoir area;

[0051] The image construction module is used to construct hyperspectral remote sensing images from all spectral remote sensing images according to spectral bands; and to construct vegetation remote sensing images from spectral remote sensing images corresponding to near-infrared and infrared bands.

[0052] The spectral detection module is used to input the hyperspectral remote sensing image into the spectral detection network, detect ground features, and obtain ground feature categories and spatial distribution information.

[0053] The network reconstruction module is used to construct a leaf area network using a Bayesian network; the leaf area network is used to detect the leaf area of ​​vegetation using spectral remote sensing images.

[0054] The leaf area index detection module is used to obtain multiple leaf area indices based on the spatial distribution information of the ground features and the vegetation remote sensing image through a leaf area network; one pixel in the vegetation remote sensing image corresponds to one leaf area index.

[0055] The assessment module is used to assess ecosystems based on leaf area index, land cover categories, and spatial distribution information of land cover.

[0056] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0057] This invention also provides a method and system for evaluating the aquatic ecosystem in reservoir areas, particularly suitable for the complex environments of reservoir surfaces, drawdown zones, and terrestrial ecological spaces. Since the Normalized Difference Vegetation Index (NDI), Specific Vegetation Index (SDI), and Vertical Vegetation Index (VT) are all determined by remote sensing data, they are correlated. Therefore, a Bayesian network is constructed to find a conditional probability table to determine these correlations. A conditional probability structure is constructed using this table, and combining it with a deep neural network allows for more accurate identification of features related to different correlations. An image is then constructed using multiple NDIs, SDIs, and VT. A first convolutional neural network detects the pixel value, NDI, SDI, and VT leaf area index corresponding to the band reflectance of a given pixel. This allows for more accurate extraction of the features of the vegetation or water body corresponding to that area. This enables a more accurate assessment of vegetation growth status and the technical effectiveness of the ecosystem. This invention effectively captures the vegetation growth status and spatial heterogeneity of the aquatic ecosystem in reservoir areas, providing reliable technical support for reservoir ecological environment assessment, vegetation dynamic monitoring, and ecological protection. Attached Figure Description

[0058] Figure 1 This is a flowchart of an evaluation method for a reservoir water ecological space ecosystem provided in an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram of the leaf area network structure in the flowchart of a reservoir water ecological space ecosystem evaluation method provided in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1

[0062] like Figure 1 As shown in the figure, this invention provides a method for evaluating the aquatic ecosystem of a reservoir area, the method comprising:

[0063] S101: Acquire multiple spectral remote sensing images corresponding to multiple bands.

[0064] The spectral remote sensing image is a spectral remote sensing image constructed by covering the water ecological space range of the reservoir area.

[0065] The reservoir water ecological space includes the reservoir water surface, the drawdown zone, and the terrestrial ecological space from the first ridgeline of the reservoir to the check flood level;

[0066] Among them, the unique spectral data are formed by the difference in reflectance of different land features in different bands. For example, vegetation has high reflectance in the near-infrared band, and water bodies have strong absorption in the near-infrared band. This characteristic helps to distinguish the types of land features in the reservoir area.

[0067] In this embodiment, the wavebands used are blue (450nm), green (550nm), red (670nm), infrared, and near-infrared (800nm) to comprehensively capture the spectral information of water bodies and vegetation.

[0068] S102: Construct hyperspectral remote sensing images from all spectral remote sensing images according to the bands; construct vegetation remote sensing images from the spectral remote sensing images corresponding to the near-infrared and infrared bands.

[0069] The hyperspectral remote sensing image refers to a remote sensing image of the reservoir area.

[0070] The spectral remote sensing image is a two-dimensional image, while the hyperspectral remote sensing image and the vegetation remote sensing image are three-dimensional images.

[0071] Specifically, a hyperspectral remote sensing image dataset of the reservoir area was constructed by aligning and integrating all spectral remote sensing images according to bands; at the same time, in order to meet the needs of vegetation information enhancement, near-infrared bands and infrared bands were extracted and fused to construct thematic remote sensing images of vegetation in the reservoir area.

[0072] The above methods are used to enhance the ability to identify vegetation information.

[0073] S103: Input the hyperspectral remote sensing image into the spectral detection network to detect ground features and obtain information on ground feature categories and spatial distribution.

[0074] The land feature categories include typical reservoir area types such as reservoir water bodies, drawdown zone vegetation, reservoir perimeter forests, farmland, and bare land; spatial distribution information of land features.

[0075] In this embodiment, the spectral detection network is a three-dimensional convolutional neural network (3D-CNN), which can effectively process high-dimensional remote sensing data, improve the accuracy of land cover classification, and realize the accurate classification and spatial distribution identification of land cover in the reservoir area.

[0076] Wherein, "land feature" refers to a fixed object on the Earth's surface, either naturally formed or man-made. "Land feature category" indicates the type of land feature; in this embodiment, the land features are water bodies and vegetation. "Land feature spatial distribution information" indicates the area where the land features are located.

[0077] S104: Construct a leaf area network using a Bayesian network; the leaf area network is used to detect the leaf area of ​​vegetation through spectral remote sensing images.

[0078] The structural diagram of the leaf area network is as follows: Figure 2 As shown, it integrates conditional probability structures and deep neural networks, which improves the accuracy of leaf area inversion.

[0079] The leaf area network is suitable for complex underlying surface conditions in the reservoir area. It is used to detect the leaf area index of vegetation based on multispectral remote sensing images of the reservoir area and to couple the topography and ecological characteristics of the reservoir area.

[0080] S105: Based on the spatial distribution information of the ground features and the vegetation remote sensing image, multiple leaf area indices are obtained through the leaf area network; one pixel in the vegetation remote sensing image corresponds to one leaf area index.

[0081] In this image, one pixel in the vegetation remote sensing image corresponds to one area on the earth's surface.

[0082] The leaf area index refers to the leaf area index of the vegetation in the reservoir area.

[0083] S106: Assess the ecosystem based on leaf area index, land cover category, and spatial distribution information of land cover.

[0084] The assessment includes aspects such as vegetation cover, ecosystem structural integrity, and ecological health.

[0085] In this embodiment, vegetation and water areas are distinguished by land cover type, with vegetation areas closer to water bodies exhibiting better ecology. Leaf area index (LAI) is used to assess vegetation status; a higher LAI indicates a more stable ecosystem.

[0086] Optionally, constructing the leaf area network using a Bayesian network includes:

[0087] The training data set and labeled data at multiple time points in the reservoir area are obtained; the training data set includes the training normalized vegetation index, the training ratio vegetation index, and the training vertical vegetation index; the labeled data is the labeled leaf area; the training data is the data used to train the leaf area network; the leaf area network is a convolutional neural network, specifically used to retrieve the vegetation leaf area index based on multispectral remote sensing images of the reservoir area.

[0088] The training Normalized Difference Vegetation Index (NDVI) is obtained by (NIR-R) / (NIR+R). The training Specific Vegetation Index (RVI) is obtained by NIR / R. The training Vertical Vegetation Index (PVI) is obtained by... Obtain. NIR is the reflectance value in the near-infrared band, and R is the reflectance value in the red band. This indicates the reflectance of vegetation in the near-infrared band. This indicates the reflectance of soil in the near-infrared band. This indicates the reflectivity of vegetation in the red light band. This indicates the reflectivity of the soil in the red light band.

[0089] Using a Bayesian network, based on the training and labeled data, correlations are detected, resulting in multiple conditional probability tables; these conditional probability tables represent the correlation between wavebands and vegetation.

[0090] Specifically, the conditional probability table represents the correlation between reflection characteristics of different wavebands and vegetation parameters.

[0091] Obtain multiple conditional probability structures; each conditional probability structure stores a conditional probability table.

[0092] In the conditional probability table for unrelated training and labeled data, all values ​​are 0.5, indicating that the two data have the same probability and have a low degree of mutual influence.

[0093] The inclusion of conditional probability tables for unrelated training and labeled data is intended to represent a low degree of mutual influence between the two in the connection structure of the deep neural network, as described later.

[0094] Based on the conditional probability structure, the deep neural network is adjusted to obtain the leaf area network.

[0095] In this process, the structure and parameters of the deep neural network are adjusted to obtain an optimized leaf area network.

[0096] Among them, the Deep Nueral Network (DNN) has multiple hidden layers, resulting in a leaf area network suitable for the complex environment of the reservoir area.

[0097] Optionally, the step of adjusting the deep neural network based on the conditional probability structure to obtain the leaf area network includes:

[0098] The conditional probability structure corresponds to two neurons in the input layer of a deep neural network; the outputs of the two neurons in the input layer are the inputs of a conditional probability structure; the output of the conditional probability structure is the input of one neuron in the hidden layer.

[0099] The values ​​corresponding to the two neurons in the input layer are multiplied by the corresponding probabilities in the conditional probability structure.

[0100] Among them, the method used to enhance the physical rationality and regional adaptability of feature representation is particularly suitable for the complex underlying surface and diverse vegetation types of reservoir water ecological space.

[0101] Optionally, the step of detecting associations using a Bayesian network based on the training and labeled data yields multiple conditional probability tables, including:

[0102] The training normalized vegetation index at multiple time points is classified to obtain the training normalized vegetation level and the corresponding training normalized vegetation level interval; the training normalized vegetation level is divided into two levels, low and high; the normalized vegetation level interval represents the range of normalized vegetation index corresponding to different levels.

[0103] The above methods are used to adapt to the spatiotemporal variation characteristics of vegetation in the reservoir area.

[0104] Among them, the range of the training normalized vegetation level interval corresponding to the low training normalized vegetation level is smaller than the range of the training normalized vegetation level interval corresponding to the high training normalized vegetation level.

[0105] The training ratio vegetation index corresponds to the training ratio vegetation level and the corresponding training ratio vegetation level range; the training vertical vegetation index corresponds to the training vertical vegetation level and the corresponding training vertical vegetation level range; and the labeled data corresponds to the labeled data level and the corresponding labeled data level range.

[0106] The labeled data levels are divided into large leaf and small leaf.

[0107] Based on the training normalized vegetation level and its corresponding training normalized vegetation level interval, the training ratio vegetation level and its corresponding training ratio vegetation level interval, the training vertical vegetation level and its corresponding training vertical vegetation level interval, and the labeled data level and its corresponding labeled data level interval, a conditional probability table is obtained.

[0108] Optionally, the step of obtaining a conditional probability table based on the training normalized vegetation level and its corresponding training normalized vegetation level interval, the training ratio vegetation level and its corresponding training ratio vegetation level interval, the training vertical vegetation level and its corresponding training vertical vegetation level interval, and the labeled data level and its corresponding labeled data level interval includes:

[0109] By using maximum likelihood estimation, we can find the degree to which training data influence each other and how training data influences labeled data, and obtain conditional probability values.

[0110] Among them, the conditional probability values ​​are the values ​​in the conditional probability table. Table 1 is the conditional probability table between the normalized vegetation index and leaf area during the early stage of wheat growth.

[0111] Table 1

[0112]

[0113] Among them, maximum likelihood estimation (MLE) calculates the conditional probability of each node.

[0114] In this embodiment, for continuous data, it is assumed that node Xi is given its parent node. It follows a known distribution, a common assumption being that the conditional distribution is Gaussian. For each node Xi, given its parent node... The formula for calculating conditional probability is:

[0115]

[0116]

[0117]

[0118] Among them, the This represents the conditional probability value, and N represents the sample size. The term "sample" refers to either training data or labeled data in this embodiment. In this embodiment, there are four types of samples: training normalized vegetation level, training ratio vegetation level, training vertical vegetation level, and labeled data level. The total sample size is four.

[0119] Based on the Bayesian criterion and the conditional probability value, a first Bayesian network is obtained. The first Bayesian network includes multiple directed edges. The directed edges of the first Bayesian network represent the correlation between multiple training data and labeled data, and between multiple training data.

[0120] In this embodiment, the formula for the Bayes criterion is shown in the figure below:

[0121]

[0122] in, This is the online rating score. Let be the conditional probability value, where D is the observed data, G is the network structure, N is the sample size, and k is the number of parameters in the Bayesian network, i.e., the number of edges.

[0123] The algorithm uses a greedy approach to repeatedly find the relationships between training data and between training data and labeled data. It finds the directed edge with the largest network score and adds it to the Bayesian network. Then it checks the network score of other directed edges. If the network score no longer increases, the search stops.

[0124] Based on the first Bayesian network, the conditional probability values ​​are recalculated to obtain the conditional probability table; one conditional probability table corresponds to one directed edge in the first Bayesian network.

[0125] Among them, through Calculate the conditional probability table in the first Bayesian network.

[0126] The above methods improve the accuracy of leaf area index inversion in the reservoir area.

[0127] Optionally, the training method for the leaf area network includes:

[0128] The trained normalized vegetation index, the trained ratio vegetation index, and the trained vertical vegetation index are input into the conditional probability structure of the leaf area network to obtain conditional probability features.

[0129] The conditional probability features are input into the neurons of the deep neural network in the leaf area network to obtain the training leaf area value.

[0130] In this context, the neurons in the deep neural network of the conditional probability feature input leaf area network are neurons that are connected one-to-one with the conditional probability structure, and are used to extract features after the correlation degree is determined by conditional probability. The neurons connected to the conditional probability structure before it are used to extract features from the training data itself.

[0131] The loss is calculated using the training leaf area values ​​and labeled data, and then the training leaf area network is backpropagated.

[0132] In this embodiment, the cross-entropy loss function is used to calculate the loss.

[0133] The above methods are used to adapt to the complex environment of the reservoir's aquatic ecological space.

[0134] Optionally, based on the spatial distribution information of ground features and the vegetation remote sensing image, multiple leaf area indices are obtained through a leaf area network, including:

[0135] In the spatial distribution information of land cover category 'vegetation', obtain reflectance values ​​of multiple near-infrared bands, multiple red bands, multiple soil reflectance values, and multiple vegetation reflectance values.

[0136] Among them, the reflectance values ​​of the near-infrared band, the reflectance values ​​of the red band, the soil reflectance, and the vegetation reflectance of the land surface area corresponding to one pixel in the vegetation remote sensing image are measured.

[0137] Soil reflectance can be measured directly: spectral data of exposed soil can be collected in the field using a spectrometer, or it can be referenced from spectral libraries such as the USGS spectral library and the JHU spectral library, which contain typical soil spectral curves. Vegetation reflectance can be measured from the spectra of healthy vegetation (such as leaves or canopy), or obtained from spectral libraries (such as leaf spectra simulated by the PROSPECT model). Near-infrared and red light reflectance values ​​can be obtained using sensors.

[0138] Based on the reflectance values ​​of the near-infrared band, the red band, the soil reflectance, and the vegetation reflectance, the normalized vegetation index, the specific vegetation index, and the vertical vegetation index are obtained; multiple pixels in the vegetation remote sensing image correspond to multiple normalized vegetation indices, multiple specific vegetation indices, and multiple vertical vegetation indices.

[0139] The Normalized Difference Vegetation Index (NDVI) is obtained by (NIR-R) / (NIR+R), representing the vegetation cover of the area corresponding to a pixel in a vegetation remote sensing image. When vegetation cover is at medium to low levels, this index increases rapidly with increasing cover, but its growth slows down after reaching a certain coverage level, making it suitable for dynamic monitoring of vegetation in the early and middle growth stages. The Relative Vegetation Index (RVI) is obtained by NIR / R, based on the principle that leaves absorb more red light than infrared light, representing the vegetation cover of the area corresponding to a pixel in a vegetation remote sensing image. NIR represents the reflectance value in the near-infrared band, and R represents the reflectance value in the red band, detected by a spectrometer. The Vertical Vegetation Index (PVI) is obtained by... Obtain. This indicates the reflectance of vegetation in the near-infrared band. This indicates the reflectance of soil in the near-infrared band. This indicates the reflectivity of vegetation in the red light band. This represents the reflectivity of the soil in the red light band. , , , This refers to soil and vegetation data measured in a remote sensing image of vegetation at the region corresponding to one pixel using a ground-based spectrometer. A ground-based spectrometer is an instrument that can measure the reflectance of natural objects (such as vegetation, soil, and water bodies) at different wavelengths. The leaf area is obtained using a leaf area meter.

[0140] The normalized vegetation index, the specific vegetation index, and the vertical vegetation index are input into the leaf area network to obtain the first leaf area index feature.

[0141] Based on the normalized vegetation index, the ratio of vegetation index to vertical vegetation index, the second leaf area index feature is obtained through the first convolutional neural network.

[0142] In this example, the first convolutional neural network is a convolutional neural network (CNN).

[0143] The first leaf area index feature and the second leaf area index feature are input into a fusion neural network to obtain the leaf area index; multiple normalized vegetation indices, multiple ratio vegetation indices and multiple vertical vegetation indices are used to obtain multiple leaf area indices.

[0144] The fusion neural network is a fully-connected neural network (FCNN).

[0145] Optionally, the step of obtaining the second leaf area index feature based on the normalized vegetation index, the ratio of vegetation index to vertical vegetation index, and through the first convolutional neural network includes:

[0146] Based on the position of the corresponding pixel, multiple normalized vegetation indices are used to construct a normalized vegetation image; the relative vegetation index is used to obtain a relative vegetation image; the vertical vegetation index is used to obtain a vertical vegetation image.

[0147] The vegetation remote sensing image, normalized vegetation image, average vegetation image, and vertical vegetation image are superimposed to form a three-dimensional image to obtain a fused vegetation image.

[0148] The pixels are then superimposed according to their corresponding positions.

[0149] The fused vegetation image is input into a first convolutional neural network to obtain the second leaf area index feature.

[0150] The above methods enhance the model's ability to perceive the spatial heterogeneity of vegetation distribution in the reservoir area.

[0151] Optionally, the training method for the first convolutional neural network includes:

[0152] Acquire training fused vegetation images; the training fused vegetation images represent a three-dimensional image composed of the training normalized vegetation index, the training ratio vegetation index, and the training vertical vegetation index used to train the first convolutional neural network.

[0153] The training fused vegetation image is input into the first convolutional neural network to obtain the training second leaf area index feature;

[0154] The second leaf area index feature is trained and input into a fully connected neural network to obtain the second leaf area index.

[0155] The fully-connected neural network (FCNN) is used for classification.

[0156] The loss is calculated using the second leaf area index and labeled data, and the fully connected neural network and the first convolutional neural network are trained by backpropagation.

[0157] The above methods improve the prediction accuracy of leaf area index in complex reservoir environments.

[0158] Example 2

[0159] Based on the above-mentioned method for evaluating the water ecological space ecosystem of reservoir areas, this invention also provides a system for evaluating the water ecological space ecosystem of reservoir areas, the system comprising:

[0160] The acquisition module is used to acquire multiple spectral remote sensing images corresponding to multiple bands;

[0161] The image construction module is used to construct hyperspectral remote sensing images from all spectral remote sensing images according to spectral bands; and to construct vegetation remote sensing images from spectral remote sensing images corresponding to near-infrared and infrared bands.

[0162] The spectral detection module is used to input the hyperspectral remote sensing image into the spectral detection network, detect ground features, and obtain ground feature categories and spatial distribution information.

[0163] The network reconstruction module is used to construct a leaf area network using a Bayesian network; the leaf area network is used to detect the leaf area of ​​vegetation using spectral remote sensing images.

[0164] The leaf area index detection module is used to obtain multiple leaf area indices based on the spatial distribution information of the ground features and the vegetation remote sensing image through a leaf area network; one pixel in the vegetation remote sensing image corresponds to one leaf area index.

[0165] The assessment module is used to assess ecosystems based on leaf area index, land cover categories, and spatial distribution information of land cover.

[0166] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0167] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0168] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components in the apparatus according to embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

Claims

1. A method for evaluating an ecological system of a water ecological space in a reservoir area, characterized in that, The method comprises the following steps: acquire a plurality of spectral remote sensing images corresponding to a plurality of wave bands covering the water ecological space of the reservoir area; construct all the spectral remote sensing images into a hyperspectral remote sensing image according to the wave bands; and construct the spectral remote sensing images corresponding to the near-infrared wave band and the infrared wave band into a vegetation remote sensing image; input the hyperspectral remote sensing image into a spectral detection network to detect the ground objects and obtain ground object categories and ground object spatial distribution information; construct a leaf area network through a Bayesian network; the leaf area network is used to detect the leaf area of the vegetation through the spectral remote sensing image; obtain a plurality of leaf area indices through the leaf area network based on the ground object spatial distribution information and the vegetation remote sensing image; one pixel point in the vegetation remote sensing image corresponds to one leaf area index; evaluate the ecological system according to the leaf area index, the ground object category and the ground object spatial distribution information.

2. The method according to claim 1, wherein, The method of constructing the leaf area network through the Bayesian network comprises the following steps: acquire a training data set and labeled data at a plurality of time points; the training data set comprises training normalized vegetation indices, training fractional vegetation indices and training vertical vegetation indices; the labeled data are labeled leaf areas; the training data are data used for training the leaf area network; the leaf area network is a convolutional neural network; detect the correlation relationship based on the training data and the labeled data through the Bayesian network to obtain a plurality of conditional probability tables; the conditional probability table represents the correlation relationship between the wave band and the vegetation; acquire a plurality of conditional probability structures; the conditional probability structure stores the conditional probability table; adjust the deep neural network based on the conditional probability structure to obtain the leaf area network.

3. The method according to claim 2, wherein, The method of adjusting the deep neural network based on the conditional probability structure to obtain the leaf area network comprises the following steps: the conditional probability structure corresponds to two neurons of an input layer in the deep neural network; the output of the two neurons of the input layer is the input of one conditional probability structure; the output of the conditional probability structure is the input of one neuron of a hidden layer; the values corresponding to the two neurons of the input layer are multiplied by the corresponding probabilities in the conditional probability structure.

4. The method according to claim 2, wherein, The method of detecting the correlation relationship based on the training data and the labeled data through the Bayesian network to obtain a plurality of conditional probability tables comprises the following steps: classify the training normalized vegetation indices at a plurality of time points to obtain training normalized vegetation levels and corresponding training normalized vegetation level intervals; the training normalized vegetation level is divided into two levels, i.e., a low level and a high level; the normalized vegetation level interval represents the range of the normalized vegetation index corresponding to different levels; the training fractional vegetation indices correspond to obtain training fractional vegetation levels and corresponding training fractional vegetation level intervals; the training vertical vegetation indices correspond to obtain training vertical vegetation levels and corresponding training vertical vegetation level intervals; the labeled data correspond to obtain labeled data levels and corresponding labeled data level intervals; obtain the conditional probability table based on the training normalized vegetation levels and the corresponding training normalized vegetation level intervals, the training fractional vegetation levels and the corresponding training fractional vegetation level intervals, the training vertical vegetation levels and the corresponding training vertical vegetation level intervals, the labeled data levels and the corresponding labeled data level intervals.

5. The method according to claim 4, wherein, The conditional probability table is obtained based on the training normalized vegetation level, the corresponding training normalized vegetation level interval, the training ratio vegetation level, the corresponding training ratio vegetation level interval, the training vertical vegetation level, the corresponding training vertical vegetation level interval, the labeled data level, and the corresponding labeled data level interval, and includes: Through maximum likelihood estimation, the degree of mutual influence of the training data and the influence of the training data on the labeled data is found, and a conditional probability value is obtained; According to the Bayesian criterion, the first Bayesian network is obtained according to the conditional probability value; the first Bayesian network includes a plurality of directed edges; the directed edges of the first Bayesian network represent that there is an association relationship between the plurality of training data and the labeled data and between the plurality of training data; According to the first Bayesian network, the conditional probability value is recalculated to obtain a conditional probability table; one conditional probability table corresponds to one directed edge in the first Bayesian network.

6. The method according to claim 3, wherein, The training method of the leaf area network includes: The training normalized vegetation index, the training ratio vegetation index, and the training vertical vegetation index are input into the conditional probability structure in the leaf area network to obtain a conditional probability feature; The conditional probability feature is input into the neurons of the deep neural network in the leaf area network to obtain a training leaf area value; The training leaf area value and the labeled data are calculated to obtain a loss, and the leaf area network is back propagated and trained.

7. The method according to claim 1, wherein, The plurality of leaf area indices are obtained through the leaf area network based on the feature spatial distribution information and the vegetation remote sensing image, and include: In the feature spatial distribution information of the vegetation category, a plurality of near-infrared band reflection values, a plurality of red light band reflection values, a plurality of soil reflectances, and a plurality of vegetation reflectances are obtained; According to the near-infrared band reflection value, the red light band reflection value, the soil reflectance, and the vegetation reflectance, a normalized vegetation index, a ratio vegetation index, and a vertical vegetation index are obtained; a plurality of pixels in the vegetation remote sensing image correspond to a plurality of normalized vegetation indices, a plurality of ratio vegetation indices, and a plurality of vertical vegetation indices; The normalized vegetation index, the ratio vegetation index, and the vertical vegetation index are input into the leaf area network to obtain a first leaf area index feature; Based on the normalized vegetation index, the ratio vegetation index, and the vertical vegetation index, a second leaf area index feature is obtained through a first convolutional neural network; The first leaf area index feature and the second leaf area index feature are input into a fusion neural network to obtain a leaf area index; a plurality of normalized vegetation indices, a plurality of ratio vegetation indices, and a plurality of vertical vegetation indices correspond to obtain a plurality of leaf area indices.

8. The method according to claim 7, wherein, The second leaf area index feature is obtained through the first convolutional neural network based on the normalized vegetation index, the ratio vegetation index, and the vertical vegetation index, and includes: According to the positions of the corresponding pixels, the plurality of normalized vegetation indices are constructed into a normalized vegetation image; the ratio vegetation index corresponds to obtain a ratio vegetation image; and the vertical vegetation index corresponds to obtain a vertical vegetation image; The vegetation remote sensing image, the normalized vegetation image, the ratio vegetation image, and the vertical vegetation image are superimposed into a three-dimensional image to obtain a fusion vegetation image; The fusion vegetation image is input into the first convolutional neural network to obtain the second leaf area index feature.

9. The method according to claim 8, wherein, The training method of the first convolutional neural network comprises: Obtaining a training fusion vegetation image; the training fusion vegetation image represents a three-dimensional image composed of a training normalized vegetation index, a training ratio vegetation index and a training vertical vegetation index for training the first convolutional neural network; Inputting the training fusion vegetation image into the first convolutional neural network to obtain a training second leaf area index feature; Inputting the training second leaf area index feature into the full connection neural network to obtain a second leaf area index; Obtaining a loss by using the second leaf area index and labeled data, and backward propagating to train the full connection neural network and the first convolutional neural network.

10. A reservoir water ecological space ecosystem evaluation system, characterized in that, Comprise: The acquisition module is used for acquiring a plurality of spectral remote sensing images corresponding to a plurality of bands covering the water ecological space of the reservoir area; The image construction module is used for constructing all the spectral remote sensing images into a hyperspectral remote sensing image according to the bands; and constructing the spectral remote sensing images corresponding to the near-infrared band and the infrared band into a vegetation remote sensing image; The spectral detection module is used for inputting the hyperspectral remote sensing image into a spectral detection network to detect the ground objects and obtain ground object categories and ground object spatial distribution information; The network reconstruction module is used for constructing a leaf area network through a Bayesian network; the leaf area network is used for a network for detecting the leaf area of the vegetation through the spectral remote sensing image; The leaf area index detection module is used for obtaining a plurality of leaf area indexes through the leaf area network based on the ground object spatial distribution information and the vegetation remote sensing image; one pixel point in the vegetation remote sensing image corresponds to one leaf area index; The evaluation module is used for evaluating the ecosystem according to the leaf area index, the ground object categories and the ground object spatial distribution information.

Images