CN-122023828-A - Lake and reservoir blue algae characteristic extraction and prediction system of multi-source remote sensing and neural network

CN122023828ACN 122023828 ACN122023828 ACN 122023828ACN-122023828-A

Abstract

The invention relates to the technical field of remote sensing monitoring of water area environments, and discloses a lake and reservoir blue algae characteristic extraction and prediction system of a multi-source remote sensing and neural network. The system comprises a data preprocessing module, a feature fusion module, a dynamic prediction module and a situation generation module. The multi-source remote sensing image and the on-site monitoring data are deeply fused through an attention mechanism, a blue algae biomass sequence is generated by using a recursive prediction model, and three-dimensional dynamic simulation is driven to simulate the vertical migration and horizontal diffusion processes of the blue algae biomass sequence. Based on simulation results, the system realizes pixel level identification of the blue algae gathering area, and generates a grading early warning instruction and a situation deduction report by combining a risk probability field. The invention improves the precision of multi-source data fusion characterization, realizes the simulation and early warning of the three-dimensional dynamic behavior of the blue algae, and enhances the space-time coverage capability and early warning prospective of monitoring.

Inventors

LIN LI
DONG LEI
ZHANG JIAMEI
PAN XIONG
ZHANG YUTING
DENG SHANSHAN
GAO YU
HUANG HUAWEI
HAN CHENG

Assignees

长江水利委员会长江科学院

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. The system for extracting and predicting the characteristics of the blue algae in the lakes and reservoirs by using the multi-source remote sensing and the neural network is characterized by comprising the following components: The data preprocessing module acquires a synthetic aperture radar image and a multispectral image of a target water area, performs space enhancement processing to generate an enhanced remote sensing image set, acquires field water quality parameters and microclimate parameters, performs outlier screening and interpolation processing to generate a calibrated field data set; The feature fusion module inputs the enhanced remote sensing image set into a space-time feature encoder based on an attention mechanism to extract primary feature codes, and performs feature intersection and cascade operation on the primary feature codes and the calibrated field data set in a code dimension to generate a fusion feature matrix; The dynamic prediction module inputs the fusion feature matrix into a blue algae biomass recursion prediction model to iteratively output blue algae biomass prediction sequences of a plurality of time points in the future, and drives a blue algae community dynamic simulation unit to simulate the vertical migration and horizontal diffusion processes of the blue algae community dynamic simulation unit to generate a three-dimensional dynamic distribution simulation result; And the situation generating module is used for carrying out pixel-level semantic segmentation on the enhanced remote sensing image set by utilizing the three-dimensional dynamic distribution simulation result, identifying a blue algae gathering region, a diffusion region and a background water body region to generate a blue algae space distribution detailed diagram, further combining a water bloom risk probability field model and the blue algae biomass prediction sequence to calculate risk probability and generate a grading early warning instruction, and finally, aggregating to form a lake and reservoir blue algae monitoring and situation deduction report.
2. The system for extracting and predicting characteristics of blue algae in lakes and reservoirs by using multi-source remote sensing and neural networks according to claim 1, wherein the generating of the enhanced remote sensing image set specifically comprises: dispatching a synthetic aperture radar satellite, collecting a back scattering intensity image of a target water area, and performing speckle noise suppression and terrain correction on the back scattering intensity image to generate a surface roughness characteristic image; Dispatching a multispectral satellite, synchronously collecting visible light to short wave infrared band images of a target water area, and carrying out atmospheric scattering correction and water normalization on the visible light to short wave infrared band images to generate reflectivity characteristic images; and carrying out pixel-level spatial registration on the surface roughness characteristic image and the reflectivity characteristic image, and fusing the characteristics of the two images by adopting a multi-scale image fusion method to generate an enhanced remote sensing image set with high spatial resolution and rich spectral information.
3. The system for extracting and predicting characteristics of blue algae in lakes and reservoirs by using multi-source remote sensing and neural network according to claim 2, wherein the generating of the calibrated field data set specifically comprises: A sensor node network arranged in a target water area acquires water quality parameters consisting of chlorophyll concentration, phycocyanin concentration, water temperature and turbidity, and microclimate parameters consisting of illumination intensity, water surface wind speed, air temperature and humidity in real time; carrying out time sequence analysis on the water quality parameters and the microclimate parameters of each sensor node, and identifying and screening out abnormal measured values by using a statistical outlier detection method; Filling the data blank corresponding to the abnormal measured value by adopting a space-time-based Kriging interpolation method, and generating a calibrated field data set with continuous time and complete space.
4. The system for extracting and predicting blue algae characteristics in lakes and reservoirs by using multi-source remote sensing and neural networks according to claim 3, wherein the step of inputting the enhanced remote sensing image set into a space-time characteristic encoder based on an attention mechanism to extract primary characteristic codes comprises the following steps: the space-time feature encoder outputs a primary feature code comprising blue algae spectral response features and spatial texture features, Constructing a space-time feature encoder, wherein the space-time feature encoder is formed by connecting a spectrum attention sub-network and a space convolution sub-network in parallel; the spectrum attention sub-network performs weight learning on multispectral wave bands of the enhanced remote sensing image set, and strengthens blue algae characteristic sensitive wave bands; the space convolution sub-network carries out multi-level convolution operation on the space domain of the enhanced remote sensing image set, and extracts blue algae space textures and morphological characteristics under different scales; And splicing the outputs of the spectrum attention sub-network and the space convolution sub-network in the dimension of the characteristic channel, and outputting the primary characteristic code through a characteristic compression layer.
5. The system for extracting and predicting blue-green algae characteristics in lakes and reservoirs of a multi-source remote sensing and neural network according to claim 4, wherein performing characteristic crossing and cascading operations on the primary characteristic codes and the calibrated field data set in the code dimension to generate a fusion characteristic matrix comprises: Mapping the calibrated field data set into a field feature code with the same dimension as the primary feature code through a fully-connected encoder; Constructing a feature cross layer, and performing element-by-element multiplication operation on the primary feature code and the field feature code to generate cross features; Sequentially cascading the cross feature, the primary feature code and the field feature code in feature dimensions to form a high-dimensional cascading feature; And carrying out principal component analysis on the high-dimensional cascade features, and generating a fused feature matrix with regular dimensions after dimension reduction, wherein the fused feature matrix contains mixed information of remote sensing and ground observation.
6. The system for extracting and predicting characteristics of blue algae in lakes and reservoirs by using multi-source remote sensing and neural network according to claim 5, wherein inputting the fusion characteristic matrix into a blue algae biomass recursion prediction model iteratively outputs blue algae biomass prediction sequences at a plurality of time points in the future comprises: constructing a cyanobacteria biomass recursion prediction model which takes a gating circulating unit as a core and comprises a feedforward output layer; taking the fusion feature matrix as input of initial time, and simultaneously inputting weather parameters in the calibrated field data set corresponding to the current time as a condition variable; The gating circulation unit updates memory according to the current input and hidden state, and the feedforward output layer maps the updated hidden state into a blue algae biomass predicted value at the current moment; and taking the predicted value at the current moment as part of the next moment to be input, and iteratively executing the prediction process to generate a group of future blue algae biomass prediction sequences which are arranged in time sequence.
7. The system for extracting and predicting characteristics of blue algae in lakes and reservoirs by using multi-source remote sensing and neural network according to claim 6, wherein the driving the blue algae community dynamic simulation unit to simulate the vertical migration and horizontal diffusion processes thereof to generate the three-dimensional dynamic distribution simulation result comprises: Constructing a blue algae community dynamic simulation unit, wherein a fluid dynamics model and a blue algae growth migration model are built in the blue algae community dynamic simulation unit; The fluid dynamic model simulates a three-dimensional flow field and a temperature layered structure of a target water area according to on-site water quality parameters and microclimate parameters; Inputting the blue algae biomass prediction sequence serving as an initial biomass field into the blue algae growth migration model, and calculating a motion track of a blue algae population under the driving of buoyancy adjustment and water flow by combining the three-dimensional flow field and a temperature layered structure; And simulating the vertical distribution change and the horizontal transportation process of the blue algae biomass in the future period in a three-dimensional grid space, and outputting a three-dimensional dynamic distribution simulation result taking time as a dimension.
8. The system for extracting and predicting blue algae characteristics in lakes and reservoirs by using multi-source remote sensing and neural networks according to claim 7, wherein the method for generating a detailed spatial distribution diagram of blue algae by using the three-dimensional dynamic distribution simulation result to perform pixel-level semantic segmentation on the enhanced remote sensing image set and identifying blue algae aggregation areas, diffusion areas and background water areas comprises the following steps: Extracting the spatial distribution of the blue algae biomass of the surface water body from the three-dimensional dynamic distribution simulation result, and taking the spatial distribution as a priori probability map of semantic segmentation; constructing a multi-task semantic segmentation network which shares an encoder but has three decoder heads for distinguishing an aggregation area, a diffusion area and a background water area respectively; inputting the enhanced remote sensing image set and the prior probability map into a shared encoder of the multi-task semantic segmentation network; And the three decoder heads respectively output pixel level probability diagrams of corresponding categories, determine the final category of each pixel by comparing the probability sizes of the categories, and synthesize blue algae space distribution detailed diagrams with space category labels.
9. The system for extracting and predicting characteristics of blue algae in lakes and reservoirs by using multi-source remote sensing and neural network according to claim 8, further calculating risk probability by combining a bloom risk probability field model and the blue algae biomass prediction sequence and generating a hierarchical early warning instruction, comprising: dividing a target water area into a plurality of gridding subareas; Building a water bloom risk probability field model, wherein the water bloom risk probability field model is input into a category composition of each partition in the blue algae space distribution detailed diagram, a predicted value of a corresponding partition in a blue algae biomass predicted sequence and a site water quality parameter corresponding to the partition; the water bloom risk probability field model calculates the conditional probability of each subarea generating water bloom at a specific moment in the future, namely the instantaneous risk probability, through a feedforward neural network; Presetting a plurality of risk probability thresholds, comparing the instantaneous risk probability of each partition with the threshold, and generating early warning instructions of corresponding levels if the instantaneous risk probability exceeds a certain threshold, wherein the early warning instructions of all the partitions are summarized as hierarchical early warning instructions; the water bloom risk probability field model is constructed by the following steps: collecting historical water bloom event records of a target water area, and taking the water bloom occurrence condition of each meshed partition as tag data; Selecting detailed blue algae space distribution pattern type characteristics, blue algae biomass prediction sequence characteristics and on-site water quality parameter characteristics of each meshed partition before the occurrence of a historical water bloom event as model input characteristics; establishing a water bloom risk probability field model based on a depth fully connected neural network, wherein the depth fully connected neural network comprises a plurality of hidden layers, and a batch normalization layer and an activation function are accessed after each hidden layer; Training a water bloom risk probability field model by using historical data, and optimizing network weights by using a back propagation algorithm to enable probability values output by the model to be matched with actual water bloom occurrence conditions; And deploying the trained water bloom risk probability field model into a system, and calculating the instantaneous risk probability of each meshed partition for generating water bloom at a specific future moment.
10. The system for extracting and predicting blue algae characteristics in lakes and reservoirs by using multi-source remote sensing and neural networks according to claim 4, wherein the spatial convolution sub-network performs multi-level convolution operation on the spatial domain of the enhanced remote sensing image set, and extracts the blue algae spatial textures and morphological characteristics under different scales, and the system comprises: constructing a space convolution sub-network with a parallel multi-branch architecture, wherein each branch uses convolution kernels with different sizes to carry out convolution operation; The first branch uses a large-size convolution kernel to carry out convolution on the enhanced remote sensing image set, and captures macroscopic space texture characteristics formed by blue algae bloom in a large range; The second branch uses a medium-size convolution kernel to carry out convolution on the enhanced remote sensing image set, and morphological contour features of the blue algae gathering area under the medium scale are extracted; the third branch uses a small-size convolution kernel to carry out convolution on the enhanced remote sensing image set to obtain local detail texture features of blue algae plaques under a small scale; and splicing the multi-scale feature graphs output by the three branches in the channel dimension, and generating blue algae space texture and morphological feature codes containing multi-scale space information through a cross-channel information fusion layer.

Description

Lake and reservoir blue algae characteristic extraction and prediction system of multi-source remote sensing and neural network Technical Field The invention relates to the technical field of remote sensing monitoring of water area environments, in particular to a lake and reservoir blue algae characteristic extraction and prediction system of a multi-source remote sensing and neural network. Background The existing lake and reservoir cyanobacteria monitoring technology mainly depends on remote sensing inversion of a single data source or water quality monitoring of discrete sites. The remote sensing method is mainly based on band combination calculation chlorophyll index of multispectral images, has insufficient detection capability on calm water surface, cloud cover or blue algae initial aggregation period, can obtain accurate parameters in site monitoring, but has limited space representativeness, and is difficult to reflect the overall condition of large-area water body. The conventional technical scheme generally carries out simple superposition or statistical analysis on remote sensing inversion results and point measurement data in space, and the processing mode belongs to shallow data integration, and cannot fully mine deep association and complementary value of multi-source heterogeneous data on a blue algae growth characterization driving mechanism. In the aspect of blue algae prediction, the existing method is mainly focused on constructing a statistical or machine learning model by using historical time sequence data to predict the blue algae biomass concentration of a specific point or area in the future. The prediction results are presented in the form of numerical values or two-dimensional concentration fields, and the consideration of the physiological and ecological behaviors of the blue algae is lacked. The vertical migration and horizontal diffusion of the blue algae community are key three-dimensional dynamic processes for forming and dissipating water bloom, and only the biomass predicted value is output, so that the spatial movement track and three-dimensional distribution evolution of the blue algae community cannot be intuitively reflected, and the accuracy of early warning and the pertinence of prevention and control measure formulation are limited. Disclosure of Invention The invention aims to provide a lake and reservoir blue algae characteristic extraction and prediction system of a multi-source remote sensing and neural network, so as to solve the problems in the background technology. In order to achieve the above purpose, the invention provides a system for extracting and predicting characteristics of blue algae in lakes and reservoirs by using a multi-source remote sensing and neural network, which comprises: The data preprocessing module is used for taking a synthetic aperture radar image and a multispectral image of a target water area, performing space enhancement processing to generate an enhanced remote sensing image set, acquiring field water quality parameters and microclimate parameters, performing outlier screening and interpolation processing, and generating a calibrated field data set; The feature fusion module inputs the enhanced remote sensing image set into a space-time feature encoder based on an attention mechanism to extract primary feature codes, and performs feature intersection and cascade operation on the primary feature codes and the calibrated field data set in a code dimension to generate a fusion feature matrix; The dynamic prediction module inputs the fusion feature matrix into a blue algae biomass recursion prediction model to iteratively output blue algae biomass prediction sequences of a plurality of time points in the future, and drives a blue algae community dynamic simulation unit to simulate the vertical migration and horizontal diffusion processes of the blue algae community dynamic simulation unit to generate a three-dimensional dynamic distribution simulation result; And the situation generating module is used for carrying out pixel-level semantic segmentation on the enhanced remote sensing image set by utilizing the three-dimensional dynamic distribution simulation result, identifying a blue algae gathering region, a diffusion region and a background water body region to generate a blue algae space distribution detailed diagram, further combining a water bloom risk probability field model and the blue algae biomass prediction sequence to calculate risk probability and generate a grading early warning instruction, and finally, aggregating to form a lake and reservoir blue algae monitoring and situation deduction report. Preferably, generating the enhanced remote sensing image set specifically includes: dispatching a synthetic aperture radar satellite, collecting a back scattering intensity image of a target water area, and performing speckle noise suppression and terrain correction on the back scattering intensity image to generate a surfac