A high-resolution algal bloom recognition method based on deep learning

Abstract

The present invention relates to the technical field of water body monitoring, in particular to a high-resolution water bloom recognition method based on deep learning. The present invention solves the problems of recognition lag and low efficiency caused by the manual recognition of water bloom in the prior art. A water bloom detection model is established by using multispectral images from remote sensing satellites, and the classification of water bloom features and types is combined. Progress training and verification of the images are performed through a deep learning model, thereby improving the overall recognition accuracy of water blooms in water bodies, and providing an efficient, accurate and scientific basis for the recognition, prevention and timely treatment of water blooms in water bodies.

Description

Technical Field

[0001] The present invention relates to the technical field of water body monitoring, and in particular to a high-resolution algal bloom recognition method based on deep learning. Background Art

[0002] In recent decades, global warming and intensified urbanization have posed a huge threat to the sustainable use of terrestrial freshwater resources. In particular, the frequent occurrence of algal blooms is related to the increasing eutrophication of water bodies. Algal blooms are caused by the excessive growth of certain cyanobacteria in water bodies with long retention times due to the serious enrichment of nutrients in water bodies. It seriously threatens activities such as drinking, irrigation, fishing and recreation.

[0003] In the existing technology, the treatment of algal bloom is based on manual visual judgment or regular inspections of rivers and lakes. Although this method has an intuitive feeling for the outbreak of algal bloom, for water bodies in large areas and long river basins, this method obviously has recognition lags and low efficiency, which in turn affects the timely treatment and early warning of algal blooms.

[0004] Therefore, it is necessary to design a high-resolution algal bloom identification method based on deep learning, which uses high-resolution satellite data combined with subsequent identification models to achieve accurate identification of algal blooms in water bodies. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-resolution algal bloom identification method based on deep learning, which utilizes high-resolution satellite data combined with subsequent identification models to achieve accurate identification of algal blooms in water bodies.

[0006] In order to achieve the above object, the present invention provides a high-resolution algal bloom recognition method based on deep learning:

[0007] S1, data download and processing:

[0008] S1-1, using four 10-m bands and five 20-m bands of multispectral imager images from remote sensing satellites to monitor algal blooms in rivers;

[0009] S1-2, download the images with cloud cover less than 10% in previous years, of which 2 / 3 images are used to construct the training set to establish the water bloom detection model, and the remaining 1 / 3 images are used to test the model performance;

[0010] S1-3, using the field monitoring data of algal cell density and chlorophyll a concentration of the Ministry of Ecology and Environment to classify algal bloom types;

[0011] S1-4, using remote sensing satellite surface temperature products as auxiliary data for identifying algal blooms;

[0012] S1-5, generate the remote sensing reflectance of the water surface through atmospheric correction, and then use the Fmask (Function ofmask) mask function to remove clouds and cloud shadows;

[0013] S1-6, use the Normalized Difference Water Index (NDWI) algorithm to extract water areas and exclude the two pixels adjacent to the edge of the water body:

[0014] NDWI = (G-NIR) / (G+NIR);

[0015] Among them, G is the green light band, NIR is the near infrared band, and the threshold of NDWI is 0;

[0016] S1-7, resample the remote sensing satellite surface temperature product in S1-4 to a spatial resolution of 20 m;

[0017] S2, identification of the spatial extent of water blooms:

[0018] S2-1, Analysis of water bloom characteristics:

[0019] S2-1-1, Classification of algal bloom types:

[0020] The bloom types are divided into three categories: no bloom, mild bloom and severe bloom;

[0021] According to the historical survey records of water blooms, the classification standards for water blooms are:

[0022] No water bloom: when the algae cell density is less than 2.0×106 / L and the chlorophyll a concentration is less than 10μg / L;

[0023] Mild algal bloom: when the algae cell density is between 2.0-10×106 / L and the chlorophyll a concentration is between 10-50μg / L;

[0024] Severe algal bloom: when the algae cell density exceeds 1.0×107 / L and the chlorophyll a concentration exceeds 50μg / L;

[0025] S2-1-2, Construction of water bloom spectrum index:

[0026] The condition of the bloom is determined by its spectral shape and reflectivity:

[0027]

[0028] Among them, NRER is the algae bloom spectral index, and are the normalized values ​​of the red edge band and the red light band respectively;

[0029] Normalized difference vegetation index NDVI and floating algae index FAI:

[0030]

[0031] in, and are the surface reflectance values ​​in the near infrared and short-wave infrared bands, respectively, red , NIR and λ SWIR They are the wavelengths of the red band, near infrared band, and short-wave infrared band;

[0032] S2-2, Construction of deep learning model for water bloom identification:

[0033] S2-2-1, Architecture design: Based on the convolutional neural network, a water bloom recognition model consisting of 13 convolutional layers and 3 fully connected layers was designed according to the structure of the VGG (Visual Geometry Group)-16 network;

[0034] S2-2-2, network structure:

[0035] S2-2-2-1, input variables include 9 spectral bands of remote sensing satellite images, spectral index NRER, NDVI, FAI and temperature auxiliary variables;

[0036] S2-2-2-2, each input variable will pass through 5 VGG blocks consisting of 3 consecutive convolutional layers and one maximum pooling layer;

[0037] S2-2-2-3, each convolutional layer is followed by a linear rectification function ReLU function to increase nonlinearity and a maximum pooling layer to reduce the patch size of the output features;

[0038] S2-2-2-4, after the maximum pooling layer processing, the model expands the image into a one-dimensional vector and classifies it through three fully connected layers;

[0039] S2-2-2-5, in the fully connected layer, each neuron is connected to all neurons in the previous layer to calculate the weighted sum and perform nonlinear mapping through the ReLU function;

[0040] S2-2-2-6, the last layer converts the output of the convolution block into the probability of each category, and the loss function is the cross entropy loss, and its formula is:

[0041]

[0042] Among them, N is the number of training data for one iteration, M is the number of categories, and y ic is the label of sample i, p ic is the probability that sample i is correctly classified as category c;

[0043] S2-2-3, sample preparation and training:

[0044] Select training samples and validation samples from satellite images based on historical survey records and visual reflectance spectra;

[0045] The selected sample pixels corresponding to different water bloom types are used as center pixels to generate sample points for model input; 70% of the samples are randomly selected from 2 / 3 of the image frames in S1-2 for model training; the remaining 30% of the samples are randomly selected from 1 / 3 of the image frames in S1-2 for accuracy evaluation;

[0046] S3, accuracy verification:

[0047] S3-1, accuracy verification indicators:

[0048]

[0049] UA=p jj / p j+ ;

[0050] PA=p jj / p +j ;

[0051] Among them, OA is the overall accuracy, UA is the user accuracy, PA is the producer accuracy, and p jj is the proportion of pixels in the jth layer that are correctly classified, p j+ is the ratio of correctly classified pixels to the total number of pixels in the jth layer, p +j is the ratio of correctly classified pixels to the total number of corresponding pixels;

[0052] S3-2, accuracy verification results: Use the test set to verify the results.

[0053] The remote sensing satellite surface temperature product in S1-4 has a spatial resolution of 30m.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention establishes an algal bloom detection model by utilizing multispectral images from remote sensing satellites, combines the classification of algal bloom characteristics and types, and performs progress training and verification of the images through a deep learning model, thereby improving the overall recognition accuracy of algal blooms in water bodies. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention;

[0057] Figure 2 Schematic diagram of identifying the water body of Taihu Lake in an embodiment of the present invention; DETAILED DESCRIPTION

[0058] The present invention will now be further described with reference to the accompanying drawings.

[0059] See also Figures 1-2 The present invention provides a high-resolution algal bloom recognition method based on deep learning, comprising the following steps:

[0060] S1, data download and processing:

[0061] S1-1, the Sentinel-2 multispectral imager images were downloaded and used from the Copernicus Open Access Hub (https: / / scihub.copernicus.eu), where four 10m bands and five 20m bands were used to monitor algal blooms in rivers;

[0062] S1-2, we used Taihu Lake, the third largest freshwater lake in China, as a case area for the experiment, and further downloaded 42 images with cloud cover less than 10% from 2017 to 2021, of which 32 images were used to construct a training set to establish an algal bloom detection model, and the remaining 10 images were used to test the model performance;

[0063] S1-3, using the field monitoring data of algal cell density and chlorophyll a concentration of the Ministry of Ecology and Environment to classify algal bloom types;

[0064] S1-4, the surface temperature product (30 m spatial resolution) of Landsat8 (the 8th satellite of the Landsat program) is used as auxiliary data for identifying water blooms;

[0065] S1-5, generate the remote sensing reflectance of the water surface through atmospheric correction, and then use the Fmask (Function ofmask) mask function to remove clouds and cloud shadows;

[0066] S1-6, use the Normalized Difference Water Index (NDWI) algorithm to extract water areas and exclude the two pixels adjacent to the edge of the water body:

[0067] NDWI = (G-NIR) / (G+NIR);

[0068] Among them, G is the green light band, NIR is the near infrared band, and the threshold of NDWI is 0;

[0069] S1-7, resample the Landsat8 (the 8th satellite of the Landsat program) surface temperature product in S1-4 to a spatial resolution of 20 m;

[0070] S2, identification of the spatial extent of water blooms:

[0071] S2-1, Analysis of water bloom characteristics:

[0072] S2-1-1, Classification of algal bloom types:

[0073] The bloom types are divided into three categories: no bloom, mild bloom and severe bloom;

[0074] According to the historical survey records of water blooms in Taihu Lake, the classification standards for water blooms are as follows:

[0075] No water bloom: when the algae cell density is less than 2.0×106 / L and the chlorophyll a concentration is less than 10μg / L;

[0076] Mild algal bloom: when the algae cell density is between 2.0-10×106 / L and the chlorophyll a concentration is between 10-50μg / L;

[0077] Severe algal bloom: when the algae cell density exceeds 1.0×107 / L and the chlorophyll a concentration exceeds 50μg / L;

[0078] S2-1-2, Construction of water bloom spectrum index:

[0079] According to prior knowledge, the two spectral features of the local maximum reflectance in the visible light band and the reflection peak at 705nm in the red edge channel are the key parameters for detecting water blooms. Therefore, the condition of water blooms is determined based on their spectral shape and reflectance:

[0080]

[0081] Among them, NRER is the algae bloom spectral index, and are the normalized values ​​of the red edge band and the red light band respectively;

[0082] Normalized difference vegetation index NDVI and floating algae index FAI:

[0083]

[0084]

[0085] in, and are the surface reflectance values ​​in the near infrared and short-wave infrared bands, respectively, red , NIR and λ SWIR They are the wavelengths of the red band, near infrared band, and short-wave infrared band;

[0086] S2-2, Construction of deep learning model for water bloom identification:

[0087] S2-2-1, Architecture design: Based on the convolutional neural network, a water bloom recognition model consisting of 13 convolutional layers and 3 fully connected layers was designed according to the structure of the VGG (Visual Geometry Group-16) network;

[0088] S2-2-2, network structure:

[0089] S2-2-2-1, input variables include 9 spectral bands of remote sensing satellite images, spectral index NRER, NDVI, FAI and temperature auxiliary variables;

[0090] S2-2-2-2, each input variable will pass through 5 VGG blocks consisting of 3 consecutive convolutional layers and one maximum pooling layer;

[0091] S2-2-2-3, each convolutional layer is followed by a linear rectification function ReLU function to increase nonlinearity and a maximum pooling layer to reduce the patch size of the output features;

[0092] S2-2-2-4, after the maximum pooling layer processing, the model expands the image into a one-dimensional vector and classifies it through three fully connected layers;

[0093] S2-2-2-5, in the fully connected layer, each neuron is connected to all neurons in the previous layer to calculate the weighted sum and perform nonlinear mapping through the ReLU function;

[0094] S2-2-2-6, the last layer converts the output of the convolution block into the probability of each category, and the loss function is the cross entropy loss, and its formula is:

[0095]

[0096] Among them, N is the number of training data for one iteration, M is the number of categories, and y ic is the label of sample i, p ic is the probability that sample i is correctly classified as category c;

[0097] S2-2-3, sample preparation and training:

[0098] Based on historical survey records and visual reflectance spectra, training samples and validation samples were selected from Sentinel-2 satellite images;

[0099] The selected sample pixels corresponding to different water bloom types are used as center pixels to generate sample points for model input; in this embodiment, 760 sample points are produced, including 165 water bloom-free sample points, 325 light water bloom sample points and 270 heavy water bloom sample points, of which 70% of the samples are randomly selected from 32 images used for training for model training; the remaining 30% of the samples are randomly selected from 10 images used for testing for accuracy evaluation.

[0100] S3, accuracy verification:

[0101] S3-1, accuracy verification indicators:

[0102]

[0103] UA=p jj / p j+ ;

[0104] PA=p jj / p +j ;

[0105] Among them, OA is the overall accuracy, UA is the user accuracy, PA is the producer accuracy, and p jj is the proportion of pixels in the jth layer that are correctly classified, p j+ is the ratio of correctly classified pixels to the total number of pixels in the jth layer, p +j is the ratio of correctly classified pixels to the total number of corresponding pixels;

[0106] S3-2, accuracy verification results: The results were verified using the test set: the accuracy of the present invention was between 91-96%, and the integral accuracy was 93.4%, showing high applicability, as shown in the following table:

[0107]

[0108] The above are only preferred embodiments of the present invention, which are only used to help understand the method and core idea of ​​the present application. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the idea of ​​the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

[0109] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0110] The present invention solves the problems of delayed recognition and low efficiency caused by manual recognition of algal blooms in the prior art. An algal bloom detection model is established by utilizing multispectral images from remote sensing satellites, and combined with the classification of algal bloom characteristics and types, the images are trained and verified through a deep learning model. This improves the overall recognition accuracy of algal blooms in water bodies, and provides an efficient, accurate and scientific basis for the recognition, prevention and timely treatment of algal blooms in water bodies.

Claims

1. A high-resolution algal bloom recognition method based on deep learning, characterized in that: The following steps are involved: S1, data download and processing: S1-1, monitoring algal blooms in rivers using four 10-m bands and five 20-m bands of the multispectral imager on the remote sensing satellite Sentinel-2; S1-2, download the images with cloud cover less than 10% in previous years, of which 2 / 3 images are used to construct the training set to establish the water bloom detection model, and the remaining 1 / 3 images are used to test the model performance; S1-3, using the field monitoring data of algal cell density and chlorophyll a concentration of the Ministry of Ecology and Environment to classify algal bloom types; S1-4, using remote sensing satellite surface temperature products as auxiliary data for identifying algal blooms; S1-5, generate the water surface remote sensing reflectance through atmospheric correction, and then use the Fmask mask function to remove clouds and cloud shadows; S1-6, use the Normalized Difference Water Index (NDWI) algorithm to extract water areas and exclude the two pixels adjacent to the edge of the water body: NDWI = (G-NIR) / (G+NIR); Among them, G is the green light band, NIR is the near infrared band, and the threshold of NDWI is 0; S1-7, resample the remote sensing satellite surface temperature product in S1-4 to a spatial resolution of 20 m; S2, identification of the spatial extent of water blooms: S2-1, Analysis of water bloom characteristics: S2-1-1, Classification of algal bloom types: The bloom types are divided into three categories: no bloom, mild bloom and severe bloom; According to the historical survey records of water blooms, the classification standards for water blooms are: No water bloom: When the algae cell density is less than 2.0×10 6 / L and the chlorophyll a concentration is less than 10μg / L; Mild algal bloom: When the algae cell density is between 2.0-10×10 6 / L and the chlorophyll a concentration is in the range of 10-50μg / L; Severe algal bloom: When the density of algae cells exceeds 1.0×10 7 / L and the chlorophyll a concentration exceeds 50μg / L; S2-1-2, construction of water bloom spectrum index: The condition of the bloom is determined by its spectral shape and reflectivity: Among them, NRER is the algae bloom spectral index, and They are the surface reflectance values ​​of the red edge band and the red light band respectively; Normalized difference vegetation index NDVI and floating algae index FAI: in, and are the surface reflectance values ​​in the near infrared and short-wave infrared bands, respectively, red , NIR and λ SWIR They are the wavelengths of the red band, near infrared band, and short-wave infrared band; S2-2, Construction of deep learning model for water bloom identification: S2-2-1, Architecture design: Based on the convolutional neural network, a water bloom recognition model consisting of 13 convolutional layers and 3 fully connected layers was designed according to the structure of VGG (Visual Geometry Group-16) network; S2-2-2, network structure: S2-2-2-1, input variables include 9 spectral bands of remote sensing satellite images, spectral index NRER, NDVI, FAI and temperature auxiliary variables; S2-2-2-2, each input variable will pass through 5 VGG blocks; S2-2-2-3, each convolutional layer is followed by a linear rectification function ReLU function to increase nonlinearity; S2-2-2-4, after the maximum pooling layer processing, the model expands the image into a one-dimensional vector and classifies it through 3 fully connected layers; S2-2-2-5, in the fully connected layer, each neuron is connected to all neurons in the previous layer to calculate the weighted sum and perform nonlinear mapping through the ReLU function; S2-2-2-6, the last layer converts the output of the convolution block into the probability of each category, and the loss function is the cross entropy loss, and its formula is: Among them, N is the number of training data for one iteration, M is the number of categories, and y ic is the label of sample i, p ic is the probability that sample i is correctly classified as category c; S2-2-3, sample preparation and training: Select training samples and validation samples from satellite images based on historical survey records and visual reflectance spectra; The selected sample pixels corresponding to different water bloom types are used as center pixels to generate sample points for model input; 70% of the samples are randomly selected from 2 / 3 of the image frames in S1-2 for model training; The remaining 30% of samples are randomly selected from 1 / 3 of the images in S1-2 for accuracy evaluation; S3, accuracy verification: S3-1, accuracy verification through error matrix statistical calculation; S3-2, Accuracy verification results: Use the test set to verify the results.

2. The high-resolution algal bloom recognition method based on deep learning according to claim 1, characterized in that: The remote sensing satellite surface temperature product in S1-4 has a spatial resolution of 30m.

Images