Lake remote sensing image algae bloom coverage degree inversion method based on machine learning

Abstract

The invention relates to a lake remote sensing image algae bloom coverage degree inversion method based on machine learning. Compared with the prior art, the defect that algae bloom coverage degree inversion is difficult to carry out on lake remote sensing images is overcome. The method comprises the following steps: acquiring and preprocessing a remote sensing data source; carrying out algae bloom extraction and processing on the Sentinel-2 multispectral image to obtain a coverage sample set; performing feature selection and combination on the MODIS multispectral image to obtain a feature index data set; constructing a lake algae bloom coverage degree data set; constructing a model and training based on machine learning; and obtaining an inversion result of the algae bloom coverage. According to the method, the algae bloom coverage degree of the MODIS pixel scale can be obtained by utilizing Sentinel-2 data with high spatial resolution, and the quantitative estimation model is constructed by combining MODIS spectral characteristics, so that high-temporal-spatial-precision monitoring and change analysis of the lake algae bloom are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a method for inverting algal bloom coverage of lake remote sensing images based on machine learning. Background Art

[0002] Eutrophication is a water pollution phenomenon. The main reason for its occurrence is the excessive content of nutrients such as N and P in the water, which causes the imbalance of nutrient input and output. This will affect the aquatic ecosystem, leading to the crazy growth of a single species and the imbalance of species distribution in the aquatic ecosystem. This phenomenon often occurs in areas with slow water flow such as lakes and bays.

[0003] The study of algal bloom coverage is an important task to monitor and analyze the density and distribution of algae growth in water bodies. It is of great significance to ecological environmental monitoring, water body management and protection. The corresponding research can be used to monitor the ecological and environmental status of water bodies. The growth and distribution of algal blooms are affected by many factors such as water quality, nutrients, temperature, and light. Therefore, algal bloom coverage can be used as one of the indicators of the health of water ecosystems. High coverage may indicate that the water body is affected by pollution or excessive nutrient input, and measures need to be taken to improve water quality. Through research, problem areas in water bodies can be identified, and the formulation and implementation of relevant management measures can be guided to reduce pollution, improve water quality, and protect water ecosystems. It plays a role in maintaining the health of water ecosystems, ensuring drinking water safety, promoting sustainable water resources management, and providing data for scientific research.

[0004] At present, there are three technical difficulties in the inversion of algal bloom coverage: First, the spectral characteristics of algal blooms are easily confused with those of other substances such as water bodies and floating leaf plants. In order to accurately separate algal blooms, it is usually necessary to rely on hyperspectral data or fine band calculations (such as NDVI, phytoplankton index, etc.). The complex conditions of different water bodies also make spectral separation more difficult; second, the inversion of algal bloom coverage usually uses machine learning algorithms (such as random forests, support vector machines, XGboost, etc.) or physical models (such as radiation transfer models). Different algorithms and models have different sensitivities to data. Choosing a suitable algorithm and tuning parameters have a great impact on the inversion accuracy; third, the spatial and temporal dynamics of algal blooms are very variable, and the characteristics of algal blooms in different regions and seasons are significantly different. Establishing a model with good generalization ability that can adapt to changes in different waters and time is also one of the difficulties in the inversion work.

[0005] Therefore, how to leverage the advantages of satellite remote sensing based on multi-source satellites supplemented by ground surveys, and gradually achieve accurate estimation of long-term algal bloom coverage based on high spatial resolution remote sensing data through multiple remote sensing data and machine learning methods has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the present invention is to solve the defect in the prior art that it is difficult to invert algal bloom coverage for lake remote sensing images, and to provide a method for inverting algal bloom coverage of lake remote sensing images based on machine learning to solve the above problem.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for inverting algal bloom coverage from lake remote sensing images based on machine learning includes the following steps:

[0009] Acquire remote sensing data sources and perform preprocessing: obtain long-term Sentinel-2 and MODIS images of the lake, perform radiation correction and denoising preprocessing on the images, and obtain preprocessed Sentinel-2 multispectral images and MODIS multispectral images;

[0010] Extract and process algal blooms from Sentinel-2 multispectral images to obtain a coverage sample set;

[0011] Feature selection and combination of MODIS multispectral images are performed to obtain feature index datasets;

[0012] Construct lake algal bloom coverage dataset: perform pixel and spatial matching on the obtained coverage sample set and characteristic index dataset to obtain a MODIS satellite-scale algal bloom coverage dataset that matches the Sentinel-2 satellite pixels;

[0013] Model building and training based on machine learning: Based on XGBoost machine learning, a satellite-scale algal bloom coverage model was built. The algal bloom coverage dataset was divided into a training set for model training and a validation set for accuracy verification in a ratio of 7:3. The training set was input into the algal bloom coverage model for training.

[0014] Obtaining the inversion results of algal bloom coverage: By inputting 63 index data of MODIS images on different dates into the trained algal bloom coverage model, the inversion results of algal bloom coverage on different dates of the lake were obtained.

[0015] The extraction and processing of algal blooms from Sentinel-2 multispectral images to obtain a coverage sample set includes the following steps:

[0016] Use the FAI index to separate the water cover area:

[0017] When the FAI of a pixel is greater than FAI_t, the pixel is judged as an overwater area, where FAI_t is the universal threshold of FAI. FAI_t is obtained by the maximum gradient histogram statistics of the selected image. The formula is as follows:

[0018] FAI_t=μ-2σ

[0019] Among them, μ is the mean of the maximum gradient histogram, σ is the standard deviation of the maximum gradient histogram;

[0020] TWI is used to remove the influence of high turbidity water on the classification of water cover. When FAI>0.005 and TWI>0.107, the pixel is determined as turbid water. The calculation formula of TWI is:

[0021] TWI=ρ R -ρ SWIR

[0022] Among them, ρ R and ρ SWIR are the reflectances of the red band and the short-wave infrared band, respectively;

[0023] The above-water cover area includes algal blooms and floating leaf vegetation, and the segmentation threshold is obtained by the Otsu algorithm;

[0024] The principal component analysis method PCA and MNDWI are combined to separate algal blooms and floating leaf vegetation in the water cover area. The segmentation threshold is obtained by the Otsu algorithm. The Otsu algorithm divides the gray value of the image into foreground and background, and then calculates all possible thresholds to maximize the inter-class variance between the two categories. The maximum inter-class variance means that the difference between the two categories is the largest, thus obtaining an optimal threshold.

[0025] Assuming the between-class variance is v, the calculation formula of the between-class variance is:

[0026] v(t)=ω0(t)×ω1(t)×[μ0(t)-μ1(t)] 2 ,

[0027] Where t is the current segmentation threshold, ω0(t) and ω1(t) are the probabilities of the image being segmented into foreground and background pixels by the threshold t, respectively, which are calculated by the histogram, and μ0(t) and μ1(t) are the average grayscale values ​​of these two types of pixels, respectively;

[0028] The goal of the Otsu algorithm is to find a threshold t that maximizes v(t) and calculate the segmentation threshold PC2 in each image. t ;

[0029] A classification decision tree is constructed:

[0030] When FAI<0.005 and TWI>0.107, the water cover area is separated, and based on this threshold, the water cover area is divided into two categories, C1 and C2. Then the MNDWI means of all pixels in C1 and C2 are calculated, which are M1 and M2 respectively; by calculating the universal threshold M_t of the MNDWI mean, the category corresponding to M1 and M2 greater than M_t is algal bloom, otherwise it is floating leaf vegetation.

[0031] The feature selection and combination of MODIS multispectral images to obtain a feature index data set includes the following steps:

[0032] Spectral index based on satellite remote sensing data: Seven spectral indices based on MODIS data were selected, namely NDVI, FAI, AFAI, NDAVI, NDWI, RVI, and RDVI;

[0033] Texture feature parameters based on satellite remote sensing data: 8 texture features of MODIS data band 1 to band 7 are selected, namely Mean, Contrast, Variance, Homogeneity, Dissimilarity, Entropy, Second Moment, and Correlation.

[0034] Beneficial Effects

[0035] Compared with the prior art, the method for inverting algal bloom coverage from lake remote sensing images based on machine learning can obtain algal bloom coverage at the MODIS pixel scale by using Sentinel-2 data with high spatial resolution, and build a quantitative estimation model in combination with MODIS spectral characteristics, thereby realizing high temporal and spatial precision monitoring and change analysis of lake algal blooms.

[0036] The present invention uses Sentinel-2 images and MODIS images as research objects, effectively solving the problem that considering the rapid spatiotemporal changes of algal blooms, the spatiotemporal changes of lake algal blooms are usually monitored using MODIS satellites with high temporal resolution, but due to the low spatial resolution, only "presence or absence" can be judged at the pixel scale, and the area and coverage of algal blooms in the pixel cannot be quantitatively obtained; the method constructs an automatic identification algorithm for algal blooms in eutrophic lakes based on the second principal component of principal component analysis and widely carries out accuracy verification and robustness evaluation, which has the advantages of simple calculation and good effect compared with traditional image processing algorithms; the method obtains the MODIS pixel-scale algal bloom coverage data set from the high spatial resolution Sentinel satellite, and combines the pixel spectral feature data set of the MODIS satellite to construct a quantitative coverage estimation model. Compared with the traditional method of inverting algal bloom coverage, the method can quantitatively obtain algal bloom coverage information in a long time series. It can help solve the problem that it is difficult to obtain algal bloom coverage data before the satellite cycle arrives, which also provides a new idea for algal bloom inversion at the regional and even national scales.

[0037] Compared with the traditional method of inverting algal bloom coverage, the present invention can timely, quickly and accurately achieve the inversion results of lake algal bloom coverage. At the same time, the algorithm in the present invention has the advantages of short running time and high inversion accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a method sequence diagram of the present invention;

[0039] Figure 2 A schematic diagram of the sample structure for obtaining algal bloom coverage that matches the MODIS pixel scale in the present invention;

[0040] Figure 3 This is a diagram of the algal bloom inversion results of West Taihu Lake achieved using the method described in the present invention. DETAILED DESCRIPTION

[0041] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0042] like Figure 1 and Figure 2As shown, the present invention describes a method for inverting algal bloom coverage from lake remote sensing images based on machine learning, which aims to solve the problem that it is difficult to quantitatively evaluate algal bloom coverage in traditional methods. Although traditional remote sensing technologies such as MODIS have high temporal resolution, their spatial resolution is low and they cannot accurately reflect the distribution of algal blooms within pixels. By introducing machine learning technology and combining it with high spatial resolution Sentinel-2 data, the present invention can accurately extract algal bloom coverage information. The technical significance of this design is to improve the spatial and quantitative accuracy of algal bloom monitoring, which helps to more accurately analyze the health of lake water bodies.

[0043] The first technical difficulty lies in the fact that the spectral characteristics of algal blooms are difficult to distinguish from those of other substances such as water bodies and floating leaf plants; the second is that different algorithms and models have different sensitivities to data, and it is necessary to select appropriate algorithms and perform parameter tuning; the third is that the spatial and temporal dynamics of algal blooms are very variable, and the characteristics of algal blooms in different regions and seasons are significantly different.

[0044] It includes the following steps:

[0045] The first step is to obtain the remote sensing data source and perform preprocessing: obtain the long-term Sentinel-2 and MODIS images of the lake, perform radiation correction and denoising preprocessing on the images, and obtain the preprocessed Sentinel-2 multispectral images and MODIS multispectral images.

[0046] The second step is to extract and process algal blooms from Sentinel-2 multispectral images to obtain coverage sample sets. Sentinel-2 images have a high spatial resolution and can more clearly capture the detailed distribution of algal blooms in lakes. The combined analysis of multispectral bands helps to distinguish algal blooms from other landforms such as water bodies and vegetation, and to accurately extract algal bloom areas. The technical significance of this step is that Sentinel-2 is a satellite with high spatial resolution. The coverage sample set generated using high-resolution data can provide accurate algal bloom coverage references for the training of subsequent machine learning models. This not only improves the accuracy and reliability of the model, but also provides a solid foundation for large-scale, high-temporal and spatial resolution algal bloom monitoring and prediction, which helps promote the refined management and protection of the lake's ecological environment.

[0047] (1) Using the FAI index to separate the water cover areas:

[0048] When the FAI of a pixel is greater than FAI_t, the pixel is judged as an overwater area, where FAI_t is the universal threshold of FAI. FAI_t is obtained by the maximum gradient histogram statistics of the selected image. The formula is as follows:

[0049] FAI_t=μ-2σ

[0050] Among them, μ is the mean of the maximum gradient histogram, σ is the standard deviation of the maximum gradient histogram;

[0051] TWI is used to remove the influence of high turbidity water on the classification of water cover. When FAI>0.005 and TWI>0.107, the pixel is determined as turbid water. The calculation formula of TWI is:

[0052] TWI=ρ R -ρ SWIr

[0053] Among them, ρ R and ρ SWIR are the reflectances of the red band and the short-wave infrared band, respectively.

[0054] (2) The above-water cover area includes algal blooms and floating leaf vegetation, and the segmentation threshold is obtained by the Otsu algorithm;

[0055] The principal component analysis method PCA and MNDWI are combined to separate the algal bloom and floating leaf vegetation in the water cover area. The segmentation threshold is obtained by the Otsu algorithm. The Otsu algorithm divides the gray value of the image into foreground and background, and then calculates all possible thresholds to maximize the inter-class variance between the two classes. The maximum inter-class variance means that the difference between the two classes is the largest, thus obtaining an optimal threshold.

[0056] Assuming the between-class variance is ν, the calculation formula of the between-class variance is:

[0057] ν(t)=ω0(t)×ω1(t)×[μ0(t)-μ1(t)] 2 ,

[0058] Where t is the current segmentation threshold, ω0(t) and ω1(t) are the probabilities of the image being segmented into foreground and background pixels by the threshold t, respectively, which are calculated by the histogram, and μ0(t) and μ1(t) are the average grayscale values ​​of these two types of pixels, respectively;

[0059] The goal of the Otsu algorithm is to find a threshold t that maximizes v(t) and calculate the segmentation threshold PC2 in each image. t .

[0060] (3) Constructed a classification decision tree:

[0061] When FAI<0.005 and TWI>0.107, the water cover area is separated, and based on this threshold, the water cover area is divided into two categories, C1 and C2. Then the MNDWI means of all pixels in C1 and C2 are calculated, which are M1 and M2 respectively; by calculating the universal threshold M_t of the MNDWI mean, the category corresponding to M1 and M2 greater than M_t is algal bloom, otherwise it is floating leaf vegetation.

[0062] The third step is to select and combine features of MODIS multispectral images to obtain a feature index dataset.

[0063] Although MODIS images have a low spatial resolution, their high temporal resolution and wide-area coverage provide strong support for long-term, large-scale monitoring of algal blooms. By performing feature selection and combination on the multispectral bands of MODIS images, spectral features related to algal blooms can be extracted and a characteristic index data set can be constructed. The technical significance of this step is that by selecting features with high correlation with algal bloom coverage, the sensitivity of the model to the distribution of algal blooms can be enhanced. The characteristic index data set provides input variables for the machine learning model, enabling it to effectively combine the time series characteristics of MODIS images and further improve the inversion accuracy of algal bloom coverage. In this way, not only the problem of insufficient accuracy or temporal coverage of a single data source is solved, but also the model's ability to adapt to the dynamic changes of large-scale algal blooms is enhanced.

[0064] (1) Spectral indices based on satellite remote sensing data: Seven spectral indices based on MODIS data were selected, namely NDVI, FAI, AFAI, NDAVI, NDWI, RVI, and RDVI.

[0065] (2) Texture feature parameters based on satellite remote sensing data: Eight texture features of MODIS data bands 1 to 7 are selected, namely Mean, Contrast, Variance, Homogeneity, Dissimilarity, Entropy, Second Moment, and Correlation.

[0066] The fourth step is to construct a lake algal bloom coverage dataset: the obtained coverage sample set and characteristic index dataset are matched pixel by pixel and space to obtain a MODIS satellite-scale algal bloom coverage dataset that matches the Sentinel-2 satellite pixels.

[0067] The fifth step is to build a model and conduct training based on machine learning: Based on XGBoost machine learning, a satellite-scale algal bloom coverage model is built. The algal bloom coverage dataset is divided into a training set for model training and a validation set for accuracy verification in a ratio of 7:3. The training set is input into the algal bloom coverage model for training.

[0068] Step 6. Obtaining the inversion results of algal bloom coverage: By inputting 63 index data of MODIS images on different dates into the trained algal bloom coverage model, the inversion results of algal bloom coverage on different dates of the lake are obtained.

[0069] like Figure 3As shown in the figure, it shows the inversion effect of algal bloom coverage on August 17, 2019. Figure 3 It can be seen that the algal bloom of West Taihu Lake on that day was mainly concentrated in some areas along the western coast, which can clearly reflect the spatial distribution characteristics of the algal bloom, especially in the high coverage area of ​​nearshore waters. The inversion effect of the algal bloom verifies the advantages of the present invention in high-precision spatial monitoring and can provide important spatiotemporal information for lake ecological management.

[0070] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for inverting algal bloom coverage from lake remote sensing images based on machine learning, characterized in that: The following steps are involved: 11) Obtain remote sensing data sources and perform preprocessing: obtain long-term Sentinel-2 and MODIS images of the lake, perform radiation correction and denoising preprocessing on the images, and obtain preprocessed Sentinel-2 multispectral images and MODIS multispectral images; 12) Extract and process algal blooms from Sentinel-2 multispectral images to obtain a coverage sample set; 13) Feature selection and combination of MODIS multispectral images to obtain feature index data sets; 14) Construct lake algal bloom coverage dataset: perform pixel and spatial matching on the obtained coverage sample set and characteristic index dataset to obtain a MODIS satellite-scale algal bloom coverage dataset that matches the Sentinel-2 satellite pixels; 15) Model building and training based on machine learning: Based on XGBoost machine learning, a satellite-scale algal bloom coverage model was built. The algal bloom coverage dataset was divided into a training set for model training and a validation set for accuracy verification in a ratio of 7:

3. The training set was input into the algal bloom coverage model for training. 16) Obtaining the inversion results of algal bloom coverage: By inputting 63 index data of MODIS images on different dates into the trained algal bloom coverage model, the inversion results of algal bloom coverage on different dates of the lake were obtained.

2. The method for inverting algal bloom coverage from lake remote sensing images based on machine learning according to claim 1 is characterized in that: The extraction and processing of algal blooms from Sentinel-2 multispectral images to obtain a coverage sample set includes the following steps: 21) Use the FAI index to separate the water cover areas: When the FAI of a pixel is greater than FAI_t, the pixel is judged as an overwater area, where FAI_t is the universal threshold of FAI. FAI_t is obtained by the maximum gradient histogram statistics of the selected image. The formula is as follows: FAI_t=μ-2σ Among them, μ is the mean of the maximum gradient histogram, σ is the standard deviation of the maximum gradient histogram; TWI is used to remove the influence of high turbidity water on the classification of water cover. When FAI>0.005 and TWI>0.107, the pixel is determined as turbid water. The calculation formula of TWI is: TWI=ρ R -ρ SWIR Among them, ρ R and ρ SWIR are the reflectances of the red band and the short-wave infrared band, respectively; 22) The above-water cover area includes algal blooms and floating leaf vegetation, and the segmentation threshold is obtained by the Otsu algorithm; The principal component analysis method PCA and MNDWI are combined to separate algal blooms and floating leaf vegetation in the water cover area. The segmentation threshold is obtained by the Otsu algorithm. The Otsu algorithm divides the gray value of the image into foreground and background, and then calculates all possible thresholds to maximize the inter-class variance between the two categories. The maximum inter-class variance means that the difference between the two categories is the largest, thus obtaining an optimal threshold. Assuming the between-class variance is v, the calculation formula of the between-class variance is: v(t)=ω0(t)×ω1(t)×[μ0(t)-μ1(t)] 2 , Where t is the current segmentation threshold, ω0(t) and ω1(t) are the probabilities of the image being segmented into foreground and background pixels by the threshold t, respectively, which are calculated by the histogram, and μ0(t) and μ1(t) are the average grayscale values ​​of these two types of pixels, respectively; The goal of the Otsu algorithm is to find a threshold t that maximizes v(t) and calculate the segmentation threshold PC2 in each image. t ; 23) A classification decision tree was constructed: When FAI<0.005 and TWI>0.107, the water cover area is separated, and based on this threshold, the water cover area is divided into two categories, C1 and C2. Then the MNDWI means of all pixels in C1 and C2 are calculated, which are M1 and M2 respectively; by calculating the universal threshold M_t of the MNDWI mean, the category corresponding to M1 and M2 greater than M_t is algal bloom, otherwise it is floating leaf vegetation.

3. The method for inverting algal bloom coverage from lake remote sensing images based on machine learning according to claim 1 is characterized in that: The feature selection and combination of MODIS multispectral images to obtain a feature index data set includes the following steps: 31) Spectral index based on satellite remote sensing data: Seven spectral indices based on MODIS data were selected, namely NDVI, FAI, AFAI, NDAVI, NDWI, RVI, and RDVI; 32) Texture feature parameters based on satellite remote sensing data: 8 texture features of MODIS data band 1 to band 7 are selected, namely Mean, Contrast, Variance, Homogeneity, Dissimilarity, Entropy, Second Moment, and Correlation.

Images