A water plant and algal bloom identification method based on spectrum and relative temperature fusion

Abstract

A water vegetation and algal bloom recognition method based on spectrum and relative temperature fusion, a plurality of remote sensing data sets are obtained first, then a six-dimensional feature vector F is obtained by splicing NDVI index, NDWI index, FAI index, GR ratio index, CMI index and normalized absolute lake surface temperature, then the classifier is trained according to the six-dimensional feature vector F and the label of the remote sensing data set, then the classifier including the recognition rule set of water vegetation and algal bloom is obtained; the six-dimensional feature vector F of the remote sensing data set to be recognized is applied to the classifier including the recognition rule set of water vegetation and algal bloom to output the recognition result of water vegetation and algal bloom; six indexes are used for recognition, the six indexes not only verify each other, but also complement each other, and the normalized absolute lake surface temperature feature is introduced to enhance the recognition accuracy. Therefore, the design uses multiple indexes for recognition, and the recognition accuracy is high.

Description

Technical Field

[0001] This invention relates to a method for identifying aquatic vegetation and algal blooms, belonging to the field of remote sensing image processing and environmental monitoring technology, and particularly to a method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature. Background Technology

[0002] Algal blooms can deteriorate lake water quality and threaten ecological security, while aquatic vegetation (including submerged plants, emergent plants, and floating-leaved plants) can purify water quality and inhibit algal growth, which is key to maintaining the health of lakes. Therefore, it is necessary to identify both to complete the ecological monitoring, water quality assessment, or environmental governance decision-making of lakes.

[0003] Chinese patent application number 201510074116.4, filed on February 11, 2015, discloses a remote sensing extraction method for aquatic vegetation in eutrophic water bodies based on the algal bloom frequency method. The method includes: S1, selecting MODIS images from several time periods where aquatic vegetation is lush and its area is stable in eutrophic waters; S2, calculating the Facilitation Index (FAI) of the selected MODIS images; S3, calculating the vegetation signal occurrence frequency (VPF) at each pixel in several time periods based on the FAI; and S4, defining VPF thresholds for all time periods and calculating the aquatic vegetation area (Aav). Although this design can identify aquatic vegetation and algal blooms, it still has the following drawbacks:

[0004] The design uses only the FAI index to calculate the threshold for evaluation. However, although the single FAI index can separate water bodies, aquatic vegetation and algal blooms to a certain extent, it is easily affected by water turbidity, bottom sediment reflection and differences in vegetation type. It is difficult to distinguish between planktonic algal blooms and different types of aquatic plants at the same time, so the recognition accuracy of this design is not high.

[0005] The information disclosed in this background section is intended only to enhance understanding of the overall background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and problems of existing technologies that use a single index for identification and have low identification accuracy, and to provide a method for identifying aquatic vegetation and algal blooms based on the fusion of spectrum and relative temperature that uses multiple indices for identification and has higher identification accuracy.

[0007] To achieve the above objectives, the technical solution of the present invention is:

[0008] A method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature, the method comprising the following steps:

[0009] Step 1: First, obtain multiple remote sensing datasets, then annotate the aquatic vegetation and algal blooms in the multiple remote sensing datasets, and finally obtain multiple annotated samples that correspond one-to-one with the remote sensing datasets;

[0010] The second step involves sequentially acquiring vectors from multiple remote sensing datasets. This process involves calculating the NDVI, NDWI, FAI, GR ratio, and CMI indices pixel by pixel based on a single remote sensing dataset, and then obtaining the normalized absolute lake surface temperature from that dataset. The NDVI, NDWI, FAI, GR ratio, CMI, and normalized absolute lake surface temperatures are then concatenated to obtain the six-dimensional feature vector F of the dataset. Finally, multiple six-dimensional feature vectors F corresponding to each of the multiple remote sensing datasets are obtained.

[0011] The third step is to take multiple six-dimensional feature vectors F, which correspond one-to-one with multiple remote sensing datasets, as inputs, and the labels in the labeled samples as the corresponding inputs of the six-dimensional feature vectors F. Then, the classifier model is trained based on the inputs and corresponding inputs, and then the set of recognition rules for aquatic vegetation and algal blooms is obtained. Finally, a classifier model including the set of recognition rules for aquatic vegetation and algal blooms is obtained.

[0012] Step 4: First, perform vector acquisition on the remote sensing dataset to be identified, then obtain the six-dimensional feature vector F to be identified. Next, apply a classifier model that includes the set of identification rules for aquatic vegetation and algal blooms to the six-dimensional feature vector F to be identified, then obtain the annotations for aquatic vegetation and algal blooms, then generate a preliminary classification map, then perform post-processing on the preliminary classification map to improve data accuracy, and finally output the identification results of aquatic vegetation and algal blooms.

[0013] In the third step, the classifier model refers to a random forest model with pre-set parameters.

[0014] In the third step, the parameters refer to: the number of trees is 150 to 200, and the maximum depth is 20 to 30.

[0015] In the first step, the labeled samples include a training set and a validation set; in the third step, the classifier model that obtains the set of recognition rules for aquatic vegetation and algal blooms refers to: a classifier model that has been validated and optimized and includes the set of recognition rules for aquatic vegetation and algal blooms.

[0016] In the third step, the verification refers to: verifying the XGBoost model and the SVM model sequentially; the optimization refers to: optimizing the number of trees and the maximum depth of the random forest model using five-fold cross-validation.

[0017] In the first step, the remote sensing dataset refers to: multiple single-scene remote sensing images are preprocessed to improve data accuracy, and then multiple remote sensing datasets corresponding one-to-one with the multiple single-scene remote sensing images are obtained.

[0018] The preprocessing refers to: first, performing atmospheric correction on the data in a single scene of remote sensing imagery, then obtaining the corrected data, and finally performing geometric registration and cropping on the corrected data.

[0019] In the first step, atmospheric correction of the data in a single scene of remote sensing imagery refers to: performing atmospheric correction on the visible light data, near-infrared data, and short-wave infrared data in the single scene of remote sensing imagery using LEDAPS or LaSRC; then performing brightness-temperature conversion on Band10 of the thermal infrared data in the remote sensing dataset, applying NASA radiometric calibration coefficients and NASA atmospheric correction models for atmospheric correction, and finally obtaining the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data.

[0020] In the first step, the geometric registration and cropping of the corrected data refers to: making the resolution of the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data consistent; then resampling the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data to the same grid; then obtaining the same grid data; then cropping the same grid data; and generating a water mask to retain the water portion.

[0021] In the fourth step, the post-processing refers to: first, performing morphological filtering on the preliminary classification map to obtain a denoised classification map, and then performing multi-scale segmentation and smoothing processing on the denoised classification map.

[0022] In the fourth step, the morphological filtering refers to applying an opening / closing operation to the preliminary classification map to remove isolated noise, and then obtaining a denoised classification map.

[0023] The multi-scale segmentation and smoothing process refers to using SLIC or Mean-Shift superpixel segmentation on the obtained denoised classification map to fuse and smooth the boundaries of the denoised classification map.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. The present invention discloses a method for identifying aquatic vegetation and algal blooms based on spectral and relative temperature fusion. The method includes the following steps: First, obtaining multiple remote sensing datasets, and then obtaining multiple labeled samples including annotations of aquatic vegetation and algal blooms; Second, performing vector acquisition steps on the multiple remote sensing datasets sequentially, and then obtaining a six-dimensional feature vector F composed of NDVI index, NDWI index, FAI index, GR ratio index, CMI index, and normalized absolute lake surface temperature; Third, using the six-dimensional feature vector F as input and the annotations of the labeled samples as corresponding inputs, and then training a classifier model, and then obtaining a classifier model including a set of identification rules for aquatic vegetation and algal blooms; Fourth, performing vector acquisition steps on the remote sensing dataset to be identified, and then obtaining the six-dimensional feature vector F to be identified, and then applying the classifier model including a set of identification rules for aquatic vegetation and algal blooms, and then obtaining the identification result of aquatic vegetation and algal blooms. The advantages of the present invention also include:

[0026] Firstly, this invention can output the identification results of aquatic vegetation and algal blooms;

[0027] Secondly, this invention inputs a six-dimensional feature vector F, composed of NDVI index, NDWI index, FAI index, GR ratio index, CMI index, and normalized absolute lake surface temperature, into the classifier model for identification. These feature vectors can complement each other's information, avoiding misjudgment caused by spectral confusion due to a single index in a specific environment. Moreover, the fusion of these feature vectors can reduce redundancy and form a "mutual verification" effect in classification learning, thus canceling out random noise and threshold uncertainty. This reduces the accumulation of errors and enhances the robustness and generalization ability of the model under different conditions.

[0028] Thirdly, since this invention uses a six-dimensional feature vector F for identification, that is, this invention identifies objects by the feature differences between various classification objects, this invention has cross-temporal and spatial transferability, that is, it can maintain stable identification performance in different lakes, different shooting times and different sensor conditions.

[0029] Fourthly, this invention can identify aquatic vegetation and algal blooms with just a single remote sensing image captured in a single shot. It has the advantages of short time consumption, low cost, large spatial coverage, and the ability to meet large-scale monitoring needs. Furthermore, regular shooting can meet long-term monitoring needs.

[0030] Fifthly, the NDVI, NDWI, FAI, GR ratio, and CMI indices selected in this invention closely match the spectral response characteristics of typical aquatic vegetation in urban lakes, directly reflecting vegetation greenness, water content, and algal pigment absorption characteristics, making identification more targeted. Specifically, the NDVI index is sensitive to emergent plants, the GR index is sensitive to submerged plants, the FAI and CMI indices are sensitive to algal blooms, and the NDWI index is sensitive to water bodies. The combination can cover multiple types of objects, improving the distinguishability. Furthermore, the selected indices cover multiple band combinations of red, green, blue, near-infrared, and short-wave infrared, with low information redundancy and high complementarity, effectively separating the spectral differences of vegetation, water bodies, and algal blooms.

[0031] Sixth point: This invention integrates normalized absolute lake surface temperature characteristics. While the spectral differences between aquatic vegetation and algal blooms are limited, after incorporating normalized absolute lake surface temperature characteristics, the inter-class distance between aquatic vegetation and algal blooms increases and the intra-class distance decreases when using the six-dimensional feature vector F for identification and classification, making it easier to form boundaries during identification and classification. At the same time, normalized absolute lake surface temperature characteristics enable this invention to avoid temperature reference drift caused by different sensors, imaging seasons, and atmospheric conditions, ensuring consistency across different lakes and shooting conditions, thereby significantly improving the stability and cross-regional transferability of classification results.

[0032] Seventh point: In existing technologies, time-series averaging or differential processing is usually performed on images from multiple consecutive days or multiple scenes. However, in multi-temporal or long-term sequence image data, there are intermittent problems caused by cloud cover, inconsistent imaging time, or missing data. This invention only needs to use a single-scene remote sensing image taken once to complete the acquisition of all features, reducing the intermittent problems caused by cloud cover, inconsistent imaging time, or missing data, and significantly improving the completeness and stability of feature acquisition.

[0033] Eighthly, compared to existing technologies, this invention improves overall recognition accuracy by over 12% and increases the Kappa coefficient by 0.15; in complex turbid water bodies and mixed vegetation areas, the accuracy for algal bloom classification is improved from 65% to 89%.

[0034] Therefore, this invention uses multiple indices for identification, and the identification accuracy is high.

[0035] 2. In the method for identifying aquatic vegetation and algal blooms based on spectral and relative temperature fusion of the present invention, the classifier model in the third step refers to a random forest model with 150 to 200 trees and a maximum depth of 20 to 30. The random forest model is validated using XGBoost and SVM models, and then five-fold cross-validation is used to optimize the number of trees and the maximum depth. In application, 150-200 decision trees are generated by the random forest model. A moderate number of trees can reduce the randomness brought by a single tree, and it is not sensitive to noise, resulting in more stable classification results and stronger generalization ability to the six-dimensional feature vector F. The maximum depth of each decision tree is... A tree depth of 20 to 30 allows for the learning of complex spectral differences in aquatic vegetation (such as the difference between shallow, submerged and floating algal blooms), and it can also handle nonlinear relationships and mixed pixel structures. Compared to shallower tree depths, it is more suitable for environments with weak spectral differences and complex backgrounds, such as urban lakes. Overall, the model outputs more accurate recognition rules, clearer inter-class boundaries, and stronger small target recognition capabilities. The random forest model was validated using XGBoost and SVM models, and the number of trees and maximum depth were optimized through five-fold cross-validation to ensure good robustness, stability, reliability, and transferability. Therefore, the accuracy of this invention is high.

[0036] 3. In the method for identifying aquatic vegetation and algal blooms based on spectral and relative temperature fusion of the present invention, in the first step, the remote sensing dataset is obtained by preprocessing single-scene remote sensing images. The preprocessing refers to: firstly, performing atmospheric correction on the data in the single-scene remote sensing image to obtain the corrected data; then, performing geometric registration and cropping on the corrected data. In application, atmospheric correction can eliminate interference factors in the atmosphere, preventing subsequent data inaccuracies due to atmospheric errors. Geometric registration and cropping unify the spatial resolution and positional accuracy of the data, and also crop areas unrelated to water bodies, reducing the computational burden and allowing subsequent classifier model analysis to focus on water bodies. Therefore, the preprocessing effect of the present invention is good.

[0037] 4. In the fourth step of the present invention, a method for identifying aquatic vegetation and algal blooms based on spectral and relative temperature fusion, the post-processing refers to: first, performing morphological filtering on the preliminary classification image to obtain a denoised classification image; then, performing multi-scale segmentation and smoothing processing on the denoised classification image. In application, morphological filtering can eliminate scattered misclassified pixels in the preliminary classification image, reducing salt-and-pepper noise and enhancing the continuity of category regions; multi-scale segmentation and smoothing processing can fuse adjacent pixels of the same type in the denoised classification image, weakening local misclassification caused by mixed pixels, making the boundaries smoother and the spatial pattern more consistent with the actual water body and vegetation distribution. These two types of spatial post-processing do not require changing the model structure but can significantly improve the spatial consistency, mapping quality, and overall accuracy of the results. In other words, the present invention can output high-precision identification results for aquatic vegetation and algal blooms. Therefore, the output accuracy of the present invention is high. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention.

[0039] Figure 2 It is a scatter plot of absolute lake surface temperature versus water temperature.

[0040] Figure 3 This is a schematic diagram of the architecture of the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Please see Figure 1 — Figure 3 A method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature, the method comprising the following steps:

[0043] Step 1: First, obtain multiple remote sensing datasets, then annotate the aquatic vegetation and algal blooms in the multiple remote sensing datasets, and finally obtain multiple annotated samples that correspond one-to-one with the remote sensing datasets;

[0044] The second step involves sequentially acquiring vectors from multiple remote sensing datasets. This process involves calculating the NDVI, NDWI, FAI, GR ratio, and CMI indices pixel by pixel based on a single remote sensing dataset, and then obtaining the normalized absolute lake surface temperature from that dataset. The NDVI, NDWI, FAI, GR ratio, CMI, and normalized absolute lake surface temperatures are then concatenated to obtain the six-dimensional feature vector F of the dataset. Finally, multiple six-dimensional feature vectors F corresponding to each of the multiple remote sensing datasets are obtained.

[0045] The third step is to take multiple six-dimensional feature vectors F, which correspond one-to-one with multiple remote sensing datasets, as inputs, and the labels in the labeled samples as the corresponding inputs of the six-dimensional feature vectors F. Then, the classifier model is trained based on the inputs and corresponding inputs, and then the set of recognition rules for aquatic vegetation and algal blooms is obtained. Finally, a classifier model including the set of recognition rules for aquatic vegetation and algal blooms is obtained.

[0046] Step 4: First, perform vector acquisition on the remote sensing dataset to be identified, then obtain the six-dimensional feature vector F to be identified. Next, apply a classifier model that includes the set of identification rules for aquatic vegetation and algal blooms to the six-dimensional feature vector F to be identified, then obtain the annotations for aquatic vegetation and algal blooms, then generate a preliminary classification map, then perform post-processing on the preliminary classification map to improve data accuracy, and finally output the identification results of aquatic vegetation and algal blooms.

[0047] In the third step, the classifier model refers to a random forest model with pre-set parameters.

[0048] In the third step, the parameters refer to: the number of trees is 150 to 200, and the maximum depth is 20 to 30.

[0049] In the first step, the labeled samples include a training set and a validation set; in the third step, the classifier model that obtains the set of recognition rules for aquatic vegetation and algal blooms refers to: a classifier model that has been validated and optimized and includes the set of recognition rules for aquatic vegetation and algal blooms.

[0050] In the third step, the verification refers to: verifying the XGBoost model and the SVM model sequentially; the optimization refers to: optimizing the number of trees and the maximum depth of the random forest model using five-fold cross-validation.

[0051] In the first step, the remote sensing dataset refers to: multiple single-scene remote sensing images are preprocessed to improve data accuracy, and then multiple remote sensing datasets corresponding one-to-one with the multiple single-scene remote sensing images are obtained.

[0052] The preprocessing refers to: first, performing atmospheric correction on the data in a single scene of remote sensing imagery, then obtaining the corrected data, and finally performing geometric registration and cropping on the corrected data.

[0053] In the first step, atmospheric correction of the data in a single scene of remote sensing imagery refers to: performing atmospheric correction on the visible light data, near-infrared data, and short-wave infrared data in the single scene of remote sensing imagery using LEDAPS or LaSRC; then performing brightness-temperature conversion on Band10 of the thermal infrared data in the remote sensing dataset, applying NASA radiometric calibration coefficients and NASA atmospheric correction models for atmospheric correction, and finally obtaining the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data.

[0054] In the first step, the geometric registration and cropping of the corrected data refers to: making the resolution of the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data consistent; then resampling the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data to the same grid; then obtaining the same grid data; then cropping the same grid data; and generating a water mask to retain the water portion.

[0055] In the fourth step, the post-processing refers to: first, performing morphological filtering on the preliminary classification map to obtain a denoised classification map, and then performing multi-scale segmentation and smoothing processing on the denoised classification map.

[0056] In the fourth step, the morphological filtering refers to applying an opening / closing operation to the preliminary classification map to remove isolated noise, and then obtaining a denoised classification map.

[0057] The multi-scale segmentation and smoothing process refers to using SLIC or Mean-Shift superpixel segmentation on the obtained denoised classification map to fuse and smooth the boundaries of the denoised classification map.

[0058] The following are supplementary descriptions of the present invention:

[0059] The relative temperature mentioned in this invention refers to the normalized absolute lake surface temperature.

[0060] Example 1:

[0061] Please see Figure 1 — Figure 3 A method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature, the method comprising the following steps:

[0062] Step 1: First, obtain multiple remote sensing datasets, then annotate the aquatic vegetation and algal blooms in the multiple remote sensing datasets, and finally obtain multiple annotated samples that correspond one-to-one with the remote sensing datasets;

[0063] The second step involves sequentially acquiring vectors from multiple remote sensing datasets. This process involves calculating the NDVI, NDWI, FAI, GR ratio, and CMI indices pixel by pixel based on a single remote sensing dataset, and then obtaining the normalized absolute lake surface temperature from that dataset. The NDVI, NDWI, FAI, GR ratio, CMI, and normalized absolute lake surface temperatures are then concatenated to obtain the six-dimensional feature vector F of the dataset. Finally, multiple six-dimensional feature vectors F corresponding to each of the multiple remote sensing datasets are obtained.

[0064] The third step is to take multiple six-dimensional feature vectors F, which correspond one-to-one with multiple remote sensing datasets, as inputs, and the labels in the labeled samples as the corresponding inputs of the six-dimensional feature vectors F. Then, the classifier model is trained based on the inputs and corresponding inputs, and then the set of recognition rules for aquatic vegetation and algal blooms is obtained. Finally, a classifier model including the set of recognition rules for aquatic vegetation and algal blooms is obtained.

[0065] Step 4: First, perform vector acquisition on the remote sensing dataset to be identified, then obtain the six-dimensional feature vector F to be identified. Next, apply a classifier model that includes the set of identification rules for aquatic vegetation and algal blooms to the six-dimensional feature vector F to be identified, then obtain the annotations for aquatic vegetation and algal blooms, then generate a preliminary classification map, then perform post-processing on the preliminary classification map to improve data accuracy, and finally output the identification results of aquatic vegetation and algal blooms.

[0066] Preferably, the month or season code corresponding to the acquisition time of a single remote sensing image is added to the six-dimensional feature vector F as an auxiliary temporal feature to improve the random forest model's ability to identify seasonal algal bloom changes.

[0067] Preferably, this method can be deployed on the Google Earth Engine platform or a local Python batch script environment to achieve batch identification and automatic early warning of algal blooms for multiple urban lakes.

[0068] Specifically, the feature order in the six-dimensional feature vector F is unique to ensure the stability of the model input structure and facilitate transfer and interpretation; if the feature order is disordered, it will cause dimensional mismatch and noise accumulation.

[0069] Example 2:

[0070] The basic content is the same as in Example 1, except that:

[0071] Please see Figure 1 — Figure 3 In the third step, the classifier model refers to a random forest model with pre-set parameters. The parameters in the third step refer to: 150 to 200 trees and a maximum depth of 20 to 30. In the first step, the labeled samples include a training set and a validation set. In the third step, the classifier model that yields the set of recognition rules for aquatic vegetation and algal blooms refers to a validated and optimized classifier model that includes the set of recognition rules for aquatic vegetation and algal blooms. In the third step, validation refers to sequentially using the XGBoost model and then the SVM model; optimization refers to using five-fold cross-validation to optimize the number of trees and the maximum depth of the random forest model.

[0072] In application, in the third step, the random forest model constructs a mapping relationship between the six-dimensional feature vector F and the labels. The random forest model first randomly extracts sub-samples multiple times from the training set, and then performs the training step. The training step is as follows: based on the labels of aquatic vegetation and algal blooms from the single-sample extraction, and the six-dimensional feature vector F, a decision tree is trained. Each decision tree uses a portion of random features when splitting nodes to improve generalization, until 150 to 200 decision trees with a maximum depth of 20 to 30 are generated to form a forest. Then, a set of classification rules for aquatic vegetation and algal blooms is generated by voting, resulting in a random forest model that includes the set of recognition rules for aquatic vegetation and algal blooms. The six-dimensional feature vector F includes spectral features, namely NDVI index, NDWI index, FAI index, GR ratio index, and CMI index features, as well as temperature features, namely normalized absolute lake surface temperature, which helps to generate more stable split points. A larger tree depth and a moderate number of trees are used so that the random forest model can capture aquatic vegetation-algal blooms. - Fine-grained differences in water bodies; the set of identification rules for aquatic vegetation and algal blooms serves as the optimal decision boundary automatically learned by the model, without the need for manual threshold setting; then, the random forest model, which includes the set of identification rules for aquatic vegetation and algal blooms, is validated on the validation set using XGBoost and SVM models. XGBoost emphasizes gradient boosting and feature interaction, while SVM excels at maximizing the boundary of high-dimensional features. If the random forest model, XGBoost model, and SVM model can all achieve high accuracy on the same feature set, it indicates that the feature set based on labeled samples is stable, has strong physical meaning, and good transferability; then, five-fold cross-validation is used to optimize the number of trees and maximum depth using grid search, which can avoid overfitting / underfitting caused by accidental partitioning, improve generalization ability with limited samples, and systematically evaluate the stability of model performance under different hyperparameters (i.e., number of trees and maximum depth), and make the model more reliable in cases of uneven samples.

[0073] Example 3:

[0074] The basic content is the same as in Example 1, except that:

[0075] Please see Figure 1 — Figure 3In the first step, the remote sensing dataset refers to: multiple single-scene remote sensing images undergoing preprocessing to improve data accuracy, resulting in multiple remote sensing datasets corresponding one-to-one with the single-scene remote sensing images; the preprocessing refers to: first performing atmospheric correction on the data in the single-scene remote sensing images, then obtaining the corrected data, and finally performing geometric registration and cropping on the corrected data. In the first step, the atmospheric correction of the data in the single-scene remote sensing images refers to: performing atmospheric correction on the visible light data, near-infrared data, and short-wave infrared data in the single-scene remote sensing images using LEDAPS or LaSRC, then sequentially performing brightness-temperature conversion on Band10 of the thermal infrared data in the remote sensing dataset, applying NASA radiometric calibration coefficients and the NASA atmospheric correction model for atmospheric correction, and finally obtaining the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data. In the first step, the geometric registration and cropping of the corrected data refers to: making the resolution of the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data consistent; then resampling the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data to the same grid; then obtaining the same grid data; then cropping the same grid data; and generating a water mask to retain the water portion.

[0076] In application, the visible light, near-infrared, and shortwave infrared data in a single scene of remote sensing imagery are first atmospherically corrected using LEDAPS or LaSRC. Then, the Band10 data in the thermal infrared data set is sequentially subjected to brightness-temperature conversion, and atmospheric correction is performed using NASA radiometric calibration coefficients and the NASA atmospheric correction model. This results in corrected visible light, near-infrared, shortwave infrared, and thermal infrared data with consistent resolution. The corrected data is then resampled to the same raster grid, resulting in identical raster data. This identical raster data is then cropped, and a water mask is generated to preserve the water body portion. The water mask is generated using a combination of an NDWI > 0.3 semantic threshold and Otsu automatic segmentation to remove shoreline and cloud / fog pixels.

[0077] Example 4:

[0078] The basic content is the same as in Example 1, except that:

[0079] Please see Figure 1 — Figure 3In the fourth step, the post-processing refers to: first, performing morphological filtering on the preliminary classification image to obtain a denoised classification image, and then performing multi-scale segmentation and smoothing processing on the denoised classification image. In the fourth step, the morphological filtering refers to: applying opening / closing operations to the preliminary classification image to remove isolated noise, and then obtaining the denoised classification image; the multi-scale segmentation and smoothing processing refers to: using SLIC or Mean-Shift superpixel segmentation on the obtained denoised classification image to fuse and smooth the boundaries of the denoised classification image.

[0080] In application, each pixel value of the preliminary classification image corresponds to the labeling of submerged plants, emergent plants, floating-leaved plants, and algal blooms. Then, an opening operation (structural element 3×3) is performed on the preliminary classification image to remove isolated patches (i.e., noise) smaller than 5 pixels. Then, a closing operation (structural element 5×5) is performed to smooth the target boundary, resulting in a denoised classification image. The SLIC algorithm (region size 100 pixels, compactness 10) is then used on the denoised classification image for object-level reclassification, and the boundary is merged to achieve a smooth and continuous classification surface. Finally, the recognition results of aquatic vegetation and algal blooms are output.

[0081] Example 5:

[0082] The basic content is the same as in Example 1, except that:

[0083] Please see Figure 1 — Figure 3 The first step is to obtain multiple single-scene remote sensing images of location A and location B (Landsat8OLI–TIRS Level-2 Surface Reflectance product), then preprocess the multiple single-scene remote sensing images to obtain multiple remote sensing datasets.

[0084] Next, aquatic vegetation and algal blooms were labeled on multiple remote sensing datasets. For example, visual interpretation was performed using GPS positioning in the field combined with high-resolution true-color aerial images to label four types of targets: ① Submerged vegetation (underwater vegetation in shallow water areas, 50% of samples); ② Emergent / floating-leaved vegetation (duckweed, lotus, etc., 25% of samples); ③ Planktonic algal blooms (cyanobacteria zone, 15% of samples); ④ Bare water surface (no vegetation, 10% of samples). Then, the labeled samples were divided into a training set (77 points) and a validation set (33 points) at a ratio of 70% and 30%.

[0085] The second step is to perform vector acquisition steps on multiple remote sensing datasets in sequence. The vector acquisition step refers to calculating the NDVI index, NDWI index, FAI index, GR ratio index, and CMI index pixel by pixel based on a remote sensing dataset.

[0086] The formula for calculating the NDVI index is as follows:

[0087] ;

[0088] In the formula, For near-infrared reflectivity, Reflectivity in the red light band;

[0089] The formula for calculating the NDWI index is as follows:

[0090] ;

[0091] In the formula, For shortwave infrared reflectivity, Reflectivity in the near-infrared band;

[0092] The formula for calculating the FAI index is as follows:

[0093] ;

[0094] In the formula, For red light band reflectivity, For near-infrared reflectivity, For shortwave infrared band reflectivity;

[0095] The formula for calculating the GR ratio index is as follows:

[0096] ;

[0097] In the formula, For shortwave infrared reflectivity, Reflectivity in the red light band;

[0098] The formula for calculating the CMI index is as follows:

[0099] ;

[0100] In the formula, For blue light band reflectivity, For shortwave infrared reflectivity, Reflectivity in the red light band;

[0101] To ensure stable calculations, the denominator in the above formula can be increased by adding... To prevent division by zero;

[0102] Based on this remote sensing dataset, the absolute surface temperature (LST) is obtained using the single-window algorithm (parameters: atmospheric emissivity ε=0.98, transmittance τ=0.85). Then, the normalized absolute lake surface temperature (LST) is calculated. The normalized absolute lake surface temperature The calculation formula is:

[0103] ;

[0104] In the formula, It is the absolute surface temperature. This represents the minimum temperature of the current panoramic lake surface pixels in the image. This represents the maximum temperature of the current panoramic lake surface pixel in the image.

[0105] Normalized absolute lake surface temperature The value range is [0,1], in order to eliminate the absolute temperature difference between different images and different lakes, a relative feature map of thermal anomalies can be generated;

[0106] Then, the NDVI index, NDWI index, FAI index, GR ratio index, CMI index, and normalized absolute lake surface temperature were calculated. The six single-channel graticules are resampled, aligned, and superimposed to form a six-band multidimensional image, which yields the six-dimensional feature vector F of the remote sensing dataset, used in the third step for subsequent sample extraction and model prediction.

[0107] Step 3: First, use the six-dimensional feature vector F as input, and the labels in the training set as the corresponding input to the six-dimensional feature vector F. Then, set the number of trees in the random forest model to 150, the maximum depth to 25, and the minimum sample split to 5. Then, train the random forest model to construct the recognition rules for aquatic vegetation and algal blooms, and obtain an initial random forest model that includes the recognition rule set for aquatic vegetation and algal blooms. Then, verify the accuracy of the initial model through the validation set, and then fine-tune the optimal number of trees and depth through five-fold cross-validation. Finally, obtain a random forest model that includes the recognition rule set for aquatic vegetation and algal blooms.

[0108] Step 4: First, preprocess and obtain vectors from the single-scene remote sensing images to be identified. Then, obtain the six-dimensional feature vector F to be identified. Next, apply a random forest model to the six-dimensional feature vector F, which includes a set of identification rules for aquatic vegetation and algal blooms, to obtain the labels for the four types of targets. Then, generate a preliminary classification map, with each pixel value corresponding to the labels of categories ① to ④. Then, perform opening and closing operations on the preliminary classification map to remove noise and smooth the boundaries. Then, use the SLIC algorithm (region size 100 pixels, compactness 10) to perform object-level reclassification, merge the boundaries, and achieve a smooth and continuous classification surface. Then, export the classification surface as GeoTIFF with a resolution of 30m, convert it into a vector Shapefile, and attach attribute fields: category, area (square meters), polygon ID. Generate statistical bar charts, heat maps, and spatiotemporal distribution maps for categories ① to ④. Summarize them into an automated PDF report for use by ecological management departments for decision-making.

[0109] Confusion matrix statistics were performed on the validation set (33 points), and the overall accuracy and Kappa coefficient were calculated. The overall accuracy of location A was 95.2%, and Kappa=0.93; the overall accuracy of location B was 93.8%, and Kappa=0.91. The overall accuracy of both locations was greater than 90%, which is at an extremely high level, indicating that more than 93% of the pixels were correctly classified and identified, and the ability to distinguish between water bodies and non-water bodies was strong. The Kappa coefficients were all far above 0.8 (the excellent threshold), which means that the classification consistency was extremely strong, the model had high stability in identifying water bodies, and was less affected by interference.

[0110] Compared with the NDVI / NDWI thresholding method alone, this method improves the overall accuracy by an average of 12.4% and the Kappa coefficient by 0.14. Compared with the absolute LST-assisted method, this invention reduces threshold calibration work and improves accuracy by 8%.

[0111] When the location A model obtained by this invention is directly applied to the image of location B on the same day, the overall accuracy of the validation set (20 points) is 91.0% and Kappa=0.88, which proves that the model generated by this invention has good transferability.

[0112] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature, characterized in that: The method includes the following steps: Step 1: First, obtain multiple remote sensing datasets, then annotate the aquatic vegetation and algal blooms in the multiple remote sensing datasets, and finally obtain multiple annotated samples that correspond one-to-one with the remote sensing datasets; The second step involves sequentially acquiring vectors from multiple remote sensing datasets. This process involves calculating the NDVI, NDWI, FAI, GR ratio, and CMI indices pixel by pixel based on a single remote sensing dataset, and then obtaining the normalized absolute lake surface temperature from that dataset. The NDVI, NDWI, FAI, GR ratio, CMI, and normalized absolute lake surface temperatures are then concatenated to obtain the six-dimensional feature vector F of the dataset. Finally, multiple six-dimensional feature vectors F corresponding to each of the multiple remote sensing datasets are obtained. The third step is to take multiple six-dimensional feature vectors F, which correspond one-to-one with multiple remote sensing datasets, as inputs, and the labels in the labeled samples as the corresponding inputs of the six-dimensional feature vectors F. Then, the classifier model is trained based on the inputs and corresponding inputs, and then the set of recognition rules for aquatic vegetation and algal blooms is obtained. Finally, a classifier model including the set of recognition rules for aquatic vegetation and algal blooms is obtained. Step 4: First, perform vector acquisition on the remote sensing dataset to be identified, then obtain the six-dimensional feature vector F to be identified. Next, apply a classifier model that includes the set of identification rules for aquatic vegetation and algal blooms to the six-dimensional feature vector F to be identified, then obtain the annotations for aquatic vegetation and algal blooms, then generate a preliminary classification map, then perform post-processing on the preliminary classification map to improve data accuracy, and finally output the identification results of aquatic vegetation and algal blooms. In the second step, obtaining the normalized absolute lake surface temperature based on the remote sensing dataset means: obtaining the absolute surface temperature (LST) based on the remote sensing dataset, and then calculating the normalized absolute lake surface temperature (LST). The normalized absolute lake surface temperature The calculation formula is: ； In the formula, It is the absolute surface temperature. This represents the minimum temperature of the current panoramic lake surface pixels in the image. This represents the maximum temperature of the current panoramic lake surface pixel in the image. Normalized absolute lake surface temperature The value range is [0,1], which eliminates the absolute temperature difference between different images and different lakes, thus generating a relative feature map of thermal anomalies.

2. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature as described in claim 1, characterized in that: In the third step, the classifier model refers to a random forest model with pre-set parameters.

3. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature as described in claim 2, characterized in that: In the third step, the parameters refer to: the number of trees is 150 to 200, and the maximum depth is 20 to 30.

4. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature as described in claim 3, characterized in that: In the first step, the labeled samples include a training set and a validation set; in the third step, the classifier model that obtains the set of recognition rules for aquatic vegetation and algal blooms refers to: a classifier model that has been validated and optimized and includes the set of recognition rules for aquatic vegetation and algal blooms.

5. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature according to claim 4, characterized in that: In the third step, the verification refers to: verifying the XGBoost model and the SVM model sequentially; the optimization refers to: optimizing the number of trees and the maximum depth of the random forest model using five-fold cross-validation.

6. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature as described in claim 1, characterized in that: In the first step, the remote sensing dataset refers to: multiple single-scene remote sensing images are preprocessed to improve data accuracy, and then multiple remote sensing datasets corresponding one-to-one with the multiple single-scene remote sensing images are obtained. The preprocessing refers to: first, performing atmospheric correction on the data in a single scene of remote sensing imagery, then obtaining the corrected data, and finally performing geometric registration and cropping on the corrected data.

7. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature according to claim 6, characterized in that: In the first step, atmospheric correction of the data in a single scene of remote sensing imagery refers to: performing atmospheric correction on the visible light data, near-infrared data, and short-wave infrared data in the single scene of remote sensing imagery using LEDAPS or LaSRC; then performing brightness-temperature conversion on Band10 of the thermal infrared data in the remote sensing dataset, applying NASA radiometric calibration coefficients and NASA atmospheric correction models for atmospheric correction, and finally obtaining the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data.

8. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature as described in claim 7, characterized in that: In the first step, the geometric registration and cropping of the corrected data refers to: making the resolution of the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data consistent; then resampling the corrected visible light data, near-infrared data, short-wave infrared data, and thermal infrared data to the same grid; then obtaining the same grid data; then cropping the same grid data; and generating a water mask to retain the water portion.

9. The method for identifying aquatic vegetation and algal blooms based on the fusion of spectral and relative temperature according to claim 1, characterized in that: In the fourth step, the post-processing refers to: first, performing morphological filtering on the preliminary classification map to obtain a denoised classification map, and then performing multi-scale segmentation and smoothing processing on the denoised classification map.

10. The method for identifying aquatic vegetation and algal blooms based on spectral and relative temperature fusion according to claim 9, characterized in that: In the fourth step, the morphological filtering refers to applying an opening / closing operation to the preliminary classification map to remove isolated noise, and then obtaining a denoised classification map. The multi-scale segmentation and smoothing process refers to using SLIC or Mean-Shift superpixel segmentation on the obtained denoised classification map to fuse and smooth the boundaries of the denoised classification map.

Images