Construction method of water quality parameter inversion model

Abstract

The invention provides a construction method of a water quality parameter inversion model, and relates to the field of model training.The construction method comprises the steps that three kinds of remote sensing data including visible light images, hyperspectral data and microwave remote sensing data are synchronously obtained, and actual water quality parameters of corresponding water areas are collected in a matched mode; a common feature space of image data, hyperspectral data and microwave remote sensing data is established to serve as a fusion data set, an initial water quality model is constructed, iteration is conducted on the initial water quality model on the basis of predicted water quality parameters output by the initial water quality model, the finally generated water quality parameter inversion model can overcome limitation of a single sensor, and the accuracy of water quality inversion is improved. Based on multi-dimensional information of visible light, hyperspectrum and microwave remote sensing, calculation of water quality parameters is realized, detection precision is improved, adaptability is enhanced, and the method can face a complex water body environment and meet actual detection requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of model training, in particular to a method for constructing a water quality parameter inversion model. BACKGROUND

[0002] Water quality monitoring and parameter inversion are important research directions in the intersection field of environmental science and remote sensing technology, and have irreplaceable value for guaranteeing water resource safety and maintaining ecological balance. With the increasing severity of global water environmental problems, accurate acquisition of water quality parameters has become a key support for water resource management and pollution prevention. However, current water quality inversion methods mostly rely on single light source detection algorithms, and generally have defects such as insufficient accuracy, poor adaptability, and poor response to complex water body environments, especially when facing variable water body characteristics and regional differences, existing algorithms often fail to meet actual detection needs. SUMMARY

[0003] Therefore, the present application provides a method for constructing a water quality parameter inversion model, the main purpose of which is to realize detection of water body environment based on multiple light source detection technologies at the same time and improve detection effect.

[0004] To achieve the above purpose, the present application discloses a method for constructing a water quality parameter inversion model, which comprises: obtaining image data, hyperspectral data and microwave remote sensing data obtained by respectively performing image detection, hyperspectral detection and microwave remote sensing detection on a reference water area, and actual water quality parameters corresponding to the reference water area, the reference water area containing at least one type of water area; performing cross-modal mapping on the image data, the hyperspectral data and the microwave remote sensing data by using a multi-modal feature alignment algorithm to obtain a fusion data set fused with the image data, the hyperspectral data and the microwave remote sensing data; constructing an initial water quality model by using a first fusion data set and a first actual water quality parameter, the first fusion data set and the first actual water quality parameter corresponding to each other, the first fusion data set being part of the fusion data set, and the first actual water quality parameter being part of the actual water quality parameters; inputting a second fusion data set into the initial water quality model to obtain predicted water quality parameters output by the initial water quality model, determining a parameter difference between the predicted water quality parameters and a second actual water quality parameter, the second fusion data set being part of the fusion data set, and the second fusion data set being different from the first fusion data set, the second actual water quality parameter being part of the actual water quality parameters, and the second actual water quality parameter being different from the first actual water quality parameter, the second actual water quality parameter corresponding to the second fusion data set; According to the parameter difference, the model parameters of the initial water quality model are adjusted until a water quality parameter inversion model is obtained.

[0005] In a second aspect of the present application, a device for constructing a water quality parameter inversion model is provided, and the device comprises: A first obtaining module is configured to obtain image data, hyperspectral data and microwave remote sensing data obtained by performing image detection, hyperspectral detection and microwave remote sensing detection on a reference water area, respectively, and actual water quality parameters corresponding to the reference water area, wherein the reference water area comprises at least one type of water area. A second obtaining module is configured to perform cross-modal mapping on the image data, the hyperspectral data and the microwave remote sensing data by using a multi-modal feature alignment algorithm, to obtain a fusion data set in which the image data, the hyperspectral data and the microwave remote sensing data are fused. A constructing module is configured to construct an initial water quality model by using a first fusion data set and first actual water quality parameters, wherein the first fusion data set and the first actual water quality parameters correspond to each other, the first fusion data set is part of the fusion data set, and the first actual water quality parameters are part of the actual water quality parameters. A third obtaining module is configured to input a second fusion data set into the initial water quality model, to obtain predicted water quality parameters output by the initial water quality model, to determine a parameter difference between the predicted water quality parameters and second actual water quality parameters, wherein the second fusion data set is part of the fusion data set, the second fusion data set is different from the first fusion data set, the second actual water quality parameters are part of the actual water quality parameters, the second actual water quality parameters are different from the first actual water quality parameters, and the second actual water quality parameters correspond to the second fusion data set. An adjusting module is configured to adjust model parameters of the initial water quality model according to the parameter difference, until a water quality parameter inversion model is obtained.

[0006] In a third aspect of the present application, an electronic device is provided, and the electronic device comprises: at least one processor; and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any one of the methods disclosed in the first aspect.

[0007] In a fourth aspect of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method of the first aspect.

[0008] In summary, according to the technical solution disclosed in the present application, by acquiring image data, hyperspectral data and microwave remote sensing data obtained by respectively performing image detection, hyperspectral detection and microwave remote sensing detection on a reference water area, and actual water quality parameters corresponding to the reference water area, the reference water area includes at least one type of water area; a multi-modal feature alignment algorithm is used to perform cross-modal mapping on the image data, the hyperspectral data and the microwave remote sensing data, to obtain a fusion data set in which the image data, the hyperspectral data and the microwave remote sensing data are fused; an initial water quality model is constructed using a first fusion data set and a first actual water quality parameter, the first fusion data set and the first actual water quality parameter correspond to each other, the first fusion data set is part of the fusion data set, and the first actual water quality parameter is part of the actual water quality parameters; a second fusion data set is input into the initial water quality model, to obtain predicted water quality parameters output by the initial water quality model, and a parameter difference between the predicted water quality parameters and a second actual water quality parameter is determined, the second fusion data set is part of the fusion data set, and the second fusion data set is different from the first fusion data set, the second actual water quality parameter is part of the actual water quality parameters, and the second actual water quality parameter is different from the first actual water quality parameter, the second actual water quality parameter corresponds to the second fusion data set; and model parameters of the initial water quality model are adjusted according to the parameter difference, until a water quality parameter inversion model is obtained.

[0009] The present application synchronously acquires three kinds of remote sensing data: visible light image, hyperspectral data and microwave remote sensing data, and corresponding actual water quality parameters of the water area are collected in a matched manner, a common feature space of the image data, the hyperspectral data and the microwave remote sensing data is established as a fusion data set, the construction of an initial water quality model is realized, the initial water quality model is iterated based on predicted water quality parameters output by the initial water quality model, and finally generated water quality parameter inversion model can overcome the limitation of a single sensor, realize the calculation of water quality parameters based on multi-dimensional information of visible light, hyperspectrum and microwave remote sensing, improve the detection precision, enhance the adaptability to face complex water body environment, and meet the actual detection requirements.

[0010] The above description is only a summary of the technical solution of the present application. In order to more clearly understand the technical means of the present application, the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are incorporated into the specification and constitute a part of the specification, show embodiments consistent with the present application and, together with the specification, serve to explain the principles of the present application.

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0013] Figure 1 A flow chart of a method for constructing a water quality parameter inversion model is shown. Figure 2 A structural diagram of a device for constructing a water quality parameter inversion model is shown. DETAILED DESCRIPTION

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0015] In order to solve the problem that current water quality inversion methods mostly rely on single light source detection algorithms, and the defects of insufficient accuracy, poor adaptability and poor response to complex water body environment, especially when facing variable water body characteristics and regional differences, the existing algorithms often cannot meet the actual detection requirements. The present application provides the following embodiments to solve the above problems: The present embodiment provides a method for constructing a water quality parameter inversion model, as shown in Figure 1 The present embodiment method can specifically include the following steps: Step 101, obtaining image data, hyperspectral data and microwave remote sensing data obtained by respectively performing image detection, hyperspectral detection and microwave remote sensing detection on a reference water area, and actual water quality parameters corresponding to the reference water area, the reference water area containing at least one type of water area.

[0016] Obtaining initial image data, initial hyperspectral data and initial microwave remote sensing data obtained by respectively performing image detection, hyperspectral detection and microwave remote sensing detection on the reference water area; Using a grid resampling method, projecting the initial image data, the initial hyperspectral data and the initial microwave remote sensing data into a preset spatial scale to determine the mapping image data, the mapping hyperspectral data and the mapping microwave remote sensing data contained in the preset spatial scale; Obtaining resolution requirements corresponding to the water area type; According to the resolution requirement, the alignment accuracy between the mapped image data, the mapped hyperspectral data and the mapped microwave remote sensing data in the same preset spatial scale is adjusted, and the mapped image data, the mapped hyperspectral data and the mapped microwave remote sensing data after the alignment accuracy adjustment are used as the image data, the hyperspectral data and the microwave remote sensing data.

[0017] According to the initial image data, the initial hyperspectral data and the initial microwave remote sensing data obtained through image detection, hyperspectral detection and microwave remote sensing detection, and according to the differences in spatial resolution and time coverage, a preset grid normalization method is used to perform preliminary resampling processing on the initial image data, the initial hyperspectral data and the initial microwave remote sensing data, to adjust to a unified preset scale, to obtain an intermediate data set after preliminary adjustment. According to the spectral features and band distribution differences in the intermediate data set, a data fusion tool is used to perform layer-by-layer matching on the features of the intermediate data set. If the spectral features of a certain initial image data, initial hyperspectral data and initial microwave remote sensing data do not meet the preset threshold, a feature completion is performed through an interpolation calculation tool, to determine a feature data set after fusion. Through spatial consistency detection on the texture characteristics in the feature data set, an image processing tool is used to perform smoothing processing on the texture characteristics, to obtain smoothed texture distribution data, to determine whether the preset standard of spatial consistency is met, to obtain a processing data set after consistency verification. According to the spatial scale of the processing data set, a data integration tool is used to map the processing data set to a unified grid normalization framework, to perform final calibration according to the requirements of spatial consistency and preset scale, to obtain normalized image data, hyperspectral data and microwave remote sensing data.

[0018] Exemplarily, when processing the initial image data, the initial hyperspectral data and the initial microwave remote sensing data, there are often differences in spatial resolution and time coverage due to different sensors. Assuming that the river coverage data of a certain region is processed, the data sources include a satellite A with a spatial resolution of 10 meters and a satellite B with a resolution of 30 meters, and the time coverage of the satellite A is updated once a month, while the time coverage of the satellite B is updated once a quarter. According to such differences, a grid normalization method can be used to uniformly resample the spatial resolutions of the two to a scale of 20 meters, and the time coverage is supplemented with missing time points through an interpolation method, to form an intermediate data set after preliminary adjustment. The advantage of this is that the data benchmark is unified, which is convenient for subsequent fusion analysis.

[0019] In a possible implementation, the application of the data fusion tool is crucial for the spectral features and band distribution differences of the intermediate data set.

[0020] For example, satellite A covers the visible light band, while satellite B contains the near-infrared band. For cases where the bands are inconsistent, the spectral information of the two can be aligned by matching features layer by layer. If the data of a certain band of satellite B does not meet the preset threshold, such as the reflectivity value being too low, the missing features are completed using an interpolation calculation tool, and finally a fused feature dataset is formed. This method can effectively improve the completeness of the data and provide more comprehensive spectral information for subsequent analysis.

[0021] Specifically, when detecting the spatial consistency of the texture characteristics in the feature dataset, an image processing tool can be used to smooth the texture characteristics. Assuming that in the river coverage data of a certain area, the texture characteristics present a discontinuous patchy distribution due to sensor noise, smoothing can eliminate these noises and make the texture distribution data more consistent in space, such as ensuring smooth transition of the texture boundary in the lake water area. Such processing not only improves the visual consistency of the data, but also provides a more reliable basis for subsequent classification and identification.

[0022] Preferably, in the final calibration stage, a data integration tool is used to map the spatial scale of the processed dataset to a unified grid standardization framework.

[0023] For example, the processed data is unified to a 20-meter resolution grid and calibrated for spatial consistency and preset scale requirements to ensure that the boundaries of different data sources are aligned and errors caused by scale differences are avoided. The resulting standardized first dataset can significantly improve the accuracy and consistency of the data in water source coverage monitoring, change detection, and other applications.

[0024] Among the mapped image data, mapped hyperspectral data, and mapped microwave remote sensing data, exemplary include high-resolution optical images and low-resolution microwave data with resolutions of 15 meters and 50 meters, respectively. A grid division tool can divide the data by spatial scale, such as dividing high-resolution data into fine layers and low-resolution data into coarse layers, each layer corresponding to a different spatial scale range, such as 10-20 meters for fine layers and 40-60 meters for coarse layers. Such layering helps to process different scale data separately and ensures the relevance of subsequent analysis.

[0025] Step 102, using a multi-modal feature alignment algorithm, performing cross-modal mapping on the image data, hyperspectral data, and microwave remote sensing data to obtain a fused dataset fused with the image data, hyperspectral data, and microwave remote sensing data.

[0026] Using a multi-modal feature alignment algorithm, performing cross-modal mapping on the spectral features in the image data, the band information in the hyperspectral data, and the texture features in the microwave remote sensing data to obtain a first dataset of aligned spectral features, band information, and texture features; The distribution values of the first data set in different water area types are determined by using a deep convolutional network, and the multiple distribution values and the first data set corresponding to the distribution values are combined to form a fusion data set.

[0027] The multiple distribution values and the first data set corresponding to the distribution values are combined to form a fusion data set, comprising: The first weight value of the spectral feature, the second weight value of the band information, and the third weight value of the texture feature in the first data set are calculated by using a weighted attention mechanism. The size relationship between the first weight value, the second weight value, and the third weight value and a preset weight value is compared respectively. If at least one of the first weight value, the second weight value, and the third weight value is less than the preset weight value, the first data set is subjected to data enhancement until the enhanced first weight value, second weight value, and third weight value are greater than the preset weight value. The multiple distribution values and the data-enhanced first data set corresponding to the distribution values are combined to form a fusion data set.

[0028] According to the data heterogeneity characteristics of image data, hyperspectral data, and microwave remote sensing data, a data preprocessing tool is used to classify and organize different modal data, to obtain classified distribution data and determine the feature range of each class of data. For the classified distribution data, in combination with the demand for cross-modal mapping, a feature extraction tool is used to compare the spectral feature, band information, and texture feature. If the weight value does not match the preset threshold value, data completion is performed by an interpolation calculation tool to obtain the completed distribution data. According to the completed distribution data, in combination with the adjustment target of the spectral feature, band information, and texture feature, data enhancement is performed, for example, the smoothness adjustment target of the texture feature is performed. An image smoothing tool is used to locally correct the texture feature to obtain corrected texture distribution data, and further determine whether the smoothness requirement is met. For the corrected spectral feature, band information, and texture feature distribution data, in combination with the target of spatial optimization and data alignment, a grid mapping tool is used to integrate the feature data of different modalities into a unified feature space to obtain the aligned final data set.

[0029] Exemplarily, in processing the data heterogeneity in the intermediate data set, the differences in spectral features and band information and texture features can be classified and sorted by the data preprocessing tool for different modalities of data. Assuming that the processing is a river coverage monitoring data of a certain region, different modalities of data can include image data, hyperspectral data and microwave remote sensing data obtained by optical image devices and synthetic aperture radar respectively, and their spectral features, band information and texture features are significantly different. The data preprocessing tool can classify the data obtained by the optical image device as a high-resolution spectral data group, and the data obtained by the radar as a low-resolution structure data group, and extract their spectral distribution data respectively to determine the feature range of the band information and texture feature data in different resolution structure groups.

[0030] For the classified spectral distribution data, in combination with the demand for cross-modal mapping, the feature extraction tool can be used to compare the spectral features, band information and texture features respectively. If it is found that the weight value of the spectral features, band information or texture features of a certain class of data does not meet the preset threshold value, for example, the near-infrared band response value of the optical data is lower than the expected threshold value of 80%, then the data is completed by the interpolation calculation tool. The specific completion method can be to perform spatial interpolation on the missing or low band value to simulate a response value closer to the true distribution, forming the completed distribution data.

[0031] Further, exemplarily, for the completed texture features, in combination with the smoothing adjustment goal of the texture features, the image smoothing tool can be used to locally correct the texture features.

[0032] For example, in the optical data, the river boundary area may present irregular texture due to noise interference, and the local processing of these areas by the smoothing tool makes the boundary transition more natural. After correction, it is necessary to judge whether it meets the smoothness requirement, for example, the texture change rate is required to be controlled within 5%.

[0033] In one possible implementation, assuming that the processing is a certain water environment monitoring data, the spectral distribution data can include a reflectivity of 20% to 30% in the blue band and a reflectivity of 50% to 60% in the near-infrared band, and through classification and sorting, the characteristics of different bands can be clearly distinguished to form a preliminary sorted distribution data set. For the preliminary sorted distribution data set, in combination with the texture details and the characteristics of the water environment, the image processing tool can be used to extract local features. For example, the water edge area may present irregular texture due to wave interference, and if the extracted texture detail change rate exceeds the preset threshold value of 10%, adjustment is required by the smoothing processing tool. The smoothing processing can suppress the high-frequency noise of the edge area to make the texture transition more natural, and obtain the adjusted texture distribution data, which is helpful to the accuracy of subsequent feature analysis.

[0034] For the adjusted texture distribution data, the feature encoding tool can perform hierarchical processing on spectral information and spatial distribution in combination with the target of encoding level and fusion features. In principle, hierarchical encoding is to divide data into multiple levels according to importance or spatial scale, for example, the spectral information of the center area of the water body is taken as the main level, and the edge area is taken as the secondary level. Assuming that the blue light reflectivity data of the center area of the water body is encoded as high priority, and the texture data of the edge area is encoded as low priority, the final hierarchical encoded feature set is formed. This hierarchical method facilitates subsequent differential processing.

[0035] Preferably, for the hierarchical encoded feature set, the data comparison tool can record the performance difference of the feature set in different water body environments in combination with the requirement of distribution difference and representation result.

[0036] The spectral reflectivity in still water environment and turbulent flow environment may differ by 15%, if the difference exceeds the preset condition such as 10%, the data correction tool needs to be adjusted.

[0037] The correction tool can perform smoothing processing on the data in the turbulent flow environment, so that it is closer to the feature distribution in the still water environment, and finally determine the feature representation data. This adjustment method helps to improve the adaptability of the data in different environments.

[0038] According to the preliminary representation result of the feature set, the data integration tool is used to classify the features of different sources according to the characteristics of multi-source information, to obtain the classified feature distribution data, and to obtain the preliminary integrated feature group. For the preliminary integrated feature group, the data comparison tool is used to calculate the importance score of each feature according to the requirement of weight evaluation, if the score of a certain feature is lower than the preset threshold, the data enhancement tool is used to adjust its weight, to determine the adjusted feature weight data. According to the adjusted feature weight data, the feature mapping tool is used to supplement the details of the features with low weight according to the risk of information loss, to obtain the supplemented feature distribution set. For the supplemented feature distribution set, the feature fusion tool is used to integrate multi-source information hierarchically according to the target of feature optimization, to obtain the final optimized feature data.

[0039] Exemplarily, for the preliminary representation result of the feature set, the data integration tool can be used to classify the characteristics of image data, hyperspectral data and microwave remote sensing data when classifying the characteristics of image data, hyperspectral data and microwave remote sensing data. The principle of the data integration tool is to group and arrange features of different sources such as spectrum, space and texture, so as to facilitate subsequent analysis.

[0040] For example, in water body environment monitoring, spectral features can be classified by wave band range, while spatial features can be classified by regional distribution. In a possible implementation, assuming that the processing is a lake monitoring data, the spectral features can be classified into two categories of visible light and infrared wave band, and the spatial features can be classified into two categories of central region and edge region, to form a preliminary integrated feature group through classification processing. This classification manner facilitates subsequent targeted analysis.

[0041] In an embodiment, for the preliminary integrated feature group, in combination with the requirement of weight evaluation, a data comparison tool can be used to evaluate the importance of each feature. In principle, weight evaluation is scored by comparing the contribution of the feature to the overall representation.

[0042] For example, assuming that the spectral feature has a greater impact on the result in water body monitoring, the score of the spectral feature can be 80, while the score of the texture feature is only 30. If the preset threshold is 50, the texture feature needs to be adjusted by the data enhancement tool.

[0043] Specifically, the adjustment can be to increase the sampling density of the texture feature to increase its weight, and finally the score of the texture feature can be increased to 55 to form the adjusted feature weight data. This manner is helpful to balance the influence of each feature.

[0044] It should be noted that for the adjusted feature weight data, to cope with the risk of information loss, a feature mapping tool can supplement details of a feature with a lower weight. In principle, detail supplement is to improve the data by mining the potential information of the low-weight feature.

[0045] For example, in water body monitoring, if the texture feature weight of the edge region is low, the representation capability can be enhanced by supplementing the detail change data of the edge region.

[0046] Preferably, assuming that the texture change data of the edge region originally covers only 20% of the details, the mapping tool can supplement it to 40% to form a supplemented feature distribution set. This supplement is helpful to retain more key information. For the supplemented feature distribution set, in combination with the goal of feature optimization, a feature fusion tool can perform hierarchical integration on multi-source information. In principle, hierarchical integration is to integrate different features by importance or spatial scale, and finally form optimized data.

[0047] In step 103, an initial water quality model is constructed by using a first fusion data set and a first actual water quality parameter, the first fusion data set and the first actual water quality parameter correspond to each other, the first fusion data set is part of the fusion data set, and the first actual water quality parameter is part of the actual water quality parameter.

[0048] The data mapping tool is used for nonlinear processing of the first fused data set to obtain a preliminary mapping relationship corresponding to the first actual water quality parameter, i.e., an initial water quality model.

[0049] The nonlinear processing of the data mapping tool can be based on the complex relationship between the first fused data set and the first actual water quality parameter. In principle, the nonlinear processing is for the non-direct correspondence between the first fused data set and the first actual water quality parameter.

[0050] For example, turbidity and water pollution degree may not be simply linearly related, but may present a curve change. Through the mapping tool, a mapping relationship of turbidity value to pollution index can be initially constructed to obtain a mapped data distribution, which better fits the complexity of the actual water body scene.

[0051] In step 104, the second fused data set is input into the initial water quality model to obtain a predicted water quality parameter output by the initial water quality model, and a parameter difference between the predicted water quality parameter and a second actual water quality parameter is determined. The second fused data set is part of the fused data set, and the second fused data set is different from the first fused data set. The second actual water quality parameter is part of the actual water quality parameter, and the second actual water quality parameter is different from the first actual water quality parameter. The second actual water quality parameter corresponds to the second fused data set.

[0052] If the predicted water quality parameter output by the initial water quality model in response to the input second fused data set exceeds a preset threshold, the initial water quality model parameters are adjusted by a deviation correction tool to obtain a corrected predicted water quality parameter, and it is determined whether the corrected predicted water quality parameter meets an expected range. According to the corrected predicted water quality parameter, a data integration tool is used to associate and match the corrected predicted water quality parameter with the initial water quality model to obtain a final prediction value related to the water quality parameter.

[0053] It should be noted that if the response value of the mapped data distribution under the complex water body scene exceeds the preset threshold, the application of the deviation correction tool is particularly important. In principle, the deviation correction aims to adjust the initial water quality model to make it more consistent with the actual monitoring requirements.

[0054] Before the second fused data set is input into the initial water quality model, the following steps are further included: The data correlation values between the plurality of image data, the plurality of hyperspectral data, and the plurality of microwave remote sensing data are determined. According to the data correlation values, relevant image data, relevant hyperspectral data, and relevant microwave remote sensing data are selected, and the data correlation values between the relevant image data, the relevant hyperspectral data, and the relevant microwave remote sensing data are higher than a relevant threshold value. The hyperspectral data is data obtained by fine measurement of light of different wavelengths, and can provide rich material composition information; the microwave remote sensing data is information obtained by using electromagnetic waves in the microwave band, and has advantages in obtaining information hidden under the ground or less affected by weather. The correlation values between multiple image data, multiple hyperspectral data and multiple microwave remote sensing data are calculated, for example, the Pearson correlation coefficient between two data variables is calculated, to reflect the correlation between them. For example, the acquisition source of data with high correlation, such as the acquired water area position and the acquisition time, may be the same.

[0055] After obtaining the correlation values between various data, a correlation threshold is determined, which is a standard value determined according to specific requirements and actual conditions. Only when the data correlation values between image data, hyperspectral data and microwave remote sensing data are higher than the set threshold, these data will be selected. For example, if the correlation value between image data A and hyperspectral data B is greater than the threshold, and the correlation value between hyperspectral data B and microwave remote sensing data C is also greater than the threshold, and the correlation value between image data A and microwave remote sensing data C is also greater than the threshold, then image data A, hyperspectral data B and microwave remote sensing data C will be selected. These selected data are the relevant image data, relevant hyperspectral data and relevant microwave remote sensing data. The selected relevant image data, relevant hyperspectral data and relevant microwave remote sensing data are integrated and combined into a new data set, i.e. a second fusion data set. The second fusion data set contains different types of data with high correlation between each other, which will be used for subsequent input into the initial water quality model to improve the accuracy and reliability of the water quality model.

[0056] Step 105, adjusting the model parameters of the initial water quality model according to the parameter difference until the water quality parameter inversion model is obtained.

[0057] Determine the data error value of any one of the image data, hyperspectral data and microwave remote sensing data; According to the data error value in different water area types, the model parameters of the initial water quality model are adjusted by using the back propagation method until the water quality parameter inversion model is obtained.

[0058] Considering that the data set as the training model may have errors due to the acquisition problem itself, the embodiment can match the predicted value and the true value one by one according to the comparison of the predicted water quality parameter deviation and the tolerance, obtain the deviation data of the model output exceeding the preset threshold in the complex water body environment, and determine the feature distribution range of the first fusion data set corresponding to the deviation data, according to the influence of the regional difference.

[0059] Exemplarily, in the field of water body monitoring, the predicted value and the true value can be matched one by one by the data comparison tool according to the comparison of the response deviation and the deviation tolerance. In principle, such comparison is to identify the accuracy difference of the prediction result in the complex water body environment.

[0060] For example, when monitoring a river, the predicted pollution index is 75, and the true value is 70, the deviation is 5. If the preset deviation tolerance is 3, the data is marked as exceeding the threshold, and the influence of regional difference needs to be further analyzed. Such comparison helps to find the potential abnormal distribution range and provides a basis for subsequent correction.

[0061] According to the parameter difference value and the pre-training parameter, the model parameters of the initial water quality model are adjusted until the water quality parameter inversion model is obtained. The pre-training parameter is used to represent the correlation value between the image data, the hyperspectral data and the microwave remote sensing data in the same water area type.

[0062] If the deviation of the corrected data record and the true value still exceeds the deviation tolerance range, the first fusion data set is corrected again by the data calibration tool combined with the weight distribution of the environmental influence, and whether the corrected data meets the expected range is judged. According to the corrected data, the data integration tool is used to associate the correction result with the data matching requirement, obtain the inversion correction result of the final model, and obtain the water quality parameter inversion model.

[0063] In a possible implementation, for the determination of the abnormal distribution range, the influence of the regional difference can be analyzed.

[0064] Specifically, in the above river monitoring, local water quality fluctuations may be caused by upstream industrial emissions, and the predicted value may be consistently higher in some areas. By screening the areas with a deviation exceeding the threshold of 5 using the data comparison tool, it is determined that the abnormal distribution range covers about 2 kilometers of the middle reaches of the river. The parameter optimization tool is particularly important for adjusting the parameters iteratively in a back propagation manner for the abnormal distribution range. For example, in river monitoring, the initial model parameters may cause the prediction to be too high, and through multiple iterations of adjusting the parameters, the preliminary corrected pollution index is reduced from 75 to 72. This adjustment process can effectively improve the preliminary accuracy of the data record.

[0065] Preferably, if the preliminary corrected data record still exceeds the deviation tolerance range, it is necessary to use the data calibration tool to perform secondary correction in combination with the environmental impact weight distribution.

[0066] Specifically, in river monitoring, if the corrected pollution index is 72, the true value is 70, and the deviation is still 2, which exceeds the tolerance of 1, environmental factors such as rainfall dilution of water quality need to be considered, and after giving it a weight, it is corrected to 71. This way can be closer to the actual environmental changes.

[0067] In one embodiment, for the corrected data, the data integration tool processes the correction result in association with the data matching requirements to obtain the final inversion correction result.

[0068] For example, in river monitoring, the corrected pollution index of 71 needs to match the monitoring target such as water quality grade standard, if the standard requires less than 70, the final output value may be adjusted to 70 in combination with historical trends. This associated processing ensures that the output value adapts to the complex water environment.

[0069] The determination of the abnormal feature distribution range provides a direction for parameter optimization, and the secondary correction further fits the actual environment, and the final integration ensures the applicability of the result. This multi-level processing method can significantly improve the reliability of water body monitoring.

[0070] The present application also provides a construction device of a water quality parameter inversion model, as shown in Figure 2 The construction device comprises: A first acquisition module 21 is configured to acquire image data, hyperspectral data, and microwave remote sensing data obtained by image detection, hyperspectral detection, and microwave remote sensing detection on a reference water area, respectively, and an actual water quality parameter corresponding to the reference water area, wherein the reference water area comprises at least one type of water area. A second acquisition module 22 is configured to perform cross-modal mapping on the image data, the hyperspectral data, and the microwave remote sensing data by using a multi-modal feature alignment algorithm to obtain a fusion data set in which the image data, the hyperspectral data, and the microwave remote sensing data are fused. The constructing module 23 is configured to construct an initial water quality model by using a first fusion data set and a first actual water quality parameter, the first fusion data set and the first actual water quality parameter correspond to each other, the first fusion data set is part of the fusion data set, and the first actual water quality parameter is part of the actual water quality parameter; The third obtaining module 24 is configured to input a second fusion data set into the initial water quality model, obtain a predicted water quality parameter output by the initial water quality model, and determine a parameter difference between the predicted water quality parameter and a second actual water quality parameter, the second fusion data set is part of the fusion data set, the second fusion data set is different from the first fusion data set, the second actual water quality parameter is part of the actual water quality parameter, the second actual water quality parameter is different from the first actual water quality parameter, the second actual water quality parameter corresponds to the second fusion data set, and the second actual water quality parameter is different from the first actual water quality parameter. The adjusting module 25 is configured to adjust a model parameter of the initial water quality model according to the parameter difference until a water quality parameter inversion model is obtained.

[0071] Based on the understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the method of various implementation scenarios of the present application.

[0072] Optionally, the above-mentioned entity device can further include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface can include a display screen (Display), an input unit such as a keyboard (Keyboard), etc. The optional user interface can further include a USB interface, a card reader interface, etc. The network interface can optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc.

[0073] Those skilled in the art can understand that the above-mentioned entity device structure provided by the embodiment does not constitute a limitation on the entity device, and can include more or fewer components, or combine certain components, or different component arrangements.

[0074] Based on the above Figure 1The method shown, the embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method corresponding to any embodiment. The storage medium can also include an operating system, a network communication module. The operating system is a program for managing the hardware and software resources of the above-mentioned entity device, supporting the running of information processing programs and other software and / or programs. The network communication module is used to realize the communication between the components in the storage medium and the communication between other hardware and software in the information processing entity device.

[0075] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary general hardware platform, or by hardware. By applying the scheme of the present embodiment, compared with the prior art, the present embodiment synchronously acquires three kinds of remote sensing data: visible light image, hyperspectral data and microwave remote sensing data, and collects the actual water quality parameters of the corresponding water area, establishes a common feature space of the image data, the hyperspectral data and the microwave remote sensing data as a fusion data set, realizes the construction of an initial water quality model, iterates the initial water quality model based on the predicted water quality parameters output by the initial water quality model, so that the finally generated water quality parameter inversion model can overcome the limitations of a single sensor, realize the calculation of the water quality parameters based on the multi-dimensional information of visible light, hyperspectral and microwave remote sensing, improve the detection precision, enhance the adaptability to face complex water body environment, and meet the actual detection demand.

[0076] It should be noted that, in this document, relational terms such as“first” and“second”, and the like, are used solely to distinguish one entity or action from another entity or action, without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms“comprises”,“comprising”,“includes”,“including” or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by“comprises a...” does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0077] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications of these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments described herein, but will conform to the widest scope consistent with the principles and novel features applied herein.

Claims

1. A method for constructing a water quality parameter inversion model, characterized in that: include: Obtaining image data, hyperspectral data, and microwave remote sensing data obtained when performing image detection, hyperspectral detection, and microwave remote sensing detection on a reference water area, respectively, and actual water quality parameters corresponding to the reference water area, wherein the reference water area includes at least one type of water area; Using a multimodal feature alignment algorithm, cross-modal mapping is performed on the image data, the hyperspectral data, and the microwave remote sensing data to obtain a fused data set that is fused with the image data, the hyperspectral data, and the microwave remote sensing data; An initial water quality model is constructed using a first fused data set and a first actual water quality parameter, wherein the first fused data set corresponds to the first actual water quality parameter, the first fused data set is partial data of the fused data set, and the first actual water quality parameter is partial data of the actual water quality parameter; Inputting a second fused data set into the initial water quality model, obtaining a predicted water quality parameter output by the initial water quality model, and determining a parameter difference between the predicted water quality parameter and a second actual water quality parameter, wherein the second fused data set is partial data of the fused data set, and the second fused data set is different from the first fused data set, and the second actual water quality parameter is partial data of the actual water quality parameter, and the second actual water quality parameter is different from the first actual water quality parameter, and the second actual water quality parameter corresponds to the second fused data set; According to the parameter difference, the model parameters of the initial water quality model are adjusted until a water quality parameter inversion model is obtained.

2. The method according to claim 1, characterized in that The method of performing cross-modal mapping on the image data, the hyperspectral data, and the microwave remote sensing data using a multimodal feature alignment algorithm to obtain a fused data set that fuses the image data, the hyperspectral data, and the microwave remote sensing data includes: Using a multimodal feature alignment algorithm, cross-modal mapping is performed on the spectral features in the image data, the band information in the hyperspectral data, and the texture features in the microwave remote sensing data to obtain a first data set of aligned spectral features, band information, and texture features; A deep convolutional network is used to determine the distribution values ​​of the first data set in different water types, and a plurality of the distribution values ​​and the first data sets corresponding to the distribution values ​​are combined into a fused data set.

3. The method according to claim 2, characterized in that The combining of the plurality of distribution values ​​and the first data set corresponding to the distribution values ​​into a fused data set includes: Calculating, by using a weighted attention mechanism, a first weight value of the spectral feature, a second weight value of the band information, and a third weight value of the texture feature in the first data set; Comparing the first weight value, the second weight value, and the third weight value with the preset weight value respectively; If at least one of the first weight value, the second weight value, and the third weight value is less than the preset weight value, data enhancement is performed on the first data set until the enhanced first weight value, the second weight value, and the third weight value are respectively greater than the preset weight value; The plurality of distribution values ​​and the first data set enhanced with data corresponding to the distribution values ​​are combined to form a fused data set.

4. The method according to claim 1, wherein The step of adjusting the model parameters of the initial water quality model according to the parameter difference until a water quality parameter inversion model is obtained comprises: Determine the data error value of any of the image data, hyperspectral data and microwave remote sensing data; According to the data error values ​​in different water types, the model parameters of the initial water quality model are adjusted using the back propagation method until a water quality parameter inversion model is obtained.

5. The method according to claim 1, wherein Before inputting the second fused data set into the initial water quality model, the method further includes: determining data correlation values ​​among the plurality of image data, the plurality of hyperspectral data, and the plurality of microwave remote sensing data; selecting relevant image data, relevant hyperspectral data, and relevant microwave remote sensing data according to the data correlation value, wherein the data correlation value among the relevant image data, the relevant hyperspectral data, and the relevant microwave remote sensing data is higher than a correlation threshold; The relevant image data, the relevant hyperspectral data, and the relevant microwave remote sensing data are combined into a second fused data set.

6. The method according to claim 1, characterized in that The step of adjusting the model parameters of the initial water quality model according to the parameter difference until a water quality parameter inversion model is obtained comprises: According to the parameter difference and pre-training parameters, the model parameters of the initial water quality model are adjusted until a water quality parameter inversion model is obtained. The pre-training parameters are used to represent the correlation values ​​between the image data, hyperspectral data and microwave remote sensing data in the same water area type.

7. The method according to claim 1, characterized in that The image data, hyperspectral data and microwave remote sensing data obtained when performing image detection, hyperspectral detection and microwave remote sensing detection on the reference water area respectively include: Obtaining initial image data, initial hyperspectral data, and initial microwave remote sensing data obtained when performing image detection, hyperspectral detection, and microwave remote sensing detection on the reference waters, respectively; Projecting the initial image data, the initial hyperspectral data, and the initial microwave remote sensing data into a preset spatial scale using a grid resampling method, and determining the mapped image data, the mapped hyperspectral data, and the mapped microwave remote sensing data contained in the preset spatial scale; Obtain the resolution requirements corresponding to the water type; The resolution requirement is used to adjust the alignment accuracy between the mapped image data, the mapped hyperspectral data and the mapped microwave remote sensing data in the same preset spatial scale, and the mapped image data, the mapped hyperspectral data and the mapped microwave remote sensing data after the alignment accuracy adjustment are used as image data, hyperspectral data and microwave remote sensing data.

8. A device for constructing a water quality parameter inversion model, characterized in that: include: a first acquisition module, configured to acquire image data, hyperspectral data, and microwave remote sensing data obtained when performing image detection, hyperspectral detection, and microwave remote sensing detection on a reference water area, respectively, and actual water quality parameters corresponding to the reference water area, wherein the reference water area includes at least one type of water area; a second acquisition module, configured to perform cross-modal mapping on the image data, the hyperspectral data, and the microwave remote sensing data using a multimodal feature alignment algorithm to obtain a fused data set that is a fusion of the image data, the hyperspectral data, and the microwave remote sensing data; a construction module, configured to construct an initial water quality model using a first fused data set and a first actual water quality parameter, wherein the first fused data set corresponds to the first actual water quality parameter, the first fused data set is partial data of the fused data set, and the first actual water quality parameter is partial data of the actual water quality parameter; a third acquisition module, configured to input a second fused data set into the initial water quality model, obtain a predicted water quality parameter output by the initial water quality model, and determine a parameter difference between the predicted water quality parameter and a second actual water quality parameter, wherein the second fused data set is partial data of the fused data set, the second fused data set is different from the first fused data set, the second actual water quality parameter is partial data of the actual water quality parameter, the second actual water quality parameter is different from the first actual water quality parameter, and the second actual water quality parameter corresponds to the second fused data set; The adjustment module is used to adjust the model parameters of the initial water quality model according to the parameter difference until a water quality parameter inversion model is obtained.

9. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images