Evaluation method of spatial index visualization prediction model based on hyperspectral imaging

Abstract

This invention provides an evaluation method for a spatial indicator visualization prediction model based on hyperspectral imaging. First, evaluation metrics are set for visualization prediction accuracy. Next, evaluation metrics are designed for unreasonable visualization values ​​and image quality. By analyzing the visualization evaluation results from both prediction accuracy and image quality perspectives, the visualization effect can be precisely quantified. This method analyzes the visualization effect and provides a basis for optimizing the prediction model. This method addresses the prior art issue of applying a mean spectrum-based prediction model to pixel spectrum for visualization prediction, typically resulting in sensory evaluation and inconsistent prediction accuracy. It allows for precise quantification of the visualization effect.

Description

Technical Field

[0001] The present invention is an evaluation method for a spatial indicator visualization prediction model based on hyperspectral imaging. It can quantitatively evaluate the visualization effect of spatial indicators and optimize the visualization effect based on this. It can provide a theoretical basis and method support for accurate visualization prediction and belongs to the field of hyperspectral image processing. Background Art

[0002] Hyperspectral imaging technology is a technique that can simultaneously reflect an object's external features, surface defects, and stains, as well as its internal physical structure and chemical composition. Applying a predictive model to each pixel's spectrum reveals the spatial distribution of the indicator being measured, and visualization reveals the spatial distribution of the object being measured. Examples include the Chinese patent publication number CN111579506B, "Multi-camera hyperspectral imaging method, system, and medium based on deep learning," and the Chinese patent publication number CN105158186B, "A method for detecting black core in white radish based on hyperspectral imaging."

[0003] However, currently, visualization results are rarely analyzed, and there are no dedicated metrics to evaluate them, making it difficult to verify their accuracy and determine whether optimization is necessary. Analyzing and optimizing visualization results is a pressing task in this field, and effective visualization evaluation and optimization methods will drive the development of hyperspectral imaging technology.

[0004] For example, in the prior art, when a prediction model based on mean spectrum is applied to pixel spectrum for visualization prediction, it is usually evaluated only from a sensory perspective, and the prediction accuracy of the two is inconsistent. Summary of the Invention

[0005] To address the challenges of existing technologies, this paper first establishes evaluation metrics for visualization prediction accuracy. Next, it designs evaluation metrics for unreasonable visualization values ​​and image quality. By analyzing the visualization evaluation results from both prediction accuracy and image quality perspectives, the visualization effect can be precisely quantified. This analysis of visualization results provides a basis for optimizing the prediction model.

[0006] The present invention specifically provides an evaluation method for a spatial indicator visualization prediction model based on hyperspectral imaging, comprising the following steps:

[0007] 1) Evaluate the prediction accuracy and precision of the prediction model respectively;

[0008] 2) Evaluate the prediction reliability of the prediction model;

[0009] The step 1) comprises:

[0010] 1.1) Evaluate the accuracy of the prediction model. The accuracy evaluation index expression is as follows:

[0011]

[0012] Among them, RMSE roi is the spatial indicator visualization accuracy evaluation index, m is the total number of samples, n is the total number of pixels in the i-th sample interest area, y i is the measured value of the i-th sample, f(x ij ) is the prediction model, x if The meaning of is the spectral data of the jth pixel point of the i-th sample, and j means the sequence number of the spectral data in the area of ​​interest;

[0013] If RMSE roi The actual value of is less than the threshold, then the prediction model is judged to be unqualified; if, RMSE roi If the actual value of is not less than the threshold, go to step 1.2);

[0014] 1.2) Evaluate the accuracy of the prediction model. The accuracy evaluation index expression is as follows:

[0015]

[0016] in, is the spatial indicator visualization accuracy evaluation index, m is the number of samples, n is the number of pixels in the interest area of ​​the i-th sample, y is the measured value of the i-th sample, f(x ij ) is the prediction model;

[0017] if, If the actual value of is less than the threshold, the prediction model is judged to be unqualified; if If the actual value of is not less than the threshold, the prediction reliability of the prediction model will continue to be evaluated;

[0018] The step 2) comprises:

[0019] 2.1) Evaluate the rationality of the spatial prediction of the prediction model. The expression of the unreasonable value ratio is as follows:

[0020]

[0021] Among them, urv represents the proportion of unreasonable prediction values ​​in the area of ​​interest, n ur Represents the number of unreasonable prediction values ​​of the prediction model in the area of ​​interest, n roi Represents the number of all pixels in the area of ​​interest;

[0022] If the actual value of urv is greater than the threshold, the prediction model is judged to be unqualified; if the actual value of urv is not greater than the threshold, go to step 2.2);

[0023] 2.2) The image information retention of the prediction model is used as the final indicator of spatial indicator visualization image quality evaluation to evaluate the prediction model. The expression of image information retention is expressed by the inverse coefficient of variation, which is as follows:

[0024]

[0025] Among them, icv is the spatial index visualization image quality evaluation index, R m is the mean value of the pixel values ​​in the region of interest, R s Standard deviation of pixel values ​​within the region of interest;

[0026] If the actual value of icv is greater than the threshold, the prediction model is judged to be qualified; if the actual value of icv is not greater than the threshold, the prediction model is judged to be unqualified. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 In the embodiment, the ICV value of the original spectral reflectance image;

[0028] Figure 2 This is a diagram showing the distribution of flavonoid content in Ginkgo biloba leaves in the examples;

[0029] Figure 3 In the embodiment, the visualization results of the GA-PLSR model in the short-wave near-infrared band change with the number of principal components;

[0030] Figure 4(a) to Figure 4(d) In the embodiment, the visualization index of the GA-PLSR model in the short near-infrared band changes with the number of principal components; DETAILED DESCRIPTION

[0031] The present invention will be described below with reference to the accompanying drawings and specific embodiments.

[0032] An evaluation method for a spatial indicator visualization prediction model based on hyperspectral imaging includes the following steps:

[0033] 1) Visual prediction accuracy evaluation of spatial indicators;

[0034] 2) Spatial index visualization prediction image reliability evaluation method;

[0035] In step 1), an evaluation method is designed based on the accuracy and precision of the visual prediction. The visual prediction result is a pixel index space prediction. Due to the lack of accurate pixel index, the present invention designs an evaluation method from the perspective of mean prediction. The evaluation expression of accuracy is as follows:

[0036]

[0037] Where RMSE roiis the spatial indicator visualization accuracy evaluation index, m is the number of samples, n is the number of pixels in the interest area of ​​the i-th sample, y is the measured value of the i-th sample, and f is the prediction model

[0038] The accuracy evaluation expression is as follows:

[0039]

[0040] in is the spatial indicator visualization accuracy evaluation index, m is the number of samples, n is the number of pixels in the interest area of ​​the i-th sample, y is the measured value of the i-th sample, and f is the prediction model

[0041] In step 2), an evaluation method is designed based on the unreasonable value ratio of spatial prediction for the image quality reliability of visual prediction. The unreasonable value ratio expression is as follows:

[0042]

[0043] Among them, urv represents the proportion of unreasonable prediction values ​​in the area of ​​interest, n ur Represents the number of unreasonable prediction values ​​in the region of interest, n roi Represents the number of all pixels in the region of interest.

[0044] The expression of image information retention is expressed by the inverse coefficient of variation, which is as follows:

[0045]

[0046] Among them, icv is the spatial index visualization image quality evaluation index, R m is the mean value of the pixel values ​​in the region of interest, R s The standard deviation of pixel values ​​within the region of interest.

[0047] Formula 1 is used to evaluate the overall prediction accuracy of the object under test when visualizing the prediction.

[0048] Formula 2 is used to evaluate the overall prediction accuracy of the object being tested during visualization prediction.

[0049] Formula 3 is used to evaluate the proportion of unreasonable values ​​in the visual prediction results.

[0050] Formula 4 is used to evaluate the imaging quality of the visualization prediction result image and the degree of preservation of the original spectral image during visualization prediction.

[0051] According to the results reflected by the visualization evaluation indicators, the visualization prediction model is re-established.

[0052] In this method, n ur The method to obtain is:

[0053] Count the number of unreasonable prediction values ​​in the area of ​​interest

[0054]

[0055] where y pred is the predicted value, f u (y pred ) is the unreasonable value judgment formula, as shown below.

[0056]

[0057] R m The method to obtain is:

[0058]

[0059] R s The method to obtain is:

[0060]

[0061] Among them, t is the pixel number in the area of ​​interest, y t is the predicted value of the t-th pixel, is the mean value of pixels in the region of interest.

[0062] The following describes the method by taking the visual detection of flavonoids content in Ginkgo biloba leaves as an example.

[0063] This example uses ginkgo biloba as the sample. First, the unreasonable value ratio (URV) and the spatial index visualization image quality evaluation indicator (ICV) are limited. The maximum reasonable value range of URV is set between 0 and 3 times the maximum value of the sample's physical and chemical indicators. Since the maximum total flavonoid content of the ginkgo biloba sample in this experiment is 4.9748 mg / g, 15 mg / g is set as the maximum reasonable value.

[0064] When the icv value is larger, the values ​​of the pixels in the image are closer, and the image is smoother. When the icv value is small, the image quality is poor, and when the icv value is too large, the image will lose its original texture features. In order to ensure both image quality and image texture features, this method uses the icv value of the original hyperspectral image as a reference value. When the icv value of the visualization prediction result is closer to the original image, it means that the visualization result retains more features of the original image and the image quality is better. Calculate the icv values ​​of all original spectral reflectance images, and the results are as follows: Figure 1As shown. Figure 4 (a) and Figure 4 (b) show the ICV values ​​of each wavelength in the visible near-infrared and short-wave near-infrared spectral images in the leaf area, respectively. The minimum ICV value of the spectral image in the visible near-infrared band is 3.1464, and the maximum ICV value is 26.0918. The ICV value of the image with a wavelength less than 700nm is significantly lower than that of the image greater than 700nm. This is because the visible light band of 700nm is visible to the naked eye, so a large number of leaf texture features are retained within this band. The spectral image greater than 700nm is in the near-infrared band and is an invisible band. The texture features in this band are not obvious and the image is smoother, so the ICV value in this range is higher. In the short-wave near-infrared band, the minimum ICV value of the spectral image is 1.2296, and the maximum ICV value is 20.6446. In summary, the closer the ICV value of the visible near-infrared band visualization prediction map is to the range of 3.1464-26.0918, the higher the visualization prediction performance is considered to be; the closer the ICV value of the shortwave near-infrared band visualization prediction map is to the range of 1.2296-20.6446, the higher the visualization performance is considered to be.

[0065] Table 1

[0066]

[0067] Table 1 shows the visualization indicators of the PLSR model in the visible near-infrared and short-wave near-infrared bands. The RMSE of the visualization results is roi and Parameters are analyzed. RMSE of the visible and near infrared band visualization image roi It is higher than that of short-wave near-infrared, and It is also lower, indicating that the ability of the visible near-infrared band is not as good as that of the short-wave near-infrared band in predicting the mean accuracy of the total flavonoid content in leaves; among the prediction models based on characteristic bands, the GA-PLSR model of the short-wave near-infrared band is the best model, and its RMSE roi 0.5043 mg / g, The two indicators describe the accuracy of the visual prediction mean. Since the PLSR model used in this method is a linear model, the results are consistent with those of the mean prediction model.

[0068] As shown in Table 1, the visualization image of the PLSR model based on the full visible-near-infrared (VNIR) band achieved an icv value of 0.3371. The visualization results based on the characteristic bands selected by SPA and iPLS achieved icv values ​​of 0.0203 and 0.0305, respectively. The icv values ​​of the visualization results of these two models were less than 1% of the set value, indicating that the image quality of the visualization prediction images was very poor. Even though the mean prediction model's results were similar to those of other models, it could not be used for visual distribution prediction. The visualization images of the prediction models based on the characteristic bands selected by GA and siPLS within the VNIR band had icv values ​​closest to the set value range, indicating that these two models achieved the best visualization prediction performance. Similar patterns were observed when analyzing the visualization results of the short-wavelength NIR band using the same method. These results are consistent with the sensory evaluation results, indicating that this method can quantitatively describe the imaging quality of the visualization prediction images.

[0069] Based on the URV results of the above models, it was found that the three worst models screened out were consistent with the three worst models selected using sensory evaluation, indicating that this method can effectively detect excessive unreasonable values ​​in visualization results. Comparing the visible near-infrared band and the short-wave near-infrared band, it was found that the results of the visible near-infrared band were quite different, with a large number of unreasonable pixel prediction points, generally exceeding 30% in the short-wave near-infrared results.

[0070] Now we visualize the GA-PLSR prediction model for the shortwave near-infrared band. This model is the optimal model using the spectral prediction model, and its optimal number of principal components is 13. The results are as follows: Figure 3 When the number of principal components is 1-4, the visualization image of the total flavonoids content in Ginkgo biloba leaves is clear and similar. When the number of principal components is 5-8, vertical stripes appear faintly in the visualization image. When the number of principal components is 9-11, the stripes in the image begin to increase and the noise begins to increase. When the number of principal components increases from 12 to 20, the image is completely obscured by the noise.

[0071] Figure 4(a) to Figure 4(d) As shown in Figure 4(a), the RMSE roi , Figure 4(b) reflects , Figure 4 (c) reflects icv, Figure 4 (d) reflects urv) shortwave near infrared band GA-PLSR model visualization changes, among which when the number of principal components is 13, the performance of the mean spectrum prediction model is the best, and its RMSE roi 0.5043 mg / g, is 0.8482, while the urv value is 45.41% and the icv value is 0.3034. Although the model has good accuracy, the quality of the visual image is not high, and the prediction effect of the model on the pixel spectrum is not good. As the number of principal components increases, not only the prediction accuracy decreases, but also the quality of the visual image seriously decreases. Therefore, increasing the number of principal components does not improve the visualization effect; as the number of principal components decreases, the prediction performance of the model begins to decline, but the visualization effect of the image is improved to a certain extent; when the number of principal components drops to 9, the RMSE roi 0.6073 mg / g, The URV is 8.93% and the ICV is 1.2740, which are similar to those of the original spectral image. Therefore, the image noise is within a reliable range. When the number of principal components is 10, the ICV is 0.7662, significantly different from the minimum value of the original image. Therefore, to achieve both image quality and prediction accuracy, a principal component number of 9 is the optimal choice. The same method can also be used to adjust the visualization images of other models of the present invention.

Claims

1. An evaluation method for a spatial indicator visualization prediction model based on hyperspectral imaging, characterized by the following steps: include: 1) Evaluate the prediction accuracy and precision of the prediction model respectively; 2) Evaluate the prediction reliability of the prediction model; The step 1) comprises: 1.1) Evaluate the accuracy of the prediction model. The accuracy evaluation index expression is as follows: Among them, RMSE roi is the spatial indicator visualization accuracy evaluation index, m is the total number of samples, n is the total number of pixels in the i-th sample interest area, y i is the measured value of the i-th sample, f(x ij ) is the prediction model, x ij The meaning of is the spectral data of the jth pixel point of the i-th sample, and j means the sequence number of the spectral data in the area of ​​interest; If RMSE roi The actual value of is less than the threshold, then the prediction model is judged to be unqualified; if, RMSE roi If the actual value of is not less than the threshold, go to step 1.2); 1.2) Evaluate the accuracy of the prediction model. The accuracy evaluation index expression is as follows: in, is the spatial indicator visualization accuracy evaluation index, m is the number of samples, n is the number of pixels in the interest area of ​​the i-th sample, y is the measured value of the i-th sample, f(x ij ) is the prediction model; if, If the actual value of is less than the threshold, the prediction model is judged to be unqualified; if If the actual value of is not less than the threshold, the prediction reliability of the prediction model will continue to be evaluated; The step 2) comprises: 2.1) Evaluate the rationality of the spatial prediction of the prediction model. The expression of the unreasonable value ratio is as follows: Among them, urv represents the proportion of unreasonable prediction values ​​in the area of ​​interest, n ur Represents the number of unreasonable prediction values ​​of the prediction model in the area of ​​interest, n roi Represents the number of all pixels in the area of ​​interest; If the actual value of urv is greater than the threshold, the prediction model is judged to be unqualified; if the actual value of urv is not greater than the threshold, go to step 2.2); 2.2) The image information retention of the prediction model is used as the final indicator of spatial indicator visualization image quality evaluation to evaluate the prediction model. The expression of image information retention is expressed by the inverse coefficient of variation, which is as follows: Among them, icv is the spatial index visualization image quality evaluation index, R m is the mean value of the pixel values ​​in the region of interest, R s Standard deviation of pixel values ​​within the region of interest; If the actual value of icv is greater than the threshold, the prediction model is judged to be qualified; if the actual value of icv is not greater than the threshold, the prediction model is judged to be unqualified.

2. The method for evaluating a spatial indicator visualization prediction model based on hyperspectral imaging according to claim 1, wherein n ur The method to obtain is: Count the number of unreasonable prediction values ​​in the area of ​​interest n ur =∑f u (and pred ) where y pred is the predicted value, f u (y pred ) is the unreasonable value judgment formula, as shown below: R m The method to obtain is: R s The method to obtain is: Among them, t is the pixel number in the area of ​​interest, y t is the predicted value of the t-th pixel, is the mean value of pixels in the region of interest.

Images