Method for assessing air pollution using smart hyperspectral imaging

Abstract

The present invention provides a method for assessing air pollution using smart hyperspectral images. A hyperspectral image information is obtained from a reference image, and an environmental image is converted according to the hyperspectral image information to obtain a hyperspectral image. A plurality of first eigenvalues are obtained , and a principal component analysis operation is performed according to a wave band of an air pollutant to obtain a plurality of second eigenvalues. The second feature values are subjected to a convolution operation to obtain at least one image of the object to be measured is compared with at least one sample image to generate a air pollution result, thereby reducing the time required for the detection process and increasing the detection range.

Description

Methods for assessing air pollution using intelligent hyperspectral imaging This invention relates to a method for assessing air pollution, and more particularly to a method for assessing air pollution using intelligent hyperspectral imaging. Imaging spectrometers have a wide wavelength measurement range and can be divided into multispectral images and hyperspectral images. Hyperspectral images have a spectral resolution of less than or equal to 10 nm, while multispectral images have a spectral resolution of greater than 10 nm. Hyperspectral images are characterized by nanometer-level spectral intervals, thus they can measure ranges that are not visible to the human eye. Therefore, they can be used by drones or satellites. Drones can be used for aerial photography to understand the distribution of local rivers, vegetation, and other features. The most common imaging spectrometer is the pushbroom type. Using a pushbroom, spatial and spectral information can be received simultaneously to obtain one-dimensional spatial spectral information and two-dimensional spectral information. The two-dimensional spectral information is stacked to form a cubic image of three-dimensional hyperspectral data. Principal component analysis (PCA) can be used in fields such as face recognition and image compression. Its goal is to reduce the dimensionality of a dataset while retaining the principal components. It is a statistical analysis method that uses orthogonal transformations to perform linear transformations and project the values ​​of linearly uncorrelated variables. Artificial intelligence includes machine learning and deep learning. Machine learning is related to statistics and provides different and efficient algorithms for data analysis. This method requires a high degree of expertise and is time-consuming. Deep learning integrates feature learning and model building into a single model. It uses convolution operations to transfer features from each layer to the next. The model formed in this way consists of multiple non-linear hidden layer neural networks. Features are input into the model for classification and regression, and the parameters are further optimized. Convolutional neural networks consist of convolutional layers that obtain local images as input features while preserving the spatial arrangement of the images, pooling layers that use non-linear downsampling to obtain images that conform to the size of the display area and generate corresponding image thumbnails, features are extracted and parameters are reduced through convolutional and pooling layers, and fully connected layers perform subsequent classification and obtain corresponding probabilities. Convolutional neural networks include two-dimensional and three-dimensional ones. Two-dimensional convolutional neural networks only perform convolution calculations on images, while three-dimensional ones perform convolution calculations on cubes. Compared to two-dimensional convolutional networks, three-dimensional neural networks extract both spatial and temporal dimensions, and form a dataset by convolutional kernels on multiple consecutive frames and stacking them. With the continuous development of industrialization and urbanization, a large amount of pollutants are emitted into the atmosphere, causing air quality to deteriorate and threatening human health. In addition, air pollution has also caused serious damage to the ecosystem, affecting plant growth and animal habitats. Common air pollutants include sulfur dioxide (SO2), nitrogen oxides (NOx), sulfur oxides (SOx), ammonia (NH3), carbon monoxide (CO), ozone (O3), methane (CH4), and particulate matter (PM2.5 and PM10). Methods for detecting air pollution, such as on-site sampling, have problems such as limited detection range and time consumption. Therefore, a method is needed to solve the problems of on-site sampling. To address the aforementioned problems, this invention provides a method for assessing air pollution using intelligent hyperspectral imaging. The method obtains hyperspectral image information from a reference image, converts an environmental image into a hyperspectral image based on this information, obtains a first feature value from the hyperspectral image, performs principal component analysis based on the bands of air pollutants to obtain a second feature value, performs convolution operations on the second feature value to obtain an image of the object under test, and compares this image with a sample image to generate air pollution results. This method reduces the time required for the detection process and expands the detection range. One objective of this invention is to provide a method for assessing air pollution using intelligent hyperspectral imaging, and to use this method to assist in detecting the degree of air pollution in specific areas, thereby solving the problems of time-consuming, labor-intensive, and limited detection range that previously relied solely on manual on-site sampling. To achieve the above objectives, the present invention provides a method for assessing air pollution using intelligent hyperspectral imagery. The method includes the following steps: an image capturing unit captures a reference image based on a reference object and transmits it to a host computer; the host computer converts the reference image into hyperspectral image information; the image capturing unit captures an environmental image based on an environment and transmits it to the host computer; the host computer converts the environmental image into a hyperspectral image based on the hyperspectral image information; the host computer analyzes the hyperspectral image to obtain a plurality of first feature values; the host computer performs principal component analysis on these first feature values ​​based on a band corresponding to an air pollutant to simplify the generation of a plurality of corresponding second feature values; and finally, the host computer uses a plurality of convolution kernels... At least one convolution operation is performed on the second feature values ​​for spatial and temporal considerations to obtain at least one image of a test object. The at least one image of the test object includes a plurality of surrounding images and at least one selected image. The convolution kernels include a plurality of selected feature values, and the surrounding images correspond to the surrounding feature values. The host computer compares the at least one image of the test object with at least one sample image to generate an air pollution result. The host computer retrieves the at least one sample image from a database. The at least one sample image corresponds to hyperspectral image information and corresponds to air pollutants. This achieves the effects of reducing the detection time and labor costs of air pollution detection and increasing the detection range. The present invention provides an embodiment in which, in the step of the host performing at least one convolution operation on the second feature values ​​in terms of space and time based on a plurality of convolution kernels, the host sets the plurality of convolution kernels to normalize a plurality of image values ​​of the environment image to a plurality of pixel normal values, and multiplies the plurality of convolution kernels by the plurality of pixel normal values ​​to extract the second feature values ​​in a convolutional layer. In one embodiment of the present invention, in the step of the host performing at least one convolution operation on the second feature values ​​in terms of space and time based on a plurality of convolution kernels to obtain at least one object to be measured, the host integrates the regions where the selected feature values ​​are located to obtain at least one distribution region on the environmental image. The present invention provides an embodiment in which, in the step of the image capturing unit capturing an environmental image to the host based on an environment, the image capturing unit is a drone, and the environmental image is an aerial photograph of the environment. The present invention provides an embodiment in which the step of the host generating an air pollution result by comparing the at least one image of the object under test with at least one sample image further includes at least the following steps: the host performs simplification processing in at least one pooling layer; and the host performs comparison of the image of the object under test with the at least one sample image in at least one fully connected layer. The present invention provides an embodiment in which, in the step of the host generating an air pollution result by comparing the at least one object image with at least one sample image, when the host cannot identify the air pollution image based on the at least one sample image, the host performs an approximate comparison of the at least one sample image based on the at least one sample image. The present invention provides an embodiment in which, in the step of the host converting the reference image into a hyperspectral reference image based on the reference image, the host converts the reference image into a hyperspectral reference image through a plurality of color matching functions of a standard color chart, a correction matrix and a transformation matrix. The present invention provides an embodiment in which the method for assessing air pollution using intelligent hyperspectral imaging further includes at least the following steps: the host obtains a corresponding pollution level based on the air pollution result and the environmental image, and the host stains the image of at least one object to be tested based on the pollution level. The present invention provides an embodiment in which, in the step of the host performing a principal component analysis operation on the first feature values ​​according to a band corresponding to an air pollutant to simplify the hyperspectral image and generate a plurality of corresponding second feature values, a first number of the first feature values ​​is greater than a second number of the second feature values. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application. Obviously, the described embodiments are only some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without any innovative effort are within the scope of protection of this application. In the description of this application, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. It is known that when detecting air pollution in the environment, the sampling area in a single test is relatively small, and the degree of pollution in different locations of the entire environment is not the same. Therefore, it takes a certain amount of time to obtain sufficient samples. Accordingly, the present invention proposes a method for assessing air pollution using intelligent hyperspectral imaging. The image acquisition unit captures images of the environment and then analyzes them. Based on the analysis results, it determines whether the environment is polluted and the degree of pollution, thus solving the problems of time consumption and small detection range of manual detection. The invention will be described in detail below by way of the drawings illustrating various embodiments thereof. However, the concept of the invention may be embodied in many different forms and should not be construed as being limited to the exemplary embodiments set forth herein. First, please refer to Figure 1A, which is a flowchart of image recognition according to one embodiment of the present invention. As shown in the figure, a method for assessing air pollution using intelligent hyperspectral imagery according to the present invention is applied to a host computer 10 to perform the following steps: Step S10: The image capturing unit captures a reference image to the host based on the reference object; Step S12: The host converts the reference image into a hyperspectral reference image and obtains hyperspectral image information based on the hyperspectral reference image; Step S14: The image capturing unit captures images to the host computer based on the environment; Step S16: The host computer converts the environmental image based on the hyperspectral image information to obtain a hyperspectral image; Step S18: The host analyzes the hyperspectral image to obtain the first feature value; Step S20: The host performs principal component analysis on the first eigenvalue based on the corresponding band of the air pollutant, and simplifies it to generate the corresponding second eigenvalue. Step S22: The host performs convolution operations on the second feature value in terms of space and time according to the convolution kernel to obtain the image of the object to be tested; Step S24: The host compares the image of the object to be tested with the sample image to generate air pollution results, and the host extracts the pre-stored sample image from its database. Step S24 further includes the following steps: Step S242: The host performs pooling processing. Step S244: The host performs a comparison between the image of the object under test and the sample image at the fully connected layer. Following the above, and referring to Figures 2 and 3A to 3D, which are schematic diagrams of aerial photography and some steps of the image capturing unit in one embodiment of the present invention, the detection system 1 used in conjunction with the method for assessing air pollution using intelligent hyperspectral imaging of the present invention includes a host 10 and an image capturing unit 20. In this embodiment, the host 10 is exemplified by having a processing unit 12, a memory 14, and a storage unit 16, but it is not limited to this. It can also be a server, a laptop, a tablet computer, or any electronic device with computing power as referred to in the present invention. The storage unit 16 contains a database 30, but it is not limited to this. It can also be an external storage unit of the host 10. The host 10 executes a convolution program P through the processing unit 12 to correspondingly build a convolutional neural network (CNN). In addition, the image capturing unit 20 in this embodiment is an aerial photography drone that flies above the target to be photographed. In this embodiment, the target is the surrounding landscape environment. In step S10, as shown in Figure 3A, the host 10 reads a reference image REF corresponding to the image capturing unit 20. The reference image REF includes at least an object reference image O1R and a background reference image BGR. The reference image REF may be a reference color block stored in the database 30 or captured by the image capturing unit 20 from the 24 color blocks. In step S12, the host 10 obtains a hyperspectral image information HSA based on the reference image REF. That is, it uses Visible Hyperspectral Imaging (VIS-HSI) to perform calculations on the reference image REF provided by the image capturing unit 20 to obtain a conversion equation for converting the general image color gamut space to the XYZ color gamut space (CIE 1931 XYZ color space). The hyperspectral image information HSA of this invention corresponds to 24 color patches. The 24 color patches include red, blue, green, gray and other major colors commonly found in nature, and include multiple color matching functions of a standard color card, a correction matrix and a conversion matrix. Following the above, the conversion step requires converting the reference image REF and the spectrometer (Ocean Optics, QE65000) to the same XYZ color gamut space. The conversion formula is as follows: Formula (1) in, Formula (II) Formula (3) Formula (IV) This is the gamma function, which can convert sRGB values ​​to RGB values; For the transformation matrix, The color adaptation transformation matrix can convert linear RGB values ​​to XYZ values ​​(XYZ Endoscope) according to Equation (I). The formula for converting the reflectance spectrum data captured by the spectrometer to the XYZ color space is as follows: Formula (5) Formula (VI) Formula (VII) Where k is as shown in equation (8). Formula (8) , , For color matching functions (CMF), The light source spectrum captured by the drone's camera, since the Y value in the XYZ color gamut space is proportional to the brightness, is obtained using equation (8) to determine the maximum brightness (saturation brightness) of the light source spectrum Y value. Then, by setting the upper limit of the Y value to 100, the brightness specification ratio k is obtained. Therefore, equations (5) to (7) are used to convert the reflection spectrum data to the XYZ values ​​specified in the XYZ color gamut space. XYZ Spectrum ]. Furthermore, image correction of the camera is performed using the correction matrix C in equation (9): Formula (9) The nonlinear response correction is performed using a third-order equation, and the nonlinear response correction of the camera is performed using the following equation (XI): Formula (10) In the camera's dark current section, the dark current is typically a fixed value that does not change with variations in the amount of light entering the camera; therefore, the dark current correction factor is defined as V. Dark The effect of dark current is corrected using the following formula: Formula (XI) The correction variable for inaccurate color separation of color filters and color shift is defined as follows: , , , This is a color matching function for converting the RGB color space to the XYZ color space, therefore based on... , , The correlation between the three can be expressed in the following formula (12) as a permutation and combination of the possibilities between X, Y, and Z, to correct the inaccuracy of color separation and color shift of camera images using color filters: Formula (12) By applying equations (x) to (xii) above, the variable matrix V corrected by equation (xiii) is derived: V=[X 3 Y 3 Z 3 X 2 YX 2 ZY 2 Z XY 2 XZ 2 YZ 2 XYZ X 2 Y 2 Z 2 XY XZ YZ XYZ α] T Formula (13) By combining the variable matrix V with the correction matrix C, the corrected X, Y, and Z values ​​[XYZ] are obtained. Correct As shown in equation (xiv): Formula (XIV) Calculate [XYZ] Correct ] and [XYZ Spectrum The root mean square error of both is 1.26. Before using CIE DE2000 to calculate color difference, [XYZ] must be... Correct ] and [XYZ Spectrum The conversion from the XYZ color space to the Lab color space is shown in equations (xv) to (xvii) below: Formula (15) Formula (16) Formula (17) in, As shown in equation (18): Formula (18) The XYZ values ​​obtained after camera calibration (XYZ Correct ) and the 24-color patch reflectance spectrum data measured by the spectrum analyzer (R Spectrum To analyze the data, principal component analysis was first performed on the reflectance spectrum data of the 24 color blocks to obtain the corresponding principal component scores and XYZ. Correct Multivariate regression analysis was performed, and the transformation matrix M was obtained after integration. In multivariate regression analysis, V is used. Color The variables are used to obtain the transformation matrix M through the following equation (19): Formula (19) After transformation matrix, the analog spectrum is obtained by the following equation (20): Formula (20) 24-color patch reflectance spectrum data (R Spectrum ) and 24-color block analog spectrum (S Spectrum By comparison, the root mean square error of each color block was calculated, with an average error of 0.0532 and an average color difference of 4.58. Therefore, the visible light super-spectrum technology established by the above process in this invention simulates the reflection spectrum of the RGB values ​​captured by the camera, thereby obtaining a visible light super-spectrum image. In step S14, the image capturing unit 20 is moved above an environment E, and an environmental image IMG is obtained based on the environment E and sent to the host 10, and step S16 is executed. In step S16, the host 10 confirms that the resolution of the environmental image IMG meets a resolution threshold, converts the environmental image IMG into a hyperspectral image HYI based on the hyperspectral image information HSA, the environmental image IMG includes at least one object image O1 and a plurality of background images BG, and executes step S18. In step S18, as shown in Figure 3B, the host 10 extracts a plurality of first feature values ​​F1 based on the hyperspectral image HYI. In step S20, as shown in Figure 3C, the first feature values ​​F1 are subjected to principal component analysis based on a band D1 of an air pollutant D, and the background images BG are filtered out to obtain a filtered image FM with a plurality of second feature values ​​F2. Step S22 is then executed, wherein the first number of the first feature values ​​F1 is greater than the second number of the second feature values ​​F2. Continuing from the above, in step S20, the air pollutant D and its corresponding band D1 are pre-stored in the database 30. Continuing from the above, in step S20, the air pollutant D is sulfur dioxide, and the corresponding band D1 is the infrared band with wavelengths of 18.9 µm, 8.8 µm and 7.6 µm. Continuing from the above, in step S20, the air pollutant D is nitrogen monoxide, and the corresponding band D1 is the infrared band with a wavelength of 5.3 µm. Continuing from the above, in step S20, the air pollutant D is nitrogen dioxide, and the corresponding band D1 is the infrared band with wavelengths of 6.2 µm, 7.5 µm and 13.3 µm. Continuing from the above, in step S20, the air pollutant D is carbon monoxide, and the corresponding band D1 is the infrared band with a wavelength of 4.6 µm. Continuing from the above, in step S20, the air pollutant D is ozone, and the corresponding band D1 is the infrared band with a wavelength of 9.3 µm ~ 10.4 µm and the ultraviolet band with a wavelength of 254 nm. In step S22, and referring to Figure 5, which is a schematic diagram of a convolutional neural network according to an embodiment of the present invention, the host 10 performs at least one convolution operation on the second feature values ​​F2 based on a plurality of convolution kernels C in terms of space and time. The convolution kernels C include a plurality of selected feature values ​​F22 corresponding to at least one object image O1 and a plurality of peripheral feature values. The host 10 sets the convolution kernels C to normalize a plurality of image values ​​of the environment image IMG to a plurality of pixel normal values. The convolution kernels C are multiplied by the pixel normal values ​​to extract the second feature values ​​F2 in a convolutional layer. The host 10 integrates the regions where the selected feature values ​​F22 are located in the environment image IMG to obtain at least one distribution region. The background images BG that do not contain the at least one object image O1 are filtered out by the peripheral feature values ​​to obtain the at least one object image O1 and then step S24 is executed. In step S24, and referring to Figure 1B, which is a flowchart of comparing air pollution results according to an embodiment of the present invention, as shown in Figure 3D, after the processing unit 12 locates the position of at least one object image O1 to be tested, the host 10 retrieves at least one sample image SA from the database 30. The host 10 compares the at least one object image O1 to be tested with the at least one sample image SA to obtain an air pollution result CR. Continuing from the above, in step S24, the at least one sample image SA corresponds to the hyperspectral image information HSA, and the at least one sample image SA corresponds to the air pollutant D. Following the above, in step S242, the host 10 simplifies the image O1 of at least one object under test in at least one pooling layer, reducing some parameters. In step S244, the host 10 integrates and classifies the image O1 of at least one object under test after passing through the convolution kernels C and the at least one pooling layer in at least one fully connected layer and outputs an output result. The output result is compared with the at least one sample image SA to obtain the air pollution result CR. Continuing from the above, and referring to Figures 1C and 4, this embodiment can further colorize the image according to the degree of air pollution, and the steps are as follows: Step S26: The host computer obtains the corresponding pollution level based on the air pollution results and environmental images; Step S28: The host computer stains the image of the object to be tested according to the degree of contamination. In step S26, the air pollution result CR includes the different pollution levels in different areas of the environmental image IMG. Therefore, the host 10 obtains the pollution level of the at least one distribution area in the corresponding environmental image IMG based on the air pollution result CR. In step S28, as shown in Figure 4, which is a schematic diagram of image coloring according to an embodiment of the present invention, the host 10 colors the image of the object to be tested according to the different degrees of contamination in different areas. No contamination is colored with a first color TA1, such as green or blue. The degree of contamination is lower, and a second color TA2 is colored, such as yellow or orange. The degree of contamination is higher, and a third color TA3 is colored, such as red or black. The embodiments described above provide a method for assessing air pollution using intelligent hyperspectral imaging. This method involves using a drone to capture and analyze images at high altitudes, and then coloring the images according to the level of pollution analyzed. This increases the detection range and reduces the sampling time for detecting air pollution in a specific area. The following will further explain the characteristics contained in this work: A user uses a drone to take aerial photos in an environment. A host 10 reads a reference image (REF) from the drone and obtains the corresponding hyperspectral image information (HSA). The host 10 reads an environmental image (IMG) obtained from the aerial photography area. The host 10 converts the environmental image IMG into a hyperspectral image (HYI) based on the hyperspectral image information (HAS) and extracts a plurality of first feature values ​​(F1). The host 10 performs principal component analysis on these first feature values ​​according to a band D1 corresponding to an air pollutant (D) to remove a plurality of background images (BG) from the environmental image IMG, obtaining a filtered image (FM) that retains at least one image (O1) of the object under test. The filtered image (FM) has multiple features. Several second feature values ​​F2, one of which has a second quantity less than one of the first feature values ​​F1, are used. The host 10 performs at least one convolution operation on these second feature values ​​F2 in terms of time and space based on a plurality of convolution kernels C. These convolution kernels C include a plurality of selected feature values ​​F22 corresponding to at least one object image O1 and a plurality of peripheral feature values. After the at least one convolution operation, the host 10 integrates these selected feature values ​​F22 to obtain at least one distribution area of ​​the environmental image. The host 10 retrieves at least one sample image SA from a database 30 and compares it with the at least one object image O1 to obtain an air pollution result CR. Continuing from the above, the air pollution image CR includes different pollution levels and at least one distribution area. The pollution levels include no pollution, low pollution, and high pollution. The area corresponding to no pollution is colored green, the area corresponding to low pollution is colored yellow, and the area corresponding to high pollution is colored red. The embodiments described above provide a method for assessing air pollution using intelligent hyperspectral imaging. This method involves using a drone to capture and analyze images at high altitudes, and then coloring the images according to the level of pollution analyzed. This increases the detection range and reduces the sampling time for detecting air pollution in a specific area. Therefore, this invention is indeed novel, inventive, and industrially applicable, and undoubtedly meets the requirements for patent application under the Patent Law of our country. Thus, we hereby file an invention patent application in accordance with the law, and earnestly pray that the Bureau will grant the patent as soon as possible. However, the above description is only a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes and modifications made to the shape, structure, features and spirit described in the claims of the present invention should be included in the scope of the claims of the present invention. 1: Detection System 10: Host 12: Processing Unit 14: Memory 16: Storage Unit 20: Image Capture Unit 30: Database BG: Background Image BGR: Background Reference Image C: Convolutional kernel CNN: Convolutional Neural Network CR: Air Pollution Results D: Air pollutants D1: Band E: Environment F1: First eigenvalue F2: Second eigenvalue F22: Selecting eigenvalues FM: Filter out images HSA: Hyperspectral Imaging Information HYI: Hyperspectral Imagery IMG: Environmental Imagery O1: Image of the object under test O1R: Object Reference Image P: Convolution program REF: Reference Image S10~S24: Steps S242~S244: Steps S26~S28: Steps SA: Sample Image TA1: First staining TA2: Second staining TA3: Third staining Figure 1A: A flowchart of image recognition according to one embodiment of the present invention; Figure 1B: A flowchart of comparing air pollution results according to one embodiment of the present invention; Figure 1C: A flowchart of image coloring according to one embodiment of the present invention; Figure 2: A schematic diagram of aerial photography performed by the image capturing unit according to one embodiment of the present invention; Figures 3A to 3D: Schematic diagrams of some steps of one embodiment of the present invention; Figure 4: A schematic diagram of image coloring according to one embodiment of the present invention; and Figure 5: A schematic diagram of a convolutional neural network according to one embodiment of the present invention. S10~S24: Steps

Claims

1. A method for assessing air pollution using intelligent hyperspectral imagery includes the following steps: An image capturing unit captures a reference image based on a reference object and transmits it to a host computer; The host computer converts the reference image into a hyperspectral reference image and obtains hyperspectral image information based on the hyperspectral reference image; The image capturing unit captures an environmental image based on an environment and transmits it to the host computer; The host computer converts the environmental image based on the hyperspectral image information to obtain a hyperspectral image; The host computer analyzes the hyperspectral image to obtain a plurality of first feature values; The host computer performs principal component analysis on the first feature values ​​based on a band corresponding to an air pollutant to simplify the hyperspectral image and generate a plurality of corresponding second feature values; The host computer performs at least one convolution operation on the second feature values ​​spatially and temporally using a plurality of convolution kernels to obtain at least one image of a target object, the at least one image of the target object including a plurality of surrounding images and at least one selected image, wherein... The convolutional kernels include a plurality of selected feature values ​​and a plurality of peripheral feature values, the peripheral images surrounding the at least one selected image, the at least one selected image corresponding to the selected feature values, and the peripheral images corresponding to the peripheral feature values; and the host computer generates an air pollution result by comparing the at least one object image with at least one sample image; wherein, a database of the host computer extracts the at least one sample image pre-stored, the at least one sample image corresponding to the hyperspectral image information, and the at least one sample image corresponding to the air pollutant. As described in claim 1, in the method for assessing air pollution using intelligent hyperspectral imagery, in the step where the host performs at least one convolution operation on the second feature values ​​in terms of space and time based on a plurality of convolution kernels, the host sets the plurality of convolution kernels to normalize a plurality of image values ​​of the environmental image to a plurality of pixel normal values, and multiplies the plurality of convolution kernels by the plurality of pixel normal values ​​to extract the second feature values ​​in a convolutional layer. As described in claim 1, in the method for assessing air pollution using intelligent hyperspectral imagery, in the step where the host performs at least one convolution operation on the second feature values ​​spatially and temporally based on a plurality of convolution kernels to obtain at least one object to be measured, the host integrates the regions where the selected feature values ​​are located to obtain at least one distribution area on the environmental image. As described in claim 1, in the method for assessing air pollution using intelligent hyperspectral imaging, in the step of the image capturing unit capturing an environmental image to the host based on an environment, the image capturing unit is a drone, and the environmental image is an aerial photograph of the environment. The method for assessing air pollution using intelligent hyperspectral imaging as described in claim 1, wherein the step of the host comparing the at least one image of the object to be measured with at least one sample image to generate an air pollution result further includes at least the following steps: in the step of the host comparing the at least one image of the object to be measured with at least one sample image to generate an air pollution result, the host performs simplification processing in at least one pooling layer; and the host compares the image of the object to be measured with the at least one sample image in at least one fully connected layer. As described in claim 1, in the step of the host comparing the at least one object image to at least one sample image to generate an air pollution result, when the host cannot determine that the environmental image is the air pollution image based on the at least one sample image, the host performs an approximate comparison of the at least one object image to at least one sample image based on the at least one sample image. The method for assessing air pollution using intelligent hyperspectral imagery as described in claim 1, wherein in the step of the host converting the reference image into a hyperspectral reference image based on the reference image, the host converts the reference image into a hyperspectral reference image through a plurality of color matching functions of a standard color chart, a correction matrix, and a transformation matrix. The method for assessing air pollution using intelligent hyperspectral imaging as described in claim 1, wherein in the step of the host comparing the at least one object image to the at least one copy image to generate an air pollution result, the host compares the aligned at least one object image to the at least one copy image. The method for assessing air pollution using intelligent hyperspectral imagery as described in claim 1 further includes at least the following steps: the host obtains a corresponding pollution level based on the air pollution result and the environmental image; and the host colors the image of at least one object to be tested based on the pollution level. As described in claim 1, in the method for assessing air pollution using intelligent hyperspectral imagery, in the step where the host performs a principal component analysis operation on the first feature values ​​based on a band corresponding to an air pollutant to simplify the hyperspectral image and generate a plurality of corresponding second feature values, a first number of the first feature values ​​is greater than a second number of the second feature values. As described in claim 1, the method for assessing air pollution using intelligent hyperspectral imagery, wherein in the step of the host computer performing a principal component analysis on the first feature values ​​based on a band corresponding to an air pollutant, the air pollutant and its corresponding band are pre-stored in the database, and the air pollutant includes: sulfur dioxide, the corresponding band being the infrared bands with wavelengths of 18.9 µm, 8.8 µm, and 7.6 µm; nitrogen monoxide, the corresponding band being the infrared band with wavelengths of 5.3 µm; nitrogen dioxide, the corresponding band being the infrared bands with wavelengths of 6.2 µm, 7.5 µm, and 13.3 µm; carbon monoxide, the corresponding band being the infrared band with wavelengths of 4.6 µm; and ozone, the corresponding band being the infrared band with wavelengths of 9.3 µm to 10.4 µm and the ultraviolet band with wavelengths of 254 nm.

Images