Near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing

Abstract

The present invention provides a near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, comprising the following steps: S1. Preliminary shadow detection: Preliminary shadow detection is performed on the hyperspectral image to obtain the illuminated area; S2. Shadow compensation is performed using a shadow compensation method based on cubic multilinear unmixing. The method performs a secondary shadow detection after the first SMLM model unmixing, and considers the endmembers in the shadow area to perform a second SMLM model unmixing to determine the degree of compensation; the abundance and VPEI index obtained from the second SMLM model unmixing are used to determine the shadow type, and a third SMLM model unmixing combined with VPEI is performed on areas with minimal information. Different degrees of compensation are applied to different shadows, effectively solving the problems of overcompensation and erroneous compensation in high-resolution imagery caused by the SMLM compensation model.

Description

Technical Field

[0001] The present invention relates to the field of shadow compensation of hyperspectral images, and in particular to a near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing. Background Art

[0002] High-spatial-resolution imagery refines ground object information, allowing for clearer locations, more vivid colors, and more pronounced texture and shape features, presenting promising development prospects. Hyperspectral data, with its wide spectral range and high spectral resolution, conveys richer spectral information related to the physical properties and chemical composition of ground objects compared to RGB and multispectral imagery. However, with increasing image resolution, shadows have become an unavoidable issue in high-resolution hyperspectral image analysis. During remote sensing image acquisition, shadows are created by ground objects blocking light. Partial information about objects located in shadowed areas is lost, making shadows a significant interference factor in automated remote sensing image processing, directly impacting image processing accuracy and, consequently, object detection and recognition. For vegetation, canopy shadows reduce spectral reflectance and alter the shape of the spectral curve due to indirect illumination and multiple sunlight scattering. This can lead to misclassification of shadowed areas and distort parameter inversion results. Vegetation canopy shadows can also obscure visual or spectral changes caused by stress, viral infection, and other factors, distorting interpretation and significantly impacting agricultural and environmental monitoring. Therefore, restoring shadow information from high-resolution hyperspectral vegetation canopy imagery is crucial. Most studies refer to this process of restoring shadowed areas as "shadow compensation."

[0003] Scholars have long studied shadow compensation methods. One type is based on physical models, using satellite parameters, solar altitude, and solar azimuth to simulate solar illumination conditions and employing radiative transfer models to analyze shadows. Friman et al. used LiDAR data as an additional data source based on solar altitude information, providing geometric information and employing precise digital surface models for restoration. However, the compensation effect in the bands beyond 700 nm was poor. Oduncu et al. used LiDAR and hyperspectral data within a physical radiation model for feature correction and applied machine learning to perform shadow compensation. Another type relies on image information itself, leveraging spectral and texture features to explore the relationship between illumination and shadow reflectance. These include traditional shadow compensation methods such as LCC and LSCR. For hyperspectral images, many researchers approach them as a spectral unmixing problem. Heylen et al. incorporated a shadow fraction into the MLM unmixing model and proposed the shadow-based multilinear mixture (SMLM) model, reconstructing the image by varying the shadow fraction Q. Zhang et al. proposed the ESMLM model, which incorporates neighboring light sources into the SMLM model. Such methods are mostly used in urban remote sensing to remove shadows of buildings. For hyperspectral images of vegetation and soil with higher spatial resolution, the shadows formed by vegetation canopies are more complex, and the applicability of these algorithms to vegetation canopies remains to be verified.

[0004] At present, there are still the following problems in the research on shadow compensation: (1) There is a lack of research on near-ground hyperspectral images; (2) There is little research on vegetation canopy shadows, and the building shadow compensation method will have the problem of overcompensation when directly applied to vegetation canopies; (3) Some compensation models will show incorrect compensation in areas with severe spectral information loss and fail to remove all shadows. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing to solve the problems of over-compensation and compensation errors in the vegetation canopy area of ​​the current traditional shadow compensation method.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] The near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing of the present invention includes the following steps: S1. preliminary shadow detection: performing preliminary shadow detection on the hyperspectral image to obtain the illuminated area; S2. performing shadow compensation using a shadow compensation method based on cubic multilinear unmixing.

[0008] Optionally, in the above-mentioned near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, in step S2, it includes: a. illumination area end member selection: extracting illumination area end members of vegetation and soil in the illumination area; b. shadow secondary detection and shadow area end member selection: using the illumination area end member as the input end member of the shadow-based multilinear mixture (SMLM) model to extract the shadow area, and selecting the shadow area end member using the Euclidean distance method; c. shadow compensation: inputting the shadow area end member into the shadow-based multilinear mixture (SMLM) model, and dividing the shadow into ordinary shadow area and shadow area with too little information, the ordinary shadow area is spectrally reconstructed and compensated according to the unmixed parameters, and the shadow area with too little information is unmixed based on the shadow multilinear mixture (SMLM) model in combination with VPEI and then spectrally reconstructed and compensated, wherein VPEI is the vegetation pixel extraction index considering shadows.

[0009] Optionally, in the above-mentioned near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, in a, in the initially detected illumination area, the vertex component analysis algorithm is used to extract the end members of vegetation and soil, which are regarded as the illumination area end members.

[0010] Optionally, in the above-mentioned near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, in step b, the end member of the illumination area is input into the shadow-based multilinear mixing (SMLM) model for the first shadow-based multilinear mixing model unmixing. The mixing equation is as follows and follows the least squares principle:

[0011]

[0012] Where x is the reflectivity vector of the pixel, e i 、a i are endmembers and endmember abundances, respectively. Q is the shadow score of the shadow-based multilinear mixed model. The number of endmembers is p, a i ≥0, 0≤P≤1,0≤Q≤1,

[0013] The shadow score Q after the first unmixing of the multilinear mixture model with the shadow is threshold segmented and the shadow secondary detection is performed. After the image is binarized, the morphological closing operation is used to fill the gaps in the image without changing the outline of the binary image. The process is that the structural element B first expands and then corrodes A. The formula is as follows:

[0014]

[0015] The Euclidean distance is used to measure the similarity between the two spectral curves. The pixel closest to the endmember of the initial detection illumination area is found in the shadow area of ​​the secondary detection, that is, the pixel with the smallest Euclidean distance, and it is regarded as the endmember of the shadow area. The formula of the Euclidean distance d is:

[0016]

[0017] Where x=[x1,x2,…,x n ] T ,y=[y1,y2,…,x n ] T , x and y represent the spectral vectors of two different pixels.

[0018] Optionally, in the above-mentioned near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, in c, the shadow area is subjected to a second shadow-based multilinear mixing model unmixing, and based on the result of the second shadow-based multilinear mixing model unmixing, a judgment is added to divide the shadow area of ​​the image into an ordinary shadow area and a shadow area with very little information and perform shadow compensation, including: 1) judging the type of vegetation and soil, assuming that the vegetation end member and abundance are e1 and a1, and the soil end member and abundance are e2 and a2, then the judgment basis is: a1 ≥ 0.8 and VPEI ≥ 0.3 or a1 ≤ 0.2 and VPEI <0.3, when its VPEI ≥ 0.3, it is considered a vegetation pixel, otherwise it is a soil pixel; 2) For pixels that do not meet the conditions in 1), they are regarded as "shadow areas with very little information" and their types are determined by combining the VPEI index. Vegetation pixels and soil pixels respectively meet the special case of p = 1 of the shadow-based multilinear mixed model. The third shadow-based multilinear mixed model unmixing is performed, and the original restriction condition Q∈[0,1] of the shadow-based multilinear mixed model combined with VPEI is changed to: Q∈[0.7,1]. The shadow area with very little information should meet the following formula:

[0019]

[0020] Where x is the reflectance vector of the pixel, e1 and e2 are the vegetation and soil end members in the shadow area, respectively, 0≤P≤1, 0.7≤Q≤1.

[0021] Optionally, in the above-mentioned near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, in step S1, two vegetation indices, namely the natural gas stress index (NGSI) and the converted chlorophyll absorption reflectance index (TCARI), and the threshold segmentation method are used to perform preliminary shadow detection on the hyperspectral image to obtain the illumination area, which includes the vegetation illumination area and the soil illumination area; and the Otsu method is used for threshold segmentation to define the thresholds of the vegetation shadow area and the vegetation illumination area, and the empirical threshold method is used to define the thresholds of the soil shadow area and the soil illumination area.

[0022] The beneficial effects of the present invention are:

[0023] The present invention's near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, called the triple-SMLM shadow compensation method, effectively solves the problems of over-compensation and erroneous compensation of the SMLM compensation model in high-resolution images. The shadow compensation performance of the method of the present invention is greatly improved, with obvious improvements on the vegetation shadow edges and soil shadow parts with less information, making it more suitable for the vegetation canopy area of ​​near-ground high-spatial-resolution hyperspectral images. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific implementation or the description of the prior art.

[0025] Figure 1 is a flow chart of a near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing according to the present invention;

[0026] Figure 2 It is a detailed flowchart of the near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing according to the present invention;

[0027] Figure 3 This is a comparison chart of shadow compensation results of other methods and the method of the present invention, taking SOC710-VP hyperspectral data as an example;

[0028] Figure 4 The shadow compensation results of other methods and the method of the present invention are shown, as well as the vegetation spectrum after compensation, taking the SOC710-VP hyperspectral data as an example. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0030] like Figure 1 and Figure 2 As shown, the near-ground hyperspectral shadow compensation method based on cubic multilinear unmixing of the present invention includes the following steps:

[0031] S1. Preliminary shadow detection

[0032] In this step, two vegetation indices, the natural gas index (NGSI) and the transformed chlorophyll absorption in reflectance index (TCARI), and the threshold segmentation method are used to perform preliminary shadow detection on the hyperspectral image to obtain the illumination area, which includes the vegetation illumination area and the soil illumination area.

[0033] Specifically, the Natural Gas Index (NGSI) and the Transformed Chlorophyll Absorption in Reflectance Index (TCARI) are used to extract illuminated vegetation and illuminated soil from the hyperspectral reflectance data of hyperspectral imagery. The NGSI is more effective at extracting illuminated vegetation, while the TCARI is more effective at distinguishing illuminated soil. The calculation formulas for the two indices are as follows:

[0034]

[0035]

[0036] The Otsu method was used for threshold segmentation to define the thresholds of vegetation shadow area and vegetation illumination area, and the empirical threshold method was used to define the thresholds of soil shadow area and soil illumination area.

[0037] S2.triple-SMLM shadow compensation: Shadow compensation is performed using a shadow compensation method based on cubic multilinear unmixing. The process is divided into the following three steps:

[0038] a. Illumination zone end-member selection: Extract the illumination zone end-members of vegetation and soil in the illumination area.

[0039] Specifically, endmember extraction includes extracting endmembers of vegetation and soil in the initially detected illumination area using a vertex component analysis (VCA) algorithm, which are regarded as illumination area endmembers.

[0040] b. Secondary detection of shadows and selection of end elements in shadow areas

[0041] The endmembers of the illumination area are used as the input endmembers of the SMLM mixture model to extract the shadow area (the first multilinear unmixing), and the shadow area endmembers are selected using the Euclidean distance method. Specifically, the endmembers extracted from the illumination area ( Figure 2 The endmembers in the illuminated region (i.e., the endmembers in the illuminated region) are input into the SMLM mixture model for the first SMLM model unmixing. The mixing equation of the SMLM model is as follows and follows the least squares principle:

[0042]

[0043] Where x is the reflectivity vector of the pixel, e i 、a i are endmembers and endmember abundances respectively, Q is the shadow score of the SMLM model, the number of endmembers is p, a i ≥0, 0≤P≤1, 0≤Q≤1.

[0044] The shadow score Q after the first SMLM model demixing is segmented using Otsu thresholding for secondary shadow detection. After binarization, the image is then closed using morphological closing to fill gaps in the image without changing the outline of the binarized image. This process involves dilating and then eroding the structure element A using the following formula:

[0045]

[0046] Compared with the illuminated area, the shadow area of ​​the same type of ground objects has lower components in each band. Based on this characteristic, the Euclidean distance is used to measure the similarity between the two spectral curves and used as the standard for selecting the endmembers of the shadow area. In the shadow area of ​​the secondary detection, the pixel closest to the endmember of the illuminated area of ​​the initial detection is found, that is, the pixel with the smallest Euclidean distance, and it is regarded as the endmember of the shadow area. The formula of the Euclidean distance d is:

[0047]

[0048] Where, pixel x=[x1,x2,…,x n ] T ,y=[y1,y2,…,x n ] T , x and y represent the spectral vectors of two different pixels.

[0049] c. Shadow compensation: Input the shadow area end member into the SMLM mixed model (second SMLM model unmixing), and use the abundance and VPEI index to divide the shadow into two parts, namely the ordinary shadow area and the shadow area with too little information, and perform shadow compensation. The ordinary shadow area is spectrally reconstructed and compensated according to the unmixed parameters, and the shadow area with too little information is unmixed with the SMLM model combined with the VPEI index and then spectrally reconstructed and compensated (third SMLM model unmixing).

[0050] The shadow area is subjected to a second SMLM model unmixing, i.e., the second unmixing. Based on the results of the second unmixing, judgment is added to divide the shadow area of ​​the image into ordinary shadow area and shadow area with very little information, and shadow compensation is performed. It is divided into the following two parts:

[0051] 1) Vegetation and soil type determination: Let e1 and a1 be the vegetation endmember and its abundance, and e2 and a2 be the soil endmember and its abundance. The determination criteria are: a1 ≥ 0.8 and VPEI ≥ 0.3, or a1 ≤ 0.2 and VPEI < 0.3. Since the data has a spatial resolution of 2.8 mm, there are very few mixed pixels in the imagery. Therefore, when the abundance unmixing clearly indicates that a pixel is leaning towards a certain type, the unmixing result is considered correct and has a clear classification within the data. The VPEI index is a vegetation extraction index that accounts for shadows and effectively addresses the misclassification of shaded soil from illuminated vegetation. According to the bimodal histogram method, the optimal threshold for distinguishing vegetation from soil is 0.3. When the VPEI ≥ 0.3, the pixel is considered vegetation; otherwise, it is soil. Incorporating the VPEI into the determination prevents SMLM from misidentifying pixel types. For these pixels, shadow compensation is performed using SMLM.

[0052]

[0053] 2) For pixels that do not meet the conditions in 1), they are regarded as "shadow areas with very little information" and their types are determined by combining the VPEI index. Vegetation pixels and soil pixels respectively meet the special case of p = 1 in the SMLM model, and the third SMLM model unmixing is performed, that is, the third unmixing.

[0054] The present invention believes that, except for a small number of mixed pixels at the junction of vegetation and soil, the remaining pixels do not meet the restriction conditions of SMLM spectral unmixing, and the reason why the type cannot be correctly identified is that the lighting conditions are too poor and the shadow area loses a lot of spectral and texture information. Therefore, the present invention changes the original restriction condition Q∈[0,1] of the shadow score Q of the SMLM model that takes into account the vegetation pixel extraction index (VPEI) of the shadow to: Q∈[0.7,1], indicating that the shadow degree of the pixels in this part is very heavy, thereby limiting its unmixing conditions. This method can effectively improve the limitations of the SMLM model caused by unnatural compensation due to insufficient spectral information. Therefore, the shadow area with very little information should satisfy the following formula:

[0055]

[0056] Where x is the reflectance vector of the pixel, e1 and e2 are the vegetation and soil end members in the shadow area, respectively, 0≤P≤1, 0.7≤Q≤1.

[0057] Replacing the illumination source in the shadowed area with that in the sunlit area can "illuminate" the shadowed portion of the pixel. For the SMLM model, Q = 0 can be set to simulate direct illumination in the shadowed area. Spectra of the two shadowed areas are reconstructed based on the unmixed abundance and the shadowed endmembers. The reconstruction conforms to the mixing equation, but the shadow fraction Q = 0 is set to achieve the effect of shadow compensation, that is, the compensated image.

[0058] The present invention improves the SMLM shadow compensation model and proposes a near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, called the triple-SMLM shadow compensation method. After the first SMLM model unmixing, the shadow is secondary detected, and the shadow area end members are taken into account, and the second SMLM model unmixing is performed to determine the compensation degree. This process solves the over-compensation problem; the abundance and VPEI index obtained by the second SMLM model unmixing are used to determine the shadow type, and a third SMLM model unmixing combined with VPEI is performed on areas with very little information. Different shadows are compensated to different degrees. This process solves the problem of erroneous compensation; the present invention provides an effective method for shadow compensation in near-ground hyperspectral vegetation canopy areas, solving the problem that the prior art lacks a shadow compensation method for hyperspectral vegetation canopy areas.

[0059] Specific Example The method of the present invention is described using SOC710-VP hyperspectral data as research data.

[0060] The data plot was located in farmland in Luozhuang Village 2, Changziying Town, Daxing District, Beijing (39°39′2.56″N, 116°34′33.10″E). The experimental area consisted of 24 square plots, each 2.5 m in length. The climate in this area is warm temperate, semi-humid, and monsoon, with distinct seasons, an average annual temperature of 11.6°C, and an average annual precipitation of 556 mm. This makes the climate ideal for soybean growth. Soybeans were sown on June 14, 2017.

[0061] Data were collected using a SOC710-VP imaging spectrometer between 10:00 AM and 2:00 PM local time under good weather conditions. The instrument has a spectral resolution of approximately 4.69 nm, a wavelength range of 370–1045 nm, and 128 bands. The hyperspectral camera was raised to a height of 5 m above the canopy using a mobile lift. The images achieved a spatial resolution of 2.8 mm × 2.8 mm. The data exhibited high spatial and spectral resolution. This experiment used soybean plots from July 30, 2017. The data underwent preprocessing, including radiometric correction, rotation, and cropping.

[0062] The near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing of the present invention is mainly divided into the following two steps:

[0063] S1. Preliminary shadow detection

[0064] The Natural Gas Index (NGSI) and the Transformed Chlorophyll Absorption in Reflectance Index (TCARI) were used to spectrally identify illuminated vegetation and illuminated soil. The NGSI and TCARI were more effective in distinguishing illuminated vegetation, while the TCARI was more effective in distinguishing illuminated soil. The calculation formulas for the two indices are as follows:

[0065]

[0066]

[0067] The Otsu threshold segmentation method was used to define the thresholds of vegetation shadow area and light area, and the empirical threshold method was used to define the thresholds of soil shadow area and light area.

[0068] S2.triple-SMLM shadow compensation

[0069] The triple-SMLM shadow compensation process is divided into three steps:

[0070] a. Selection of end members in the illumination area

[0071] Endmember extraction: In the initially detected illumination area, the vertex component analysis (VCA) algorithm is used to extract the endmembers of vegetation and soil, which are regarded as the illumination area endmembers.

[0072] b. Secondary detection of shadows and selection of end elements in shadow areas

[0073] The end members extracted from the illumination area are input into the SMLM model. The hybrid equation of the SMLM model is as follows, and it follows the least squares principle:

[0074]

[0075] Where x is the reflectivity vector of the pixel, e i 、a i are endmembers and endmember abundance respectively, the number of endmembers is p, a i ≥0, 0≤P≤1, 0≤Q≤1.

[0076] The shadow score Q after the first SMLM model demixing is segmented using Otsu thresholding for secondary shadow detection. After binarization, the image is then closed using morphological closing to fill gaps in the image without changing the outline of the binarized image. This process involves dilating and then eroding the structure element A using the following formula:

[0077]

[0078] Compared to illuminated areas, shadow areas of similar features have lower components in each band. Based on this characteristic, the Euclidean distance is used to measure the similarity between the two spectral curves and serves as the criterion for selecting endmembers in the shadow area. Within the shadow area of ​​the secondary detection, the pixel closest to the endmember of the illuminated area detected initially is found, that is, the pixel with the smallest Euclidean distance is considered the shadow endmember.

[0079]

[0080] Where x=[x1,x ２ ,…,x n ] T ,y=[y1,y2,…,x n ] T , x and y represent the spectral vectors of two different pixels.

[0081] c.Shadow compensation

[0082] Perform a second SMLM model unmixing on the shadow area. Based on the results of the second unmixing, judgment is added to divide the shadow area of ​​the image into ordinary shadow area and shadow area with very little information for shadow compensation. It is divided into the following two parts:

[0083] (1) Determine the type of vegetation and soil. Let the vegetation end member and abundance be e1 and a1, and the soil end member and abundance be e2 and a2. The judgment criteria are: a1 ≥ 0.8 and VPEI ≥ 0.3 or a1 ≤ 0.2 and VPEI < 0.3. Since the spatial resolution of the data is 2.8 mm, there are only a few mixed pixels in the image. Therefore, when the abundance unmixing can clearly determine that the pixel tends to a certain type, we believe that the unmixing result is correct and has a clear type meaning in the data. The VPEI index is a vegetation extraction index that takes shadows into account and can effectively solve the problem of misclassification of shadowed soil and illuminated vegetation. According to the bimodal histogram method, the optimal threshold for distinguishing vegetation and soil is 0.3. When its VPEI ≥ 0.3, it is considered a vegetation pixel, otherwise it is a soil pixel. Including VPEI in the judgment prevents SMLM from misidentifying the pixel type. For these pixels, shadow compensation under SMLM is performed.

[0084]

[0085] (2) For pixels that do not meet the conditions in (1), they are regarded as “shadow areas with very little information” and their types are determined by combining the VPEI index. Vegetation pixels and soil pixels meet the special case of SMLM model p = 1, and the third SMLM model unmixing is performed.

[0086] The present invention believes that, except for a small number of mixed pixels at the junction of vegetation and soil, the remaining pixels do not meet the restriction conditions of SMLM spectral unmixing, and the reason why the type cannot be correctly identified is that the lighting conditions are too poor and the shadow area loses a lot of spectral and texture information. Therefore, the present invention changes the original restriction condition Q∈[0,1] of the shadow score Q of the SMLM model combined with VPEI to: Q∈[0.7,1], indicating that the shadow degree of this part of the pixels is very heavy, thereby limiting its unmixing conditions. This method can effectively improve the limitations of the SMLM model caused by unnatural compensation due to insufficient spectral information. Therefore, the shadow area with very little information should satisfy the following formula:

[0087]

[0088] Where x is the reflectance vector of the pixel, e1 and e2 are the vegetation and soil end members in the shadow area, respectively, 0≤P≤1, 0.7≤Q≤1.

[0089] Replacing the illumination source in the shadowed area with that in the sunlit area can "illuminate" the shadowed portion of the pixel. For the SMLM model, Q = 0 can be set to simulate direct illumination in the shadowed area. Spectra of the two shadowed areas are reconstructed based on the unmixed abundance and the shadowed endmembers. The reconstruction conforms to the mixing equation, but the shadow fraction Q = 0 is set to achieve the effect of shadow compensation, resulting in a compensated image.

[0090] In order to compare the accuracy of the method of the present invention, the SMLM method and the triple-SMLM method of the present invention were compared on the same data. Figure 3 shown. Figure 3 (a), (b), (c), and (d) are the original image, the image compensated by the LCC method, the image compensated by the SMLM method, and the image compensated by the triple-SMLM method, respectively. Overall, the SMLM result is brighter than the original image, and even compensates for the illuminated areas that do not need compensation. LCC compensates heavily shaded soil with darker vegetation, and SMLM also compensates for some soil with vegetation, but some soil is not fully compensated and still has shadows. Triple-SMLM's compensation of the soil area has a better visual effect. Figure 4Figure 2 shows the shadow compensation results and the spectrum of the compensated vegetation. SV represents illuminated vegetation, BV represents shadow edge vegetation, and SHV represents shadow vegetation. In the LCC compensation results, the spectrum of the vegetation's shadow edge is higher than that of the illuminated area, indicating overcompensation at the shadow edge, which leads to a visually sharpened edge in the compensated area. The triple-SMLM vegetation compensation results are more natural, with a good transition effect even at the shadow edge. The spectral curves of the triple-SMLM's illuminated area, compensated shadow edge, and compensated shadow area are very similar, indicating a good compensation effect.

[0091] In order to further evaluate the calculation results of the invented algorithm, the algorithm accuracy evaluation is emphasized:

[0092] In order to evaluate the performance of shadow compensation, the present invention uses the changes in the average spectrum of the ground object, the mean square error (STD), and the root mean square error (RMSE) to observe the changes before and after shadow compensation. The formulas for the mean square error (STD) and the root mean square error (RMSE) are:

[0093]

[0094] Where, The average value of the image in the jth band.

[0095]

[0096] Where x ref 、x shw is the reflectivity of pixels in the original and shadow-compensated target image, and N is the number of pixels in the target image. In actual calculations, due to the lack of reflectivity data for the same pixel under both illuminated and shadowed conditions, the root mean square error (RMSE) between the original shadowed area of ​​the shadow-compensated result and the adjacent non-shadowed area of ​​the original image containing similar features is calculated. For the original image, the RMSE is a function of the reflectivity of the illuminated and shadowed pixels in the original image.

[0097] Table 1 Quantitative accuracy evaluation of STD and RMSE of original and compensated images

[0098]

[0099] The STD reflects the degree of dispersion. The smaller the STD value for similar objects, the smaller the spectral differences. The root mean square error (RMSE) indicates shadow compensation capability; smaller values ​​indicate stronger shadow compensation capabilities. Table 1 shows the STD and RMSE of the original and compensated images. The triple-SMLM method achieved an STD < 0.6 and an RMSE < 6%. Both the STD and RMSE were lower than those of the original image, and were the lowest among the shadow compensation methods, demonstrating its best shadow compensation capability. The SMLM method performed poorly, indicating that it is not suitable for shadow compensation near the ground in the vegetation canopy. The triple-SMLM method showed significant improvements in both vegetation and soil accuracy: the STD for vegetation and soil decreased from 0.97 and 1.84 to 0.51 and 0.44, respectively, a decrease of 0.47 and 1.40; the RMSE decreased from 13.96% and 17.80% to 5.21% and 4.64%, respectively, a decrease of 8.75% and 13.16%. In addition, LCC also shows good shadow compensation ability, but the STD of the soil is higher than that of the original image, indicating that the compensation for the soil is poor.

[0100] In summary, it is demonstrated from both qualitative and quantitative perspectives that the near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing of the present invention has high accuracy and can solve the problems of over-compensation and erroneous compensation.

[0101] The above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can modify or improve the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing, characterized in that: The steps include: S1. Preliminary shadow detection: Perform preliminary shadow detection on the hyperspectral image to obtain the illuminated area; S2. Perform shadow compensation using a shadow compensation method based on cubic multilinear unmixing; In step S2, it includes: a. Illumination zone end member selection: In the initial detection of the illumination area, the end members of vegetation and soil are extracted using the vertex component analysis algorithm, which is regarded as the illumination zone end member; b. Shadow secondary detection and shadow area endmember selection: The illuminated area endmember is used as the input endmember of the shadow-based multilinear mixture (SMLM) model to extract the shadow area, and the shadow area endmember is selected using the Euclidean distance method; c. Shadow compensation: The shadow area endmembers are input into the shadow-based multilinear mixture (SMLM) model, and the shadows are divided into ordinary shadow areas and shadow areas with too little information. The ordinary shadow areas are spectrally reconstructed and compensated based on the unmixed parameters. The shadow areas with too little information are unmixed with the shadow-based multilinear mixture (SMLM) model in combination with VPEI and then spectrally reconstructed and compensated, where VPEI is the vegetation pixel extraction index that takes shadows into account.

2. The near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing according to claim 1 is characterized in that: In step b, the illuminated area end member is input into the shadow-based multilinear mixing model to perform the first shadow-based multilinear mixing model unmixing. The mixing equation is as follows and follows the least squares principle: Where, is the reflectivity vector of the pixel, are endmembers and endmember abundances respectively, Q is the shadow score of the shadow-based multilinear mixed model, and the number of endmembers is ; The shadow score after the first unmixing of the shadow-based multilinear mixing model Perform threshold segmentation and shadow secondary detection. After binarizing the image, use morphological closing operation to fill the gaps in the image without changing the outline of the binary image. The process is that the structural element B first expands and then corrodes A. The formula is as follows: The Euclidean distance is used to measure the similarity between the two spectral curves. The pixel closest to the endmember of the initial detection illumination area is found in the shadow area of ​​the secondary detection, that is, the pixel with the smallest Euclidean distance, and it is regarded as the endmember of the shadow area. The formula of the Euclidean distance d is: Where, , x and y represent the spectral vectors of two different pixels.

3. The near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing according to claim 1 is characterized in that: In step c, the shadow area is subjected to a second unmixing process based on the shadow-based multilinear hybrid model. Based on the result of the second unmixing process based on the shadow-based multilinear hybrid model, a judgment is added to divide the shadow area of ​​the image into a normal shadow area and a shadow area with very little information, and shadow compensation is performed, including: 1) Determine the type of vegetation and soil, assuming that the vegetation end members and abundance are and , soil end members and their abundance are and , the judgment basis is: , when When , it is considered as a vegetation pixel, otherwise it is a soil pixel; 2) For pixels that do not meet the conditions in 1), they are regarded as "shadow areas with very little information" and their types are determined by combining the VPEI index. Vegetation pixels and soil pixels meet the shadow-based multilinear mixed model respectively. In the special case of , the third shadow-based multilinear mixed model unmixing is performed, and the shadow score of the shadow-based multilinear mixed model combined with VPEI is The original restrictions Change to , the shadow area with very little information should satisfy the following formula: Where, is the reflectivity vector of the pixel, are the vegetation and soil end members in the shadow area, .

4. The near-ground hyperspectral vegetation shadow compensation method based on cubic multilinear unmixing according to claim 1 is characterized in that: In step S1, two vegetation indices, the natural gas stress index (NGSI) and the transformed chlorophyll absorption reflectance index (TCARI), are used in conjunction with the threshold segmentation method to perform preliminary shadow detection on the hyperspectral image to obtain the illuminated area, which includes the vegetation illuminated area and the soil illuminated area. The threshold segmentation is then performed using the Otsu method to define the thresholds of the vegetation shadow area and the vegetation illuminated area, and the empirical threshold method is used to define the thresholds of the soil shadow area and the soil illuminated area.

Images