A clustering-hybrid based method for soil property monitoring using unmanned aerial vehicle hyperspectral images

Abstract

The present invention discloses a method for monitoring soil properties of unmanned aerial vehicles (UAVs) based on clustering-mixing. The method clusters the spectra of UAV hyperspectral images, and then divides the UAV images into type regions with different spectral characteristics. Then, a dedicated soil property prediction model is constructed for each type region. Then, a high-precision soil property model is used to realize the spatial prediction of soil properties in the corresponding spectral type region, and a spatial interpolation method is used to realize the spatial prediction of soil properties in the low-precision model region. Finally, the spatial distribution results predicted by the high-precision model are inlaid with the spatial interpolation results of the low-precision soil property prediction region to obtain a complete spatial distribution map of soil properties in the region. The present invention can effectively reduce the interference of factors such as field soil moisture and roughness during the process of UAV hyperspectral soil monitoring, improve the accuracy of soil property monitoring, and enhance the level of UAV hyperspectral soil property monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of soil property acquisition, and in particular to a clustering-mixing-based unmanned aerial vehicle hyperspectral soil property monitoring method. Background Art

[0002] Traditional laboratory-based chemical analysis methods for soil properties usually have high analytical accuracy, but the analysis cost is high and the analysis process is relatively complicated, which makes it difficult to meet the needs of fine management of soil resources under the conditions of modern agricultural development. Remote sensing technology can quickly obtain field soil property information, but satellite remote sensing technology usually has low spatial resolution. UAV remote sensing technology has high spatial resolution and flexibility, which can effectively meet the requirements of rapid monitoring of field soil properties.

[0003] The current application of soil monitoring based on drone hyperspectral technology is mainly aimed at areas with less influence from factors such as field soil moisture, soil roughness, and crop straw coverage. However, field soil is usually inevitably affected by moisture, and when the field is plowed, it will also lead to greater soil surface roughness, as well as residual straw in the field or the impact of straw returning to the field. Considering the fine resolution of drone hyperspectral images, the degree of influence of these factors in different pixels of the image varies. Directly constructing a model of soil properties is easily affected by related factors, which may reduce the accuracy of the model. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a clustering-mixing UAV hyperspectral soil property monitoring method, which can effectively reduce the interference of factors such as field soil moisture and roughness during UAV hyperspectral soil monitoring, improve the accuracy of soil property monitoring, and enhance the level of UAV hyperspectral soil property monitoring.

[0005] Technical solution: The method for monitoring soil properties using a drone with high spectral spectrum of the present invention comprises the following steps:

[0006] S1, pre-process the acquired UAV hyperspectral image to generate a ground reflectance image; then obtain a hyperspectral reflectance image covering the entire study area through image stitching processing;

[0007] S2, spectral clustering is performed on the acquired UAV hyperspectral reflectance image, and the entire image is divided into different areas according to the clustering results;

[0008] S3, constructing a dedicated soil property prediction model for each spectral clustering area, and dividing it into a high-precision soil property prediction model and a low-precision soil property prediction model according to the inversion accuracy;

[0009] S4, for areas belonging to the high-precision soil property prediction model, spatial prediction of soil properties is performed. For areas belonging to the low-precision soil property prediction model, the spatial interpolation method is used to obtain the spatial distribution of soil properties. The spatial distribution results predicted by the high-precision model are inlaid with the spatial interpolation results of the low-precision soil property prediction area to obtain a complete regional soil property spatial distribution map.

[0010] Furthermore, in step S1, the steps for obtaining a hyperspectral image covering the entire study area are as follows:

[0011] S11, according to the scope of the study area, set the lateral overlap ratio, plan and design the UAV flight route, and obtain N UAV hyperspectral images, where N ≥ 1;

[0012] S12, performing radiometric calibration and atmospheric correction processing on N drone hyperspectral images to generate ground reflectance images;

[0013] S13, performing image stitching processing on N ground reflectance images to obtain a hyperspectral reflectance image covering the entire study area.

[0014] Further, in step S2, according to the clustering result, the entire image is divided into different regions to achieve the following steps:

[0015] S21, for the obtained UAV hyperspectral reflectance image, a spectral clustering model is constructed using a Gaussian mixture model, and the number of clusters is recorded as K;

[0016] S22, using the constructed spectral clustering model, the entire UAV hyperspectral image is divided into regions, and a total of K regions with obvious spectral differences are planned.

[0017] Furthermore, in step S3, the implementation steps of constructing a dedicated soil property prediction model for each spectral clustering region are as follows:

[0018] S31, for each of the K spectral clustering regions, a machine learning algorithm was used, with soil properties as the dependent variable and the spectra at the soil sample collection points in the UAV hyperspectral reflectance image as the independent variable as input to construct a soil property prediction model, and the cross-validation or independent validation method was used to calculate the determination coefficient R of the soil property prediction model. 2 and root mean square error RMSE;

[0019] S32, based on K spectral clustering regions R 2 The value, R 2 The model with a value ≥ 0.60 is a high-precision soil property prediction model, and the corresponding area belongs to the high-precision soil property prediction area, which is recorded as L; R 2The model with a value of <0.60 is a low-precision soil property prediction model, and the corresponding area belongs to the low-precision soil property prediction area, which is recorded as M, where M≥0 and L+M=K.

[0020] Further, in step S4, the steps for obtaining a complete spatial distribution map of soil properties in the region are as follows:

[0021] S41, using K high-precision soil property prediction models to perform soil property spatial prediction for the corresponding spectral clustering areas;

[0022] If M>0, go to step S42;

[0023] S42, for the M low-precision soil property prediction areas, using the prediction results of the high-precision soil property prediction model, adopting the ordinary Kruger spatial interpolation method to obtain the soil property spatial prediction results, and then proceeding to step S43;

[0024] S43, mosaicking the spatial distribution results predicted by the high-precision soil property prediction model and the spatial interpolation results of the low-precision soil property prediction area to obtain a complete regional soil property spatial distribution map.

[0025] Compared with the prior art, the present invention has the following significant effects:

[0026] 1. The present invention performs spectral clustering on the entire study area and divides it into type areas with different spectral characteristics, among which the hyperspectral area of ​​drone images is less affected by factors such as soil moisture, roughness and straw coverage, and realizes spatial distribution prediction by constructing a high-precision soil property prediction model;

[0027] 2. For the low-precision soil property prediction area, the present invention utilizes the prediction results of the high-precision soil property model, and combines it with the spatial interpolation method to realize spatial distribution prediction; finally, the spatial distribution results predicted by the high-precision model are inlaid with the spatial interpolation results of the low-precision soil property prediction area to obtain a complete regional soil property spatial distribution map, which can effectively reduce the interference of factors such as field soil moisture and roughness during drone hyperspectral soil monitoring, improve the accuracy of soil property monitoring, and enhance the level of drone hyperspectral soil property monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is the overall flow chart of the present invention;

[0029] Figure 2 The study area and the spatial distribution map of soil collection points in the embodiment;

[0030] Figure 3 Schematic diagram of the UAV hyperspectral image in the embodiment;

[0031] Figure 4 is a schematic diagram of the spectrum clustering result in the embodiment;

[0032] Figure 5 is a schematic diagram of the region division result in the embodiment;

[0033] Figure 6 It is a scatter plot of soil organic matter prediction results in the embodiment;

[0034] Figure 7 4 is a spatial distribution diagram of soil organic matter in the embodiment. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0036] In the acquired drone hyperspectral images of the field, thanks to the fine spatial resolution, there are pixels with a high probability that are less or almost unaffected by factors such as field soil moisture, soil roughness or crop straw coverage. The soil property prediction model constructed using the soil spectra of these pixels has a higher accuracy. For pixels that are more affected, it is difficult to build a high-precision model, but the prediction results of the model built based on the spectra of the less affected pixels can be used to achieve spatial prediction using spatial interpolation methods, and then through mosaic processing, the spatial prediction of soil properties covering the entire study area can be achieved. Therefore, the drone hyperspectral images can first be spectrally clustered, and models can be constructed for each type. The models with higher accuracy can directly predict soil properties, and the low-precision model areas can be spatially interpolated and then mosaicked to obtain the high-precision spatial distribution of regional soil properties. Therefore, the present invention provides a drone hyperspectral soil property monitoring method based on a clustering-hybrid approach, which is used to obtain regional high-spatial resolution soil property information, such as Figure 1 As shown, the specific steps include:

[0037] In step A, the UAV hyperspectral image in the study area is obtained, and the obtained UAV hyperspectral image is preprocessed to generate a ground reflectance image, and then the image is stitched to obtain a hyperspectral reflectance image covering the entire study area, and then proceed to step B.

[0038] In practical application, the above step A specifically performs the following steps A1 to A3:

[0039] Step A1: according to the scope of the study area, set the lateral overlap ratio, plan and design the UAV flight route, and obtain N UAV hyperspectral images, where N ≥ 1; if the UAV hyperspectral reflectance image is obtained from other channels, directly go to step B;

[0040] Step A2, for N UAV hyperspectral images, perform radiation calibration and atmospheric correction processing to generate ground reflectance images, and then proceed to step A3;

[0041] Step A3: For N drone hyperspectral ground reflectance images, image stitching processing is performed to delete the first p and last q bands with large noise in the hyperspectral image. p and q are determined according to the quality of the drone hyperspectral image. A hyperspectral reflectance image covering the entire study area is obtained, and then step B is entered.

[0042] Step B, spectral clustering is performed on the obtained UAV hyperspectral reflectance image, and the entire image is divided into different areas according to the clustering results, and then step C is entered;

[0043] In practical application, the above step B specifically performs the following steps B1 to B3:

[0044] Step B1, for the obtained UAV hyperspectral reflectance image, select the band with large spectral reflectance difference, use Gaussian mixture model GMM to build a spectral clustering model, record the number of clusters as K, and then enter step B2;

[0045] Step B2, using the constructed spectral clustering model, the entire UAV hyperspectral reflectance image is divided into regions, and a total of K regions with obvious spectral differences are planned, and then entering step C.

[0046] Step C, constructing a dedicated soil property prediction model for each spectral clustering region, removing the model with low inversion accuracy according to the inversion accuracy, and then proceeding to step D;

[0047] In practical application, the above step C specifically performs the following steps C1 to C2:

[0048] Step C1: For each of the K spectral clustering regions, machine learning algorithms such as partial least squares regression, random forest, support vector machine, and deep neural network (DNN) can be used. The soil properties are used as dependent variables, and the spectra at the soil sample collection points in the UAV hyperspectral reflectance image are used as independent variables as input to construct a soil property prediction model. The cross-validation or independent validation method is used to calculate the determination coefficient R of the soil property prediction model. 2 (Coefficient of Determination,R 2 ) and Root Mean Square Error RMSE (RMSE), and then proceed to step C2;

[0049] Step C2: Based on the soil property prediction model determination coefficient R in K spectral clustering areas2 The value is expressed in R 2 The value ≥0.60 is used as the standard to classify it into a high-precision soil property prediction model, which is recorded as L. The spectral clustering area where the high-precision soil property prediction model is located is defined as the high-precision soil property prediction area; R 2 The models with a value of <0.60 are low-precision soil property prediction models, which are recorded as M. The spectral clustering area where the low-precision soil property prediction model is located is defined as the low-precision soil property prediction area, where M≥0 and L+M=K, and then enter step D.

[0050] Step D: For the high-precision soil property prediction area, a high-precision soil property prediction model is used to perform soil property spatial prediction. For the low-precision soil property prediction area, a spatial interpolation method is used to obtain the spatial distribution of soil properties. After mosaicking the two, a complete soil property spatial distribution map of the region is obtained.

[0051] In practical application, the above step D specifically performs the following steps D1 to D2:

[0052] Step D1, using K high-precision soil property prediction models to perform soil property spatial prediction for corresponding spectral clustering areas;

[0053] If M>0, go to step D2;

[0054] Step D2, for the M low-precision soil property prediction areas, using the prediction results of the high-precision soil property prediction model, adopting the Ordinary Kriging (OK) spatial interpolation method to obtain the soil property spatial prediction results, and then entering step D3;

[0055] Step D3, mosaicking the spatial distribution results predicted by the high-precision soil property prediction model with the spatial interpolation results of the low-precision soil property prediction area to obtain a complete regional soil property spatial distribution map.

[0056] The above-designed clustering-mixing-based UAV hyperspectral soil property monitoring method is applied in practice. The typical cultivated farmland in Fengqiu County, Henan Province is used as the UAV hyperspectral monitoring area, and soil organic matter is used as the monitoring object. The implementation of regional cultivated farmland soil organic matter UAV hyperspectral monitoring is taken as an example and further described in detail. However, it is not intended to limit the present invention and is specifically implemented as follows.

[0057] Step FA, the study area is the tilled farmland area, the scope of the study area and the distribution of soil sample points for subsequent model construction are as follows: Figure 2As shown, for this area, a HySpexMjolnir V-1240 hyperspectral sensor covering a spectral range of 400-1000nm is used to carry a drone to obtain hyperspectral images, and the obtained images are preprocessed to generate ground reflectance images, and then enter step FB.

[0058] The above step EA is specifically as follows in the embodiment:

[0059] Step FA1, according to the scope of the study area, set the lateral overlap ratio to 25%, and the UAV flight direction to the southeast-northwest direction. In order to cover the entire study area, a total of 4 UAV hyperspectral images are obtained; then enter step FA2;

[0060] Step FA2, for the four UAV hyperspectral images, perform radiometric calibration and atmospheric correction processing, delete the bands with large noise less than 451.4nm and greater than 899.1nm, generate ground reflectance images, and then enter step FA3;

[0061] Step FA3, for the four UAV hyperspectral ground reflectance images, image stitching processing is performed to obtain a hyperspectral reflectance image covering the entire study area. The processed UAV hyperspectral image is as follows: Figure 3 As shown, then go to step FB.

[0062] Step FB, perform spectral clustering on the obtained UAV hyperspectral reflectance image, divide the entire image into different areas according to the clustering results, and then enter step FC;

[0063] The above step FB is specifically as follows in the embodiment:

[0064] Step FB1, for the obtained UAV hyperspectral reflectance image, select the band between 700-899.1nm with large spectral reflectance difference, a total of 66 bands, and use Gaussian mixture model GMM to build a spectral clustering model. The optimal number of spectral clusters is 5. The average spectrum and standard deviation of these 5 spectral classes are as follows Figure 4 As shown, then enter step FB2;

[0065] Step FB2, using the constructed spectral clustering model, the entire UAV hyperspectral image is divided into five regions with obvious spectral differences. The results of the five divided regions are as follows: Figure 5 As shown, then go to step FC.

[0066] Step FC, construct a dedicated soil property prediction model for each spectral clustering area, remove the model with low inversion accuracy according to the inversion accuracy, and then enter step FD;

[0067] The above steps FC are specifically as follows in the embodiment:

[0068] Step FC1: For each of the five spectral clustering regions, the DNN algorithm was used to construct a soil organic matter prediction model with soil organic matter as the dependent variable and the spectrum at the soil sample collection point in the UAV hyperspectral reflectance image as the independent variable as input. The cross-validation method was used to calculate the model R 2 and RMSE, where types 1, 2, 3, 4, and 5 correspond to R 2 They are 0.50, 0.67, 0.41, 0.29, and 0.09, respectively, and the corresponding RMSEs are 1.17, 1.00, 1.33, 1.53, and 1.53 g / kg, respectively, and then proceed to step FC2;

[0069] Step FC2, based on the 5 spectral clustering regions R 2 The value is expressed in R 2 The value ≥0.60 is used as the standard to classify it as a high-precision soil property prediction model. According to this standard, type 2 is a high-precision model, and the other four types of regional models are low-precision soil property prediction models. The soil organic matter prediction scatter plot of type 2 is as follows Figure 6 As shown, then go to step FD.

[0070] Step FD: For the high-precision soil property prediction area, the high-precision soil property prediction model is used to perform soil property spatial prediction. For the low-precision soil property prediction area, the spatial interpolation method is used to obtain the spatial distribution of soil properties. After mosaicking the two, a complete soil property spatial distribution map of the region is obtained.

[0071] The above step FD is specifically as follows in the embodiment:

[0072] Step FD1, using a high-precision soil organic matter prediction model, that is, the model of the area where type 2 is located, to perform soil organic matter spatial prediction for the corresponding spectral clustering area; because the number of low-precision model type areas = 4, enter step D2;

[0073] Step FD2, for the four low-precision soil property prediction areas, using the prediction results of the type 2 soil organic matter prediction model, using the OK space interpolation method to obtain the soil organic matter spatial prediction results, and then proceeding to step D3;

[0074] Step FD3, the high-precision soil organic matter spatial distribution results of the area where type 2 is located are mosaicked with the soil organic matter interpolation results of the other four types of areas to obtain a complete regional soil organic matter spatial distribution map. The results are as follows: Figure 7 shown.

[0075] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A clustering-mixing based UAV hyperspectral soil property monitoring method, characterized in that: The steps include: S1, pre-process the acquired UAV hyperspectral image to generate a ground reflectance image; then obtain a hyperspectral reflectance image covering the entire study area through image stitching processing; S2, spectral clustering is performed on the acquired UAV hyperspectral reflectance image, and the entire image is divided into different areas according to the clustering results; S3, construct a dedicated soil property prediction model for each spectral clustering area, and divide it into a high-precision soil property prediction model and a low-precision soil property prediction model according to the inversion accuracy; the implementation steps are as follows: S31, for each of the K spectral clustering regions, a machine learning algorithm was used, with soil properties as the dependent variable and the spectra at the soil sample collection points in the UAV hyperspectral reflectance image as the independent variable as input to construct a soil property prediction model, and the cross-validation or independent validation method was used to calculate the determination coefficient R of the soil property prediction model. 2 and root mean square error RMSE; S32, based on K spectral clustering regions R 2 The value, R 2 The model with a value ≥ 0.60 is a high-precision soil property prediction model, and the corresponding area belongs to the high-precision soil property prediction area, which is recorded as L; R 2 The model with a value of <0.60 is a low-precision soil property prediction model, and the corresponding area belongs to the low-precision soil property prediction area, which is recorded as M, where M≥0 and L+M=K; S4, for the area belonging to the high-precision soil property prediction model, the soil property spatial prediction is performed, and for the area belonging to the low-precision soil property prediction model, the spatial interpolation method is used to obtain the spatial distribution of soil properties, and the spatial distribution results predicted by the high-precision model are mosaicked with the spatial interpolation results of the low-precision soil property prediction area to obtain a complete regional soil property spatial distribution map; the implementation steps are as follows: S41, using L high-precision soil property prediction models, to perform spatial prediction of soil properties for the corresponding spectral clustering areas; If M>0, go to step S42; S42, for the M low-precision soil property prediction areas, using the prediction results of the high-precision soil property prediction model, adopting the ordinary Kruger spatial interpolation method to obtain the soil property spatial prediction results, and then proceeding to step S43; S43, mosaicking the spatial distribution results predicted by the high-precision soil property prediction model and the spatial interpolation results of the low-precision soil property prediction area to obtain a complete regional soil property spatial distribution map.

2. The method for monitoring soil properties using a UAV using a hyperspectral image based on clustering and mixing according to claim 1, characterized in that: In step S1, the steps for obtaining a hyperspectral image covering the entire study area are as follows: S11, according to the scope of the study area, set the lateral overlap ratio, plan and design the UAV flight route, and obtain N UAV hyperspectral images, where N ≥ 1; S12, performing radiometric calibration and atmospheric correction processing on N drone hyperspectral images to generate ground reflectance images; S13, performing image stitching processing on N ground reflectance images to obtain a hyperspectral reflectance image covering the entire study area.

3. The method for monitoring soil properties using a UAV using hyperspectral images based on clustering and mixing according to claim 1, characterized in that: In step S2, according to the clustering results, the entire image is divided into different regions to achieve the following steps: S21, for the obtained UAV hyperspectral reflectance image, a spectral clustering model is constructed using a Gaussian mixture model, and the number of clusters is recorded as K; S22, using the constructed spectral clustering model, the entire UAV hyperspectral image is divided into regions, and a total of K regions with obvious spectral differences are planned.

Images