A long-term image synthesis method for soil remote sensing inversion

Abstract

The invention discloses a long-time image synthesis method for soil remote sensing inversion, comprising the following steps: S1, taking ten days as a basic unit, determining the potential bare soil window period of soil remote sensing inversion in the study area; S2, obtaining the long-time remote sensing images within the potential bare soil window period in the study area, and preprocessing the obtained remote sensing images; S3, respectively constructing a single-view image soil property prediction model for each remote sensing image after preprocessing, sorting the remote sensing images in descending order of inversion accuracy, and deleting remote sensing images with low inversion accuracy; S4, performing image synthesis on remote sensing images of different numbers with the highest inversion accuracy; S5, constructing a synthetic image soil property prediction model for each synthetic image, selecting the optimal synthetic image and its corresponding prediction model based on the model prediction result, and generating a regional soil property prediction map. The invention can improve the accuracy of remote sensing inversion of soil properties.

Description

Technical Field

[0001] The present invention relates to the technical field of soil property acquisition, and in particular to a long-time series image synthesis method for soil remote sensing inversion. Background Art

[0002] Soil is an important material basis for crop growth. High-quality soil resources can increase crop yields and improve crop quality. Accurately and quickly obtaining the spatial distribution pattern of soil properties is of great significance for formulating soil resource development and utilization plans and for the efficient use of field soil resources. The relevant technologies for rapid and accurate monitoring of soil properties are the technical prerequisite for revealing their pattern.

[0003] Although traditional soil property analysis and testing methods are highly accurate, the relevant methods generally require relatively complex sample pre-processing and have relatively low test efficiency. In addition, in order to achieve tasks such as regional soil property mapping, a large number of soil samples are often required, so the overall cost is high and it is difficult to meet the current needs for soil information in various fields. In recent years, remote sensing technology has developed rapidly, especially the emergence of some remote sensing big data analysis platforms, which has simplified the complexity of soil remote sensing in data acquisition, data analysis, model construction and other links.

[0004] Currently, remote sensing technology has been used to monitor soil properties at multiple scales, and has demonstrated high inversion accuracy. Generally speaking, for small-scale applications, a single remote sensing image can cover the entire study area, while for larger-scale studies, multi-image synthesis applications are often required. In addition, studies on the inversion of soil properties using long-term remote sensing images have shown that long-term remote sensing images can improve the accuracy of soil remote sensing inversion, and image synthesis is an important method for utilizing long-term remote sensing information. Therefore, image synthesis is required from both aspects.

[0005] The remote sensing image synthesis methods proposed in recent years generally directly perform arithmetic operations on the images within the potential bare soil window period, such as the mean or median of multiple images. However, these methods do not take into account the spatial differences in regional farmland management measures, such as differences in cultivated land types and irrigation dates. This results in the images obtained during the potential bare soil window period still being affected by factors such as soil moisture and straw coverage, which in turn interferes with the image soil spectrum and reduces the accuracy of soil remote sensing inversion. Summary of the invention

[0006] Purpose of the invention: The purpose of the present invention is to provide a long-time image synthesis method for soil remote sensing inversion, which can improve the accuracy of rapid acquisition of regional soil properties and improve the quality of actual work.

[0007] Technical solution: The long-time image synthesis method for soil remote sensing inversion of the present invention comprises the following steps:

[0008] S1, taking ten days as the basic unit, determines the potential bare soil window period for soil remote sensing inversion in the study area;

[0009] S2, obtain long-term remote sensing images within the potential bare soil window period in the study area, and preprocess the acquired remote sensing images;

[0010] S3, constructing a single-image soil property prediction model for each remote sensing image after preprocessing, sorting the remote sensing images in descending order of inversion accuracy, and deleting remote sensing images with low inversion accuracy;

[0011] S4, performing image synthesis on remote sensing images of different numbers with the highest inversion accuracy;

[0012] S5, constructing a synthetic image soil property prediction model for each synthetic image, selecting the optimal synthetic image and its corresponding prediction model based on the model prediction results, and generating a regional soil property prediction map.

[0013] Furthermore, in step S1, the steps for determining the potential bare soil window period for soil remote sensing inversion in the study area are as follows:

[0014] S11, obtain the crop planting situation in the study area, determine the sowing and harvesting dates of different crops in the year, and record the day without crop growth as T1;

[0015] S12, obtain the snowfall conditions in the study area, determine the first and last days of snowfall in the area, and record the day when the soil surface is not covered with snow as T2;

[0016] S13, calculate the intersection date of T1 and T2, denoted as T, as the potential time window for soil inversion in the study area.

[0017] Further, the step S2 comprises the following steps:

[0018] S21, taking the year of soil sample collection as the center, obtain remote sensing images in the potential bare soil window period T within a total of 5 years before and after;

[0019] If the acquired image has completed the radiometric calibration and atmospheric correction processing, directly proceed to step S3;

[0020] S22, performing radiometric calibration preprocessing on the acquired remote sensing image;

[0021] S23, performing atmospheric correction processing on the remote sensing image after radiometric calibration using the FLAASH atmospheric correction model.

[0022] Further, the step S3 comprises the following steps:

[0023] S31, for each remote sensing image after preprocessing, use machine learning algorithm to build a single image soil property prediction model, and use cross-validation or independent validation method to calculate the model determination coefficient R 2 and root mean square error RMSE;

[0024] S32, according to R 2 The value is large, delete R 2 Remote sensing images with a value ≤ 0.05.

[0025] Furthermore, in step S4, the steps for implementing image synthesis of different numbers of remote sensing images with relatively high inversion accuracy are as follows:

[0026] S41, the total number of images is recorded as K, and in the order of soil property prediction accuracy of each scene image from high to low, 2 to K-1 images are selected to form image groups, where K is greater than 3, and the image number interval between each image group is 1, and a total of K-2 image groups are constructed;

[0027] S42, for each image group, calculating a synthetic image according to the median or average value, and constructing K-2 synthetic images in total.

[0028] Furthermore, in step S5, the steps for generating a regional soil property prediction map are as follows:

[0029] S51: For each synthetic image, a machine learning algorithm is used to construct a synthetic image soil property prediction model, and a cross-validation or independent validation method is used to calculate the determination coefficient R 2 and root mean square error RMSE;

[0030] S52, according to R 2 The optimal synthetic image and its corresponding prediction model were selected based on RMSE to carry out regional soil property prediction and mapping.

[0031] Compared with the prior art, the present invention has the following significant effects:

[0032] The present invention firstly reduces the influence of factors such as crop cover and snow accumulation on the accuracy of remote sensing inversion of soil properties by selecting a potential bare soil period window; secondly, by constructing a single-image soil property prediction model for each remote sensing image, remote sensing images with low soil property inversion accuracy are deleted, and the interference of abnormal remote sensing images on soil property inversion is removed; finally, different image combinations are performed in descending order according to the soil property prediction accuracy of each image, and the optimal synthetic image and its corresponding synthetic image soil property prediction model are screened, which can improve the accuracy of remote sensing inversion of soil properties, provide a new idea for realizing efficient remote sensing monitoring application of soil properties, and promote the development of soil remote sensing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the present invention;

[0034] Figure 2 To implement the case study, the spatial distribution of soil collection points and the dryland boundary map;

[0035] Figure 3 Scatter plot of predicted soil organic matter results in the implementation case;

[0036] Figure 4 This is the spatial distribution map of soil organic matter predicted in the implementation case. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0038] In soil remote sensing inversion applications, among the remote sensing images within the potential bare soil window period, some images are severely affected. For example, irrigation leads to an increase in soil moisture content. These images cannot invert soil properties and need to be excluded before image synthesis. In addition, due to external factors or the quality of the image itself, there may be some images with low soil property inversion accuracy. This part of the influence also needs to be identified and removed during synthesis. Therefore, after removing the relevant images, the influence of synthesis can reduce the interference of soil property inversion and thus improve the accuracy. Therefore, the present invention provides a long-time series image synthesis method for soil remote sensing inversion, which is used for regional soil property remote sensing inversion. In practical applications, such as Figure 1 As shown, the specific steps include:

[0039] Step A, taking ten days as the basic unit, determines the potential bare soil window period for soil remote sensing inversion in the study area, and then proceeds to step B.

[0040] In practical application, the above step A specifically performs the following steps A1 to A3:

[0041] Step A1, obtain the crop planting situation in the study area, determine the sowing and harvesting dates of different crops in the year, and record the day without crop growth as T1;

[0042] Step A2, obtaining the snowfall conditions in the study area, determining the first and last snowfall dates in the area, recording the date when the soil surface is not covered with snow as T2, and then proceeding to step A3;

[0043] In step A3, the intersection date of T1 and T2 is calculated, recorded as T, as the potential time window for soil inversion in the study area, and then step B is entered.

[0044] Step B, obtaining long-term remote sensing images within the potential bare soil window period in the study area, and preprocessing the obtained remote sensing images, and then proceeding to step C;

[0045] In practical application, the above step B specifically performs the following steps B1 to B3:

[0046] Step B1, taking the soil sample collection year as the center, obtain remote sensing images in the potential bare soil window period T within a total of 5 years before and after; if the acquired images have completed radiation calibration and atmospheric correction processing, directly enter step C;

[0047] Step B2, performing radiometric calibration preprocessing on the acquired remote sensing image, and then proceeding to step B3;

[0048] Step B3, for the remote sensing image after radiometric calibration, the atmospheric correction processing is performed using the FLAASH (The Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) atmospheric correction model, and then the process proceeds to step C.

[0049] Step C, constructing a single-image soil property prediction model for each remote sensing image after preprocessing, sorting the remote sensing images in descending order of inversion accuracy, deleting remote sensing images with low inversion accuracy, and then entering step D;

[0050] In practical application, the above step C specifically performs the following steps C1 to C2:

[0051] Step C1: For each remote sensing image after preprocessing, a machine learning algorithm such as partial least squares regression (PLSR) or random forest can be used to build a single image soil property (such as organic matter, soil clay content, etc.) prediction model, and a cross-validation or independent validation method is used to calculate the model determination coefficient R. 2 (Coefficient of determination) and root mean square error RMSE (Root Mean Square Error), and then proceed to step C2;

[0052] Step C2, according to R 2 The value is large, delete R 2 Remote sensing images with a value ≤ 0.05, then go to step D.

[0053] Step D, performing image synthesis on different numbers of remote sensing images with the highest inversion accuracy, and then proceeding to step E;

[0054] In practical application, the above step D specifically performs the following steps D1 to D2:

[0055] Step D1, record the total number of images as K, select 2 to K-1 images to form image groups in the order of soil property prediction accuracy of each image from high to low, where K is greater than 3, and the image number interval between each image group is 1, construct a total of K-2 image groups, and then enter step D2;

[0056] Step D2: For each image group, a synthetic image is calculated according to the median or average value, and a total of K-2 synthetic images are constructed, and then the process goes to step E.

[0057] Step E: construct a synthetic image soil property prediction model for each synthetic image, select the optimal synthetic image and its corresponding prediction model based on the model prediction results, and generate a regional soil property prediction map.

[0058] In practical application, the above step E specifically performs the following steps E1 to E2:

[0059] Step E1: For each synthetic image, a machine learning algorithm such as PLSR or random forest can be used to construct a synthetic soil property (such as organic matter, soil clay content, etc.) prediction model, and a cross-validation or independent validation method can be used to calculate R 2 and RMSE, and then proceed to step E2;

[0060] Step E2, according to R 2 The two indicators of RMS and RMSE are used to screen the optimal synthetic image and its corresponding synthetic image soil property prediction model for regional soil property prediction mapping.

[0061] The long-term image synthesis method designed for soil remote sensing inversion is applied in practice. Taking Youyi County, Heilongjiang Province as the remote sensing inversion area and the soil organic matter in the dryland farmland area as the inversion object, the remote sensing inversion of soil organic matter in the dryland farmland area is taken as an example to further describe in detail, but it is not intended to limit the present invention and is specifically implemented as follows.

[0062] Step FA, taking ten days as the basic unit, determines the potential bare soil window period for soil remote sensing inversion in the study area. The distribution of sampling points for dry land in the study area and actual sampling for subsequent model construction is as follows: Figure 2 Then go to step B.

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

[0064] Step FA1: Corn and soybeans are mainly grown in the dryland farmland in the study area. Generally, the sowing date of crops is from late May to mid-June, and the latest harvest is in early November. In order to retain a longer bare soil window period, the date T1 without crop growth is recorded as mid-November to mid-June of the following year, and then enter step FA2;

[0065] Step FA2: There is a long period of snowfall in the study area. Generally, the first day of snowfall is from early October to early November and the last day is from early April. In order to retain a longer bare soil window period, the date T2 when the soil surface is not covered by snow is from early April to early November, and then enter step FB3;

[0066] In step FA3, the intersection month of T1 and T2 is calculated, with T being from early April to mid-June, as the potential time window for soil inversion in the study area, and then step B is entered.

[0067] Step FB, obtaining long-term remote sensing images within the potential bare soil window period in the study area, and preprocessing the obtained remote sensing images, and then entering step FC;

[0068] The above step FB is specifically as follows in the embodiment:

[0069] Step FB1, taking the soil sample collection year 2021 as the center, obtain the Sentinel-2 remote sensing images between early April and mid-June during the potential bare soil window period from 2019 to 2023, and select 10 bands (central wavelengths are: 490, 560, 665, 705, 740, 783, 842, 865, 1610, 2190nm) for soil organic matter prediction, a total of 25 scenes; since the acquired images have completed radiation calibration and atmospheric correction processing, directly enter step FC.

[0070] Step FC, constructing a single-image soil property prediction model for each remote sensing image after preprocessing, sorting the remote sensing images in descending order of inversion accuracy, deleting remote sensing images with low inversion accuracy, and then entering step FD;

[0071] The above steps FC are specifically as follows in the embodiment:

[0072] Step FC1: For each remote sensing image after preprocessing, use the PLSR algorithm to build a prediction model for soil organic matter, and use cross-validation or independent validation methods to calculate the model determination coefficient R 2 and root mean square error RMSE, and then proceed to step FC2;

[0073] Step FC2, according to R 2The remote sensing images with R2 values ​​≤ 0.05 are deleted, and a total of 6 images are deleted, leaving 19 images, and then go to step D.

[0074] Step FD, performing image synthesis for different numbers of remote sensing images with the highest inversion accuracy, and then entering step FE;

[0075] The above step FD is specifically as follows in the embodiment:

[0076] Step FD1, the total number of images is 19. According to the order of soil property prediction accuracy of each scene image from high to low, 2 to 18 images are selected to form image groups. The image number interval between each image group is 1, and a total of 17 image groups are constructed, and then enter step D2;

[0077] Step FD2, for each image group, calculate the synthetic image according to the average value, construct a total of 17 synthetic images, and then enter step FE.

[0078] In step FE, a synthetic image soil property prediction model is constructed for each synthetic image. Based on the model prediction results, the optimal synthetic image and its corresponding prediction model are selected to generate a regional soil property prediction map.

[0079] The above steps FE are specifically as follows in the embodiment:

[0080] Step FE1: For each synthetic image, use the PLSR algorithm to build a synthetic image soil organic matter prediction model, and use cross-validation or independent validation methods to calculate R 2 and RMSE, and then proceed to step E2;

[0081] Step FE2, according to R 2 and RMSE. The optimal synthetic image combination consists of the top 5 images in terms of prediction accuracy. The corresponding prediction model accuracy R 2 =0.74, RMSE = 0.50%, the prediction scatter plot is as follows Figure 3 As shown in Figure 2, the model is used to map regional soil organic matter. The spatial distribution map of organic matter is shown in Figure 2. Figure 4 shown.

[0082] 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 ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A long-term image synthesis method for soil remote sensing inversion, characterized in that: The steps include: S1, taking ten days as the basic unit, determines the potential bare soil window period for soil remote sensing inversion in the study area; S2, obtain long-term remote sensing images within the potential bare soil window period in the study area, and preprocess the acquired remote sensing images; S3, constructing a single-image soil property prediction model for each remote sensing image after preprocessing, sorting the remote sensing images in descending order of inversion accuracy, and deleting remote sensing images with low inversion accuracy; S4, image synthesis is performed on different numbers of remote sensing images with the highest inversion accuracy. The implementation steps are as follows: S41, the total number of images is recorded as K, and in the order of soil property prediction accuracy of each scene image from high to low, 2 to K-1 images are selected to form image groups, where K is greater than 3, and the image number interval between each image group is 1, and a total of K-2 image groups are constructed; S42, for each image group, calculating a synthetic image according to the median or average value, and constructing K-2 synthetic images in total; S5, constructing a synthetic image soil property prediction model for each synthetic image, selecting the optimal synthetic image and its corresponding prediction model based on the model prediction results, and generating a regional soil property prediction map.

2. The long-term image synthesis method for soil remote sensing inversion according to claim 1 is characterized in that: In step S1, the steps for determining the potential bare soil window period for soil remote sensing inversion in the study area are as follows: S11, obtain the crop planting situation in the study area, determine the sowing and harvesting dates of different crops in the year, and record the day without crop growth as T1; S12, obtain the snowfall conditions in the study area, determine the first and last days of snowfall in the area, and record the day when the soil surface is not covered with snow as T2; S13, calculate the intersection date of T1 and T2, denoted as T, as the potential time window for soil inversion in the study area.

3. The long-term image synthesis method for soil remote sensing inversion according to claim 1 is characterized in that: The step S2 comprises the following steps: S21, taking the year of soil sample collection as the center, obtain remote sensing images in the potential bare soil window period T within a total of 5 years before and after; If the acquired image has completed the radiometric calibration and atmospheric correction processing, directly proceed to step S3; S22, performing radiometric calibration preprocessing on the acquired remote sensing image; S23, performing atmospheric correction processing on the remote sensing image after radiometric calibration using the FLAASH atmospheric correction model.

4. The long-term image synthesis method for soil remote sensing inversion according to claim 1 is characterized in that: The step S3 comprises the following steps: S31: For each remote sensing image after preprocessing, a machine learning algorithm is used to construct a single image soil property prediction model, and the model determination coefficient R is calculated using cross-validation or independent validation methods. 2 and root mean square error RMSE; S32, according to R 2 The value is large, delete R 2 Remote sensing images with a value ≤ 0.

05.

5. The long-term image synthesis method for soil remote sensing inversion according to claim 1 is characterized in that: In step S5, the steps for generating a regional soil property prediction map are as follows: S51: For each synthetic image, a machine learning algorithm is used to construct a synthetic image soil property prediction model, and a cross-validation or independent validation method is used to calculate the determination coefficient R 2 and root mean square error RMSE; S52, according to R 2 The optimal synthetic image and its corresponding prediction model were selected based on RMSE to carry out regional soil property prediction and mapping.

Images