Method for extracting farmland low-lying waterlogging based on spectrum-process-terrain coordination

Abstract

The application discloses a kind of based on spectrum-process-terrain coordination's farmland low-lying quick extraction method of waterlogging, comprising the following steps: step one, satellite remote sensing image acquisition and pretreatment;Step two, waterlogging area grading mark;Step three, process quantity extraction;Step four, red edge area index calculation;Step five, terrain factor processing and feature combination;Step six, model training and optimal feature subset screening;The application is by constructing bare soil period soil moisture variation process quantity and growth period crop growth change process quantity, accurately depict the evolution track of waterlogging with growth stage, break through the limitation that single phase cannot reflect process difference;By constructing red edge area index, enhance the spectrum difference of waterlogging, distinguish bare soil period easily confused ground class, avoid growth period vegetation canopy shielding interference;Through terrain factor, reflect the influence of topography on water depth, holding water time length, improve the identification accuracy of waterlogging spatial heterogeneity;By using random forest algorithm, improve the accuracy of waterlogging grading.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of remote sensing, in particular to a farmland low-lying waterlogging rapid extraction method based on spectrum-process-terrain cooperation. BACKGROUND

[0002] Farmland low-lying waterlogging is the primary limiting factor of stable and high yield of crops. Current monitoring technologies for farmland low-lying waterlogging mainly include two types. One type is a manually guided monitoring technology, which relies on on-site measurement of water accumulation gauges, manual on-site delineation of waterlogging areas or small-range aerial photography by unmanned aerial vehicles, and the like. The technology can guarantee the measurement accuracy of local points, but can only obtain sample data of a small number of farmland plots, cannot realize the synchronous monitoring of a large range of farmland, and needs manual on-site operation, which is time-consuming and labor-intensive, has high labor and time costs, and has poor timeliness, and thus cannot meet the demand for rapid monitoring of large-scale farmland waterlogging. The other type is a satellite remote sensing technology for rapidly obtaining farmland water accumulation information. A single-time NDVI / NDWI threshold method is adopted to identify the waterlogging area by setting a fixed index threshold. The method has high automation, but still has the following shortcomings: 1. The spectral characteristics of dry soil and thin-layer water accumulation areas in the bare soil period are similar, and the single-time NDVI / NDWI threshold method cannot effectively distinguish between the two, resulting in misjudgment or omission of the waterlogging area; 2. The vegetation canopy in the crop growth period can block the water surface below, making it difficult for the NDWI index to accurately capture the water surface information, and greatly reducing the waterlogging identification accuracy; and 3. The existing technology does not consider the waterlogging evolution process from the bare soil period to the crop growth period, and only relies on single-time data, which cannot utilize the differences in waterlogging characteristics in different growth stages to improve the monitoring accuracy, and thus limits the overall identification effect. SUMMARY

[0003] The application aims to provide a farmland low-lying waterlogging rapid extraction method based on spectrum-process-terrain cooperation to solve the problems in the background technology.

[0004] To achieve the above-mentioned purpose, the application provides the following technical scheme: a farmland low-lying waterlogging rapid extraction method based on spectrum-process-terrain cooperation, comprising the following steps: step one, satellite remote sensing image acquisition and preprocessing; step two, waterlogging area hierarchical labeling; step three, process quantity extraction; step four, red edge area index calculation; step five, terrain factor processing and feature combination; and step six, model training and optimal feature subset screening.

[0005] In the above step one, cloud-free Sentinel-2A satellite remote sensing images in the bare soil period and the crop growth period in the target area are selected, the images are preprocessed, and the NDWI and NDVI of each image are calculated.

[0006] In the above step two, the pre-processed remote sensing image is cropped according to the target area farmland plot vector range, the image after the rainfall in the crop growth period is screened, a double window linkage mode is used for visual interpretation, a five-level label is marked for the farmland low-lying waterlogging area, and a plot-level label vector library is generated.

[0007] In the above step three, a time series curve of NDWI in the bare soil period and NDVI in the crop growth period is drawn, and a bare soil period soil moisture change process quantity ΔNDWI and a growth period crop vegetation change process quantity ΔNDVI are obtained based on the time series curve.

[0008] In the above step four, based on the red edge bands B5, B6 and B7 of the Sentinel-2A satellite remote sensing image, the red edge area index is calculated by using the trapezoidal integral principle.

[0009] In the above step five, the terrain factor is extracted from the DEM as the terrain source, and the terrain factor is resampled to the resolution consistent with the spectral characteristics of the Sentinel-2A satellite remote sensing image, and then the image bands, NDWI, NDVI, ΔNDWI, ΔNDVI obtained in step three, and the red edge area index obtained in step four are connected in series at the pixel level to form a multi-dimensional feature vector.

[0010] In the above step six, the multi-dimensional feature vector obtained in step five is divided into a training set and an independent validation set, a plurality of feature combination schemes are constructed to drive the random forest model training and out-of-bag error evaluation, and OA, Kappa coefficient and feature importance sequence are simultaneously output, the recognition efficiency under different combinations is compared, the optimal feature subset with the smallest out-of-bag error and the most significant Kappa improvement is selected, and the optimization and verification of the waterlogging grading model are completed.

[0011] Preferably, in the step one, the bare soil period is from April to May, and the crop growth period is from June to September; the pre-processing specifically includes radiation calibration, atmospheric correction and geometric correction.

[0012] Preferably, in the step one, the green band wavelength used for calculating NDWI is 0.559-0.570 μm, and the near-infrared band wavelength is 0.833-0.864 μm; the red band wavelength used for calculating NDVI is 0.650-0.680 μm, and the near-infrared band wavelength is 0.833-0.864 μm; in the Sentinel-2A satellite remote sensing image, the green band corresponds to band B3, the red band corresponds to band B4, and the near-infrared band corresponds to band B8; the calculation formulas of NDWI and NDVI are respectively:

[0013]

[0014] .

[0015] Preferably, in step two, the two windows of the dual-window linkage mode load a true color combination and a false color combination of images respectively. The true color combination is bands B4, B3, and B2, and the false color combination is bands B8, B4, and B3.

[0016] Preferably, in step two, the five-level labels are: Level 0 no waterlogging, Level 1 mild waterlogging, Level 2 moderate waterlogging, Level 3 severe waterlogging, and Level 4 extremely severe waterlogging, based on the differences in the color brightness and spatial continuity of farmland water bodies.

[0017] Preferably, in step three, the calculation formulas for ΔNDWI and ΔNDVI are as follows:

[0018]

[0019]

[0020] in, The image dates are from early April during the bare soil period. The image date is mid-April during the bare soil period. It is the bare soil period NDWI value of date It is the bare soil period NDWI value of date It represents the change in soil moisture during the bare soil period. Image dates for the growing season in early June. Image dates in late June during the growing season. It is the crop growth period NDVI value of the date It is the crop growth period NDVI value of the date It is a quantity of vegetation change during the crop growth period.

[0021] Preferably, in step four, the center wavelength of the red-edge band B5 of the Sentinel-2A satellite remote sensing image is 705 nm, the center wavelength of band B6 is 740 nm, and the center wavelength of band B7 is 783 nm; the formula for calculating the red-edge area index is:

[0022]

[0023] in, It is the red-edge area index. It is the surface reflectance of band B5. It is the surface reflectance of band B6. It is the surface reflectance of band B7. It is the difference in center wavelength between bands B5 and B6. is the center wavelength difference between the wave bands B6 and B7.

[0024] Preferably, in the step five, the spatial resolution of the DEM is 6m, the extracted terrain factors include slope, TWI and FA, and the resolution of the resampled terrain factors is 10m.

[0025] Preferably, in the step five, the terrain factor extraction is completed through the SAGA GIS platform, and the pixel-level series is completed through the Google Earth Engine platform.

[0026] Preferably, in the step six, the multi-dimensional feature vector is divided in a hierarchical random sampling manner, and the multi-group feature combination scheme includes a spectral single-factor combination, a spectral-terrain joint combination and a spectral-process-terrain full-coupling combination.

[0027] Compared with the prior art, the present application has the beneficial effects that: the present application breaks through the limitation that a single time phase cannot reflect process differences by constructing soil water variation process quantity in the bare soil period and crop growth variation process quantity in the growth period, accurately depicting the evolution trajectory of waterlogging with the growth stage; the present application enhances the spectral difference of waterlogging by constructing a red edge area index, distinguishes easily confused land types in the bare soil period, and avoids the interference of the growth period vegetation canopy; the present application reflects the influence of terrain on water depth and water holding time through terrain factors, and improves the spatial heterogeneity recognition accuracy of waterlogging; the present application improves the accuracy of waterlogging classification by using the random forest algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a method flowchart of the present application;

[0029] Figure 2 is a bare soil period NDWI time series curve diagram;

[0030] Figure 3 is a growth period NDVI time series curve diagram;

[0031] Figure 4 is an image wave band change curve diagram. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0033] Please refer to the accompanying Figure 1The application provides a method for extracting farmland low-lying waterlogging based on spectrum-process-terrain cooperation, which comprises the following steps: step 1, satellite remote sensing image acquisition and pretreatment; step 2, low-lying waterlogging area grading labeling; step 3, process quantity extraction; step 4, red edge area index calculation; step 5, terrain factor processing and feature combination; and step 6, model training and optimal feature subset screening.

[0034] In step one, the cloud-free Sentinel-2A satellite remote sensing images in the target region during the bare soil period and the crop growth period are selected, the images are pretreated through radiation calibration, atmospheric correction and geometric correction, and then the NDWI (Normalized Difference Water Index) and the NDVI (Normalized Difference Vegetation Index) of each image are calculated; the bare soil period is from April to May, and the crop growth period is from June to September; the green light band wavelength for calculating the NDWI is 0.559-0.570 μm, and the near-infrared band wavelength is 0.833-0.864 μm; the red light band wavelength for calculating the NDVI is 0.650-0.680 μm, and the near-infrared band wavelength is 0.833-0.864 μm; in the Sentinel-2A satellite remote sensing image, the green light band corresponds to band B3, the red light band corresponds to band B4, and the near-infrared band corresponds to band B8; the calculation formulas of the NDWI and the NDVI are respectively:

[0035]

[0036] ;

[0037] In step two, the pretreated remote sensing image is cropped according to the farmland plot vector range of the target area, the image after the rainfall in the crop growth period is screened, a double-window linkage mode is adopted for visual interpretation, the farmland low-lying waterlogging area is labeled with five levels of labels and a plot-level label vector library is generated; in the double-window linkage mode, two windows are loaded with the true color combination and the false color combination of the image, the true color combination is band B4, B3 and B2, and the false color combination is band B8, B4 and B3; the five levels of labels are: 0 level without waterlogging, 1 level with slight waterlogging, 2 level with moderate waterlogging, 3 level with severe waterlogging and 4 level with extremely severe waterlogging, and the labeling is based on the difference in color brightness and spatial continuity of the farmland water body;

[0038] In step three, the time sequence curves of the NDWI in the bare soil period and the NDVI in the crop growth period are drawn, and the soil moisture variation process quantity ΔNDWI in the bare soil period and the crop vegetation variation process quantity ΔNDVI in the growth period are obtained based on the time sequence curves; the calculation formulas of the ΔNDWI and the ΔNDVI are respectively:

[0039]

[0040]

[0041] wherein, is the image date of the bare soil period in early April, is the image date of the bare soil period in mid April, is the bare soil period NDWI value of the date, is the bare soil period NDWI value of the date, is the soil moisture change process quantity of the bare soil period, is the image date of the growth period in early June, is the image date of the growth period in late June, is the growth period NDVI value of the date, is the growth period NDVI value of the date, is the vegetation change process quantity of the growth period;

[0042] wherein, in the above step four, based on the red edge bands B5, B6 and B7 of the Sentinel-2A satellite remote sensing image, the red edge area index (REA) is calculated by using the trapezoidal integral principle; wherein, the center wavelength of the red edge band B5 of the Sentinel-2A satellite remote sensing image is 705 nm, the center wavelength of the band B6 is 740 nm, and the center wavelength of the band B7 is 783 nm; the calculation formula of the red edge area index is:

[0043]

[0044] wherein, is the red edge area index, is the band B5 ground reflectivity, is the band B6 ground reflectivity, is the band B7 ground reflectivity, is the center wavelength difference between the bands B5 and B6, is the center wavelength difference between the bands B6 and B7;

[0045] In step five above, a DEM (Digital Elevation Model) with a spatial resolution of 6m is used as the terrain source. Terrain factors are extracted through the SAGA GIS platform. The terrain factors include slope, TWI (Topographic Wetness Index), and FA (Flow Accumulation). The terrain factors are resampled to a resolution of 10m that is consistent with the spectral characteristics of the Sentinel-2A satellite remote sensing image. Then, the terrain factors are concatenated with the image bands, NDWI, NDVI, and ΔNDWI, ΔNDVI obtained in step three, and the red edge area index obtained in step four on the Google Earth Engine platform at the pixel level to form a multidimensional feature vector.

[0046] In step six above, a stratified random sampling method is used to divide the multidimensional feature vectors obtained in step five into a training set and an independent validation set. Multiple feature combination schemes are constructed to drive the training of the random forest model and the evaluation of out-of-bag error. The overall accuracy (OA), Kappa coefficient, and feature importance sequence are output simultaneously. By comparing the recognition performance under different combinations, the optimal feature subset with the smallest out-of-bag error and significant improvement in Kappa is selected to complete the optimization and validation of the urban flooding classification model. Among them, the multiple feature combination schemes include spectral single-factor combination, spectral-topography joint combination, and spectral-process-topography fully coupled combination.

[0047] Experimental example:

[0048] To verify the effectiveness of this method, an experiment was conducted using Friendship Farm as the target area, employing the method proposed in the embodiments. Nine different feature combination schemes were designed for the experiment. The experimental results are shown in Table 1 and Appendix 2. Figure 2 - Appendix Figure 4 The experimental results show that the feature combination of "bare soil period image + NDWI + REA + topographic factor + ΔNDWI; growing season image + NDVI + REA + topographic factor + ΔNDVI" has the best effect, with an OA of 0.73 and a Kappa coefficient of 0.66. It can most accurately identify and classify low-lying waterlogged areas in Youyi Farm.

[0049] Table 1

[0050]

[0051] Based on the above, the advantages of this invention are as follows: When used, it accurately depicts the evolution trajectory of waterlogging with different growth stages by constructing soil moisture change processes during the bare soil period and crop growth change processes during the growing season, overcoming the limitation that relying solely on a single time phase cannot reflect process differences; it constructs a red edge area index based on the Sentinel-2A red edge band, which is sensitive to leaf moisture, surface water accumulation, and crop stress response, significantly enhancing the spectral differences of different levels of waterlogging, effectively distinguishing easily confused land types during the bare soil period, while avoiding the shading interference of vegetation canopy on water surface information during the growing season; it accurately reflects the influence of topography on water depth and water retention duration by incorporating topographic factors, improving the accuracy of identifying spatial heterogeneity of waterlogging; and it drives iterative training of the random forest model through multi-feature combination, selecting the optimal feature subset, avoiding the rigidity of fixed thresholds, and significantly improving the accuracy of waterlogging classification and model robustness, providing more reliable technical support for accurately identifying low-lying waterlogging in farmland.

[0052] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy, comprising the following steps: Step 1: Satellite remote sensing image acquisition and preprocessing; Step 2: Hierarchical labeling of flooded areas; Step 3: Extraction of process quantities; Step 4: Calculation of red-edge area index; Step 5: Topographic factor processing and feature combination; Step 6: Model training and selection of optimal feature subset; Its characteristics are: In step one above, cloudless Sentinel-2A satellite remote sensing images of the bare soil period and crop growth period in the target area are selected. After image preprocessing, the NDWI and NDVI of each period of image are calculated. In step two above, the preprocessed remote sensing images are cropped according to the vector range of farmland plots in the target area, images after rainfall during the crop growing season are selected, and visual interpretation is performed using a dual-window linkage mode. Five-level labels are added to low-lying and waterlogged areas of farmland, and a plot-level label vector library is generated. In step three above, time series curves of NDWI during the bare soil period and NDVI during the crop growing period are plotted, and the soil moisture change process ΔNDWI during the bare soil period and the crop vegetation change process ΔNDVI during the growing period are obtained based on the time series curves. In step four above, the red edge area index is calculated using the trapezoidal integral principle based on the red edge bands B5, B6 and B7 of Sentinel-2A satellite remote sensing imagery. In step five above, topographic factors are extracted using DEM as the topographic source. The topographic factors are then resampled to a resolution consistent with the spectral characteristics of Sentinel-2A satellite remote sensing images. They are then concatenated with image bands, NDWI, NDVI, and the ΔNDWI, ΔNDVI obtained in step three and the red edge area index obtained in step four at the pixel level to form a multidimensional feature vector. In step six above, the multidimensional feature vectors obtained in step five are divided into training set and independent validation set. Multiple feature combination schemes are constructed to drive the training of random forest model and the evaluation of out-of-bag error. OA, Kappa coefficient and feature importance sequence are output simultaneously. By comparing the recognition performance under different combinations, the optimal feature subset with the smallest out-of-bag error and significant improvement in Kappa is selected to complete the optimization and validation of the urban flooding classification model. In step two, the five-level labels are: Level 0 no waterlogging, Level 1 mild waterlogging, Level 2 moderate waterlogging, Level 3 severe waterlogging, and Level 4 extremely severe waterlogging. The labeling is based on the differences in the color brightness and spatial continuity of farmland water bodies. In step three, the calculation formulas for ΔNDWI and ΔNDVI are as follows: ， ， in, The image dates are from early April during the bare soil period. The image date is mid-April during the bare soil period. It is the bare soil period NDWI value of date It is the bare soil period NDWI value of date It represents the change in soil moisture during the bare soil period. Image dates for the growing season in early June. Image dates in late June during the growing season. It is the crop growth period NDVI value of the date It is the crop growth period NDVI value of the date It is a quantity related to vegetation changes during the crop growing season; In step four, the center wavelength of the red-edge band B5 in the Sentinel-2A satellite remote sensing image is 705 nm, the center wavelength of band B6 is 740 nm, and the center wavelength of band B7 is 783 nm; the formula for calculating the red-edge area index is: ， in, It is the red-edge area index. It is the surface reflectance of band B5. It is the surface reflectance of band B6. It is the surface reflectance of band B7. It is the difference in center wavelength between bands B5 and B6. It is the difference in center wavelength between bands B6 and B7.

2. The rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy as described in claim 1, characterized in that: In step one, the bare soil period is from April to May, and the crop growing period is from June to September; the pretreatment specifically includes radiation calibration, atmospheric correction, and geometric correction.

3. The rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy as described in claim 1, characterized in that: In step one, the wavelengths used for calculating NDWI are 0.559–0.570 μm for the green band and 0.833–0.864 μm for the near-infrared band; the wavelengths used for calculating NDVI are 0.650–0.680 μm for the red band and 0.833–0.864 μm for the near-infrared band. In Sentinel-2A satellite remote sensing imagery, the green band corresponds to band B3, the red band to band B4, and the near-infrared band to band B8. The calculation formulas for NDWI and NDVI are as follows: ， 。 4. The rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy as described in claim 1, characterized in that: In step two, the two windows of the dual-window linkage mode load the true color combination and the false color combination of the image respectively. The true color combination is bands B4, B3, and B2, and the false color combination is bands B8, B4, and B3.

5. The rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy as described in claim 1, characterized in that: In step five, the spatial resolution of the DEM is 6m, and the extracted terrain factors include slope, TWI and FA. The resolution of the terrain factors after resampling is 10m.

6. The rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy as described in claim 1, characterized in that: In step five, terrain factor extraction is completed through the SAGA GIS platform, and pixel-level concatenation is completed through the Google Earth Engine platform.

7. The rapid extraction method for low-lying waterlogged areas in farmland based on spectral-process-topographic synergy as described in claim 1, characterized in that: In step six, the division of multidimensional feature vectors adopts a stratified random sampling method; multiple feature combination schemes include spectral single-factor combination, spectral-topography joint combination, and spectral-process-topography fully coupled combination.

Images