Saline-alkali soil distribution identification method based on remote sensing cloud computing

Abstract

The invention discloses a saline-alkali soil distribution identification method based on remote sensing cloud computing. The method comprises the following steps: acquiring soil sample data and remote sensing image data of saline-alkali soil; segmenting the remote sensing image data based on the plot sample, and aggregating segmentation results to obtain an aggregated plot-level image; constructing a feature pool based on the waveband data of the remote sensing image data and the environmental factors; training a random forest regression model based on the soil sample data in combination with the feature pool to obtain an identification model; and performing saline-alkali soil distribution identification by using the identification model. According to the method, before the random forest regression model is trained, the soil salinity and alkalinity sample is transformed through power transformation, so that the problem of deflection distribution is solved; according to the invention, a more complete characteristic pool is constructed, environmental factors more related to soil salinization are incorporated, and the monitoring precision is further improved; compared with a pixel-level method, the land parcel level-based saline-alkali land distribution identification method is shorter in time consumption and less in limitation.

Description

Technical Field

[0001] The present invention relates to the technical field of saline-alkali land distribution identification, and in particular to a saline-alkali land distribution identification method based on remote sensing cloud computing. Background Art

[0002] Soil salinization is a global land degradation process that affects soil fertility, soil structure and biodiversity. Existing saline-alkali land monitoring mostly relies on measured samples from different research institutions. The distribution of samples is uneven and has a significant skewed distribution. The accuracy of saline-alkali land monitoring needs to be improved. Existing saline-alkali land distribution identification methods mainly use a large number of ground samples, mostly from different research institutions around the world, combined with satellite data, remote sensing indices, meteorological soil and other environmental factors, based on machine learning or deep learning methods, to train and deploy models at the pixel level to obtain the spatial distribution of saline-alkali land. The existing saline-alkali land monitoring method has uneven sample distribution, which affects the accuracy of the model. Specifically, there are the following problems:

[0003] 1) Severely skewed sample distribution: Global soil salinization is primarily distributed in arid and semi-arid regions with high temperatures and low precipitation. Soil samples from different research institutions are collected over a wide range of areas, not just for soil salinity and alkalinity. Therefore, after sample integration, there is a serious skewed distribution, that is, the salinity and alkalinity of most soil samples are close to zero, while only a few samples have high salinity and alkalinity values.

[0004] 2) Incomplete feature pool: Soil salinization is closely related to soil water movement. The feature pool of existing methods only includes basic satellite band data and band combinations as well as common meteorological factors such as precipitation and temperature.

[0005] 3) Dependence on computer configuration: Existing methods are mostly pixel-level models that rely on a large amount of computer memory to run locally and consume a lot of time. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a method for identifying saline-alkali land distribution based on remote sensing cloud computing, comprising:

[0007] Collect soil sample data and remote sensing image data of saline-alkali land;

[0008] Segmenting the remote sensing image data based on the plot samples, and aggregating the segmentation results to obtain aggregated plot-level images;

[0009] Construct a feature pool based on the band data and environmental factors of remote sensing image data;

[0010] Transforming the soil sample data, and using the transformed soil sample data and the feature pool to train a random forest regression model to obtain a recognition model;

[0011] The plot-level image is input into the recognition model to obtain a saline-alkali land distribution recognition result.

[0012] Preferably, before segmenting the remote sensing image data, the method further includes:

[0013] The remote sensing image data is preprocessed to obtain time series image data; wherein the preprocessing includes: cloud removal, spatial filtering and noise removal, time series synthesis, linear interpolation, and remote sensing index calculation.

[0014] Preferably, segmenting the remote sensing image data based on land parcel samples includes:

[0015] Determining a segmentation size of the time series image data based on a land parcel sample;

[0016] According to the segmentation size, an image segmentation algorithm is used to perform image segmentation on the time series image data to generate segmented image data.

[0017] Preferably, aggregating the segmentation results includes:

[0018] Based on the cluster band, the segmented image data is aggregated to obtain the aggregated plot-level image.

[0019] Preferably, the band data of the remote sensing image data includes: original band data and band combination data;

[0020] The original band data includes 13 spectral bands of remote sensing images;

[0021] The band combination data is generated by mathematically combining the original band data;

[0022] The environmental factors include meteorological factors, soil factors, and topographic factors.

[0023] Preferably, transforming the soil sample data includes:

[0024] Performing power transformation on the soil sample data to obtain soil salinity and alkalinity sample values;

[0025] Based on the soil salinity and alkalinity sample values, a soil sample dataset was constructed.

[0026] Preferably, training the random forest regression model includes:

[0027] Integrating the transformed soil sample and the feature pool as input of a random forest model;

[0028] The hyperparameters in the random forest regression model are tuned by a grid search method to obtain the final recognition model.

[0029] Preferably, inputting the plot-level image into the recognition model to obtain the saline-alkali land distribution recognition result includes:

[0030] Inputting the plot-level image into the trained random forest regression model to obtain the initial saline-alkali land distribution result;

[0031] The initial saline-alkali land distribution result is inversely transformed to obtain a plot-level saline-alkali land distribution result.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] This paper proposes a method for identifying saline-alkali land distribution based on remote sensing cloud computing. The method first collects soil sample data and remote sensing image data. The remote sensing image data is then segmented based on plot samples, and the segmentation results are aggregated to obtain aggregated plot-level images. A feature pool is then constructed. A random forest regression model is trained based on the aggregated plot-level images, soil sample data, and feature pool to obtain a recognition model. Finally, the recognition model is used to identify saline-alkali land distribution. This method addresses the problem of skewed sample distribution and identifies saline-alkali land distribution at the plot level. By integrating with a cloud computing platform, the method overcomes the limitations of local memory. Before training the random forest regression model, the method also transforms soil salinity samples using a power transformation, resolving the skewed distribution issue. The transformed samples are closer to a normal distribution, improving the accuracy of the random forest model. The method constructs a more complete feature pool, incorporating environmental factors more relevant to soil salinization, including wet day frequency, runoff, vapor pressure, surface shortwave downward radiation, vapor pressure difference, soil evaporation, and transpiration, further improving monitoring accuracy. This method for identifying saline-alkali land distribution based on the plot level is less time-consuming and less restrictive than pixel-level methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0035] Figure 1 The present invention is a flowchart of a method for identifying saline-alkali land distribution based on remote sensing cloud computing. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] like Figure 1 As shown, this embodiment proposes a method for identifying saline-alkali land distribution based on remote sensing cloud computing, including:

[0039] Collect soil sample data and remote sensing image data of saline-alkali land;

[0040] Segmenting the remote sensing image data based on the plot samples, and aggregating the segmentation results to obtain aggregated plot-level images;

[0041] Construct a feature pool based on the band data and environmental factors of remote sensing image data;

[0042] Transforming the soil sample data, and using the transformed soil sample data and the feature pool to train a random forest regression model to obtain a recognition model;

[0043] The plot-level image is input into the recognition model to obtain a saline-alkali land distribution recognition result.

[0044] Specifically, in this embodiment, the remote sensing image data of the saline-alkali land collected is the Sentinel-2 remote sensing image of the corresponding time period acquired on the Google Earth Engine platform;

[0045] Furthermore, before segmenting the remote sensing image data, the method further includes: pre-processing the remote sensing image data to obtain time series image data.

[0046] Specifically, in this embodiment, a series of preprocessing is performed on the image, including cloud removal, spatial filtering and noise removal, time series synthesis, linear interpolation, remote sensing index calculation and other operations to obtain reconstructed time series Sentinel-2 data.

[0047] Specifically, in this embodiment, the land parcel samples can determine the size standard of the image segmentation. The land parcel samples can be obtained through ground surveys or manually outlined in Google Earth Engine based on color composite images of crop growth stages.

[0048] Furthermore, the remote sensing image data is segmented based on the land samples, including:

[0049] Determine the segmentation size of time series image data based on plot samples;

[0050] According to the determined segmentation size, the image segmentation algorithm is used to segment the time series image data to generate segmented image data.

[0051] Specifically, in this embodiment, segmenting remote sensing image data based on land samples includes the following steps:

[0052] Based on the sample plots, the variability of different indices within the plots was calculated and represented by the standard deviation of these indices. The standard deviation of the internal indices of the image segmentation objects was calculated at different segmentation scales, and visual inspection was used to ultimately determine the optimal segmentation size. The specific segmentation scale involved can be changed based on factors such as the size, shape, and planting habits of the crop plots in different regions.

[0053] Combined with the SNIC (simple non-iterative clustering) image segmentation algorithm embedded in the Google Earth Engine remote sensing cloud computing platform, the constructed time series remote sensing image data is segmented. The original image is segmented into spatially meaningful plot-level units based on the spatial relationship between pixels.

[0054] Furthermore, the segmentation results are aggregated to obtain aggregated plot-level images, including: aggregating the segmented image data based on cluster bands to obtain aggregated plot-level images.

[0055] Specifically, in this embodiment, the "cluster" band is used for image aggregation, and the remote sensing images are further aggregated based on the cluster numbers to extract the statistical characteristics of the pixels within each plot unit, and to construct plot-level image data with stronger spatial consistency. After aggregation, the pixels within the plot have similar spatial, temporal, and spectral characteristics, and finally the aggregated plot-level image is obtained.

[0056] After using the SNIC image segmentation algorithm, the value of each pixel in the cluster band indicates which segmentation block it belongs to, that is, the cluster band; the SNIC algorithm in GEE will generate a new image, and the most important part of it is the cluster band;

[0057] Furthermore, the band data of the remote sensing image data includes: original band data and band combination data; the environmental factors include meteorological factors, soil factors, and terrain factors;

[0058] The original band data includes 13 spectral bands of remote sensing images;

[0059] Band combination data is generated by mathematically combining the original band data.

[0060] Specifically, in this embodiment,

[0061] Environmental factors include meteorological factors, soil factors, and topographic factors; including wet day frequency, runoff, vapor pressure, downward shortwave radiation from the surface, vapor pressure difference, soil evaporation and transpiration.

[0062] The band combination data is new data generated by mathematically combining the original band data, such as NDVI, EVI, etc.

[0063] Furthermore, before training the random forest regression model, the following steps are also included: performing power transformation on the soil sample data to obtain soil salinity and alkalinity sample values; and constructing a soil sample data set based on the soil salinity and alkalinity sample values.

[0064] Specifically, in this embodiment, the collected and integrated soil sample data are subjected to Box-Cox transformation to obtain transformed soil salinity and alkalinity sample values. The Box-Cox transformation makes the transformed samples closer to a normal distribution.

[0065] Furthermore, a feature pool is constructed based on the band data and environmental factors of remote sensing image data;

[0066] Specifically, in this embodiment, the original bands and band combinations of Sentinel-2 images, as well as environmental factors such as meteorology, soil, and topography from other remote sensing and reanalysis datasets are integrated to form a feature pool for model training. Prior knowledge and expert knowledge are combined to incorporate features related to the mechanism of soil salinization.

[0067] Furthermore, a random forest regression model was trained, and the transformed soil samples and feature pool were integrated as the input of the random forest model; the hyperparameters in the random forest regression model were tuned through the grid search method to obtain the final recognition model.

[0068] Specifically, in this embodiment, 75% of the transformed soil samples were extracted as a training set and 25% as a test set, and the random forest regression model was trained and tested in combination with the feature pool;

[0069] The transformed soil samples and feature pools were integrated as inputs to the random forest model. Two hyperparameters in the random forest model (the number of trees and the size of the random feature subset) were optimized using a grid search method. The search was stopped when the model accuracy reached the highest, and the optimal hyperparameters were obtained. The random forest model with these two optimal hyperparameters was the final recognition model.

[0070] Furthermore, the plot-level image is input into the recognition model to obtain a saline-alkali land distribution recognition result, including:

[0071] The plot-level images were input into the trained random forest regression model to obtain the initial saline-alkali land distribution results;

[0072] The initial saline-alkali land distribution results were inversely transformed to obtain the saline-alkali land distribution results at the plot level.

[0073] Specifically, in this embodiment, the trained random forest model is deployed to identify the distribution of saline-alkali land at the plot level. The Box-Cox inverse transformation is applied to the output of the random forest regression model to obtain the saline-alkali land distribution results at the plot level. Finally, the accuracy indicators root mean square error (RMSE) and determination coefficient (R 2 ).

[0074] This embodiment breaks away from the limitations of local computer memory and completes image segmentation and random forest regression model deployment on the Google Earth Engine remote sensing cloud computing platform. In addition, the block-level recognition method takes less time than the pixel-level method and is closer to the ground truth.

[0075] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for identifying saline-alkali land distribution based on remote sensing cloud computing, characterized in that: include: Collect soil sample data and remote sensing image data of saline-alkali land; Segmenting the remote sensing image data based on the plot samples, and aggregating the segmentation results to obtain aggregated plot-level images; Construct a feature pool based on the band data and environmental factors of remote sensing image data; Transforming the soil sample data, and using the transformed soil sample data and the feature pool to train a random forest regression model to obtain a recognition model; The plot-level image is input into the recognition model to obtain a saline-alkali land distribution recognition result.

2. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 1, characterized in that: Before segmenting the remote sensing image data, the method further includes: The remote sensing image data is preprocessed to obtain time series image data; wherein the preprocessing includes: cloud removal, spatial filtering and noise removal, time series synthesis, linear interpolation, and remote sensing index calculation.

3. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 2, characterized in that: Segmenting the remote sensing image data based on the land parcel samples includes: Determining a segmentation size of the time series image data based on a land parcel sample; According to the segmentation size, an image segmentation algorithm is used to perform image segmentation on the time series image data to generate segmented image data.

4. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 3, characterized in that: Aggregation of segmentation results includes: Based on the cluster band, the segmented image data is aggregated to obtain the aggregated plot-level image.

5. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 1, characterized in that: The band data of the remote sensing image data includes: original band data and band combination data; The original band data includes 13 spectral bands of remote sensing images; The band combination data is generated by mathematically combining the original band data; The environmental factors include meteorological factors, soil factors, and topographic factors.

6. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 1, characterized in that: Transforming the soil sample data includes: Performing power transformation on the soil sample data to obtain soil salinity and alkalinity sample values; Based on the soil salinity and alkalinity sample values, a soil sample dataset was constructed.

7. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 1, characterized in that: Training a random forest regression model involves: Integrating the transformed soil sample and the feature pool as input of a random forest model; The hyperparameters in the random forest regression model are tuned by a grid search method to obtain the final recognition model.

8. The method for identifying saline-alkali land distribution based on remote sensing cloud computing according to claim 1, characterized in that: Inputting the plot-level image into the recognition model to obtain saline-alkali land distribution recognition results includes: Inputting the plot-level image into the trained random forest regression model to obtain the initial saline-alkali land distribution result; The initial saline-alkali land distribution result is inversely transformed to obtain a plot-level saline-alkali land distribution result.

Images