A method for extracting samples based on rice phenological characteristics

Abstract

The present invention discloses a method for extracting samples based on rice phenological characteristics. The method comprises the following steps: S1. calculating the rice phenological characteristic index r at each grid position based on remote sensing ecological index time series data of multi-temporal remote sensing images; S2. extracting training samples; S21. determining rice discrimination thresholds θ1 and θ2 to discriminate whether the grid position is rice; S22. stretching the rice phenological characteristic index r based on the rice discrimination threshold θ1; S23. determining a sampling interval based on the stretched rice phenological characteristic index r'; S24. determining the number n of sample points to be collected, and selecting the sample points when the r' value belongs to the interval (0.25-k, 0.25+k) and the LSWI is min In grids with values ​​less than θ2, n / 2 rice sample points are randomly selected, and in grids with r' values ​​in the interval (0.75-k, 0.75+k), n / 2 non-rice sample points are randomly selected. Together, these constitute a training sample set that can be used for machine learning. The training samples extracted using this method do not require manual field work, significantly reducing sampling costs.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image recognition, and specifically relates to a method for extracting samples based on rice phenological characteristics. The obtained samples can be used to train a machine learning model to extract the spatial distribution of rice. Background Art

[0002] Rice is one of my country's most important food crops, and understanding its spatial distribution is crucial for agricultural management and food security. Among existing methods for extracting rice spatial distribution, the most common approach is to use machine learning to mine the spectral signatures of rice fields based on remote sensing imagery and rice field samples to infer their spatial distribution. While these methods typically achieve higher inference accuracy than other approaches, they require training the machine learning model with a large number of samples, which typically rely on manual field collection, requiring significant human, material, and financial resources. Summary of the Invention

[0003] The purpose of this patent is to provide a method for extracting samples based on rice phenological characteristics. This method addresses the high cost of field sample collection, which is often required for machine learning methods to extract rice spatial distribution from remote sensing imagery. The method utilizes rice phenological characteristics to extract training samples from remote sensing imagery, enabling low-cost, high-precision estimation of the spatial distribution of rice fields. Phenology refers to the cyclical and seasonal changes in crop growth cycles. Rice exhibits unique phenological characteristics that distinguish it from other crops. For example, during the vegetative growth phase, rice fields appear as a mixture of water, rice, and soil, with vegetation cover rapidly increasing. Based on these phenological characteristics, the probability of each grid location belonging to a rice field can be determined using vegetation and water indexes calculated from remote sensing imagery, allowing for the collection of rice and non-rice samples.

[0004] The above purpose is achieved through the following technical solutions:

[0005] A method for extracting samples based on rice phenological characteristics, the method comprising the following steps:

[0006] S1. Construction of Rice Phenological Characteristic Index

[0007] S11. Obtain multi-temporal remote sensing images of the area to be inferred during the rice growing season;

[0008] S12. Calculate time series data of remote sensing ecological indices at each grid based on multi-temporal remote sensing images. The remote sensing ecological indices to be calculated include the Enhanced Vegetation Index (EVI2) and the Land Surface Water Index (LSWI).

[0009] S13. Use Savitzky-Golay filter to reconstruct the time series of remote sensing ecological indices at each grid;

[0010] S14. Determine the tillering and heading dates of rice using the reconstructed Enhanced Vegetation Index (EVI2) time series data;

[0011] S15. Calculate the rice phenological characteristic index r at each grid using the reconstructed EVI2 and LSWI time series analysis data:

[0012] S2. Training sample extraction

[0013] S21 determines the rice discrimination thresholds θ1 and θ2 to determine whether the grid position is rice;

[0014] S22. Based on the rice discrimination threshold θ1, the rice phenological characteristic index r is histogram stretched, and the part less than θ1 is stretched to 0-0.5, and the part greater than θ1 is stretched to 0.5-1

[0015] S23. Determine the sampling interval based on the stretched rice phenological characteristic index r'. Grids with r' near 0.5 are more likely to be misclassified and should not be used as sampling areas. Grids with r' near 1 and 0 are too typical and should not be used as sampling areas. A k value should be determined between (0 and 0.25) so that the r' value of candidate grids for rice sampling sites varies between (0.25-k and 0.25+k), and the r' value of candidate grids for non-rice sampling sites varies between (0.75-k and 0.75+k).

[0016] S24. Determine the number of sample points n to be collected, according to the sampling interval obtained in step S23, when the r' value belongs to the interval (0.25-k, 0.25+k) and LSWI min In the grids with r' values ​​less than θ2, n / 2 rice sample points are randomly selected, and in the grids with r' values ​​in the interval (0.75-k, 0.75+k), n / 2 non-rice sample points are randomly selected to form a training sample set that can be used for machine learning.

[0017] Furthermore, the calculation formulas for the Enhanced Vegetation Index EVI2 and the Land Surface Water Index LSWI in step S12 are as follows:

[0018]

[0019] Among them, NIR represents the reflectivity of the near-infrared band, R represents the reflectivity of the visible red band, SWIR represents the reflectivity of the short-wave infrared band, and the value of EVI2 is between -1 and 1, and its value is positively correlated with vegetation coverage; the value of LSWI is also between -1 and 1. The higher the value, the higher the soil moisture and vegetation water content.

[0020] Furthermore, in step S13, the Savitzky-Golay filter is used to reconstruct the remote sensing ecological index time series on each grid, that is, a p-order polynomial fitting is performed on the data points within a certain length window, and the formula is as follows:

[0021]

[0022] X p,smooth is the result after fitting, w represents the moving window size, and p represents the number of polynomial fitting times.

[0023] Furthermore, in step S15, the reconstructed EVI2 and LSWI time series analysis data are used to calculate the rice phenological characteristic index r on each grid, and the formula is as follows:

[0024]

[0025] LSWI max with LSWI min is the maximum and minimum value of LSWI index from tillering stage to heading stage, EVI heading With EVI tillering It is the value of EVI2 index at the tillering stage and heading stage; the r value varies from 0 to +∞. The lower the r value, the greater the possibility that the grid is a rice field, and vice versa.

[0026] Furthermore, the histogram in step S22 is stretched as follows:

[0027]

[0028] Among them, r i is the rice phenological characteristic index at grid position i, r i ′ is the rice phenological characteristic index after stretching at grid position i, The rice phenological characteristic index is less than r i The number of grids, is the number of grids whose rice phenological characteristic index is less than θ1, is the rice phenological characteristic index between θ1 and r i The number of grids between is the number of grids whose rice phenological characteristic index is greater than θ1.

[0029] Beneficial effects: The present invention has the following advantages:

[0030] This paper addresses the problem that remote sensing rice extraction using machine learning relies on a large number of training samples, while field sample collection is prohibitively expensive. A method for extracting training samples from remote sensing images based on rice phenology is designed. The training samples extracted using this method eliminate the need for field labor, significantly reducing sampling costs. Furthermore, these samples ensure the prediction accuracy of machine learning methods, enabling low-cost, high-efficiency, and high-precision remote sensing extraction of rice spatial distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flowchart of extracting rice phenological features for machine learning training samples;

[0032] Figure 2 This is the distribution map of the revised rice phenological characteristic index;

[0033] Figure 3 These are rice sample points and non-rice sample points collected based on the present invention. DETAILED DESCRIPTION

[0034] Below is Figure 1 The flowchart shown in FIG. 1 illustrates a specific embodiment of the present invention by taking the extraction of rice samples in Jin'an District, Lu'an City, Anhui Province in 2017 as an example:

[0035] Step 1. Construction of rice phenological characteristic index

[0036] Step 11: Rice cultivation in Jin'an District is mainly mid-season rice, which is raised in early May and harvested before mid-October. Therefore, this paper obtained 24 Sentinel-2A (Sentinel 2) multispectral remote sensing images covering the study area from early May to the end of October 2017, of which 17 were light cloud images with cloud cover below 20%. These images were used as the source of remote sensing ecological index time series data.

[0037] Step 12: Calculate the EVI2 and LSWI indices for each of the 17 images to obtain the remote sensing ecological index time series data for each grid.

[0038] Step 13: Use the SG filter to reconstruct the remote sensing ecological index series on each grid to obtain the smooth curves of the EVI2 and LSWI indices.

[0039] Step 14: For each grid, determine the date when the EVI2 index curve reaches its maximum value, which is used as the heading date of the rice in the grid. The tillering date in the grid is the 40th day before the heading date.

[0040] Step 15: Calculate the rice phenological characteristic index r at each grid based on the tillering date, heading date, and the reconstructed EVI2 and LSWI curves at each grid.

[0041] Step 2. Training sample extraction

[0042] Step 21: Determine the rice discrimination threshold θ1 as 0.9 and θ2 as 0.1. When r is less than 0.9 and LSWI min When θ1 is greater than 0.1, the grid location is identified as rice, otherwise it is non-rice. The value of θ1 varies with the study area and the remote sensing image used, and is generally between 0.6 and 0.9.

[0043] Step 22: Based on the rice discrimination threshold θ1, the histogram of the rice phenological characteristic index r is stretched. The part with r value less than 0.9 is stretched to 0-0.5, and the part with r value greater than 0.9 is stretched to 0.5-1. The spatial distribution of r' value after stretching is as follows: Figure 2 shown.

[0044] Step 23: Determine the sampling interval and set the k value to 0.05, so that the r' value of the candidate grid of the rice sampling point is between 0.2-0.3, and the r' value of the candidate grid of the non-rice sampling point is between 0.7-0.8.

[0045] Step 24: Determine the number of samples to be collected as 1000, randomly select 500 locations in the rice sample point candidate grid as rice sample points, and randomly select 500 locations in the non-rice sample point grid as non-rice sample points, together forming a 1000 training sample set that can be used for machine learning ( Figure 3 ).

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that, for those skilled in the art, without departing from the principles of the present invention, several improvements and variations may be made, and such improvements and variations shall also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting samples based on rice phenological characteristics, characterized in that: The method comprises the following steps: S1. Construction of rice phenological characteristic index; S11. Obtain multi-temporal remote sensing images of the area to be inferred during the rice growing season; S12. Calculate time series data of remote sensing ecological indices at each grid based on multi-temporal remote sensing images. The remote sensing ecological indices to be calculated include the Enhanced Vegetation Index (EVI2) and the Land Surface Water Index (LSWI). S13. Use Savitzky-Golay filter to reconstruct the time series of remote sensing ecological indices at each grid; S14. Determine the tillering and heading dates of rice using the reconstructed Enhanced Vegetation Index (EVI2) time series data; S15. Calculate the rice phenological characteristic index r at each grid using the reconstructed EVI2 and LSWI time series analysis data: S2. Training sample extraction; S21 determines the rice discrimination thresholds θ1 and θ2 to determine whether the grid position is rice; S22. Based on the rice discrimination threshold θ1, the rice phenological characteristic index r is histogram stretched, and the part less than θ1 is stretched to 0-0.5, and the part greater than θ1 is stretched to 0.5-1; S23. Determine the sampling interval based on the stretched rice phenological characteristic index r'. Grids with r' near 0.5 are likely to be misclassified and should not be used as sampling areas. Grids with r' near 1 and 0 are too typical and should not be used as sampling areas. A k value should be determined between (0 and 0.25) so that the r' value for candidate rice sampling points ranges from (0.25 - k to 0.25 + k), and the r' value for candidate non-rice sampling points ranges from (0.75 - k to 0.75 + k). S24. Determine the number of sample points n to be collected, according to the sampling interval obtained in step S23, when the r' value belongs to the interval (0.25-k, 0.25+k) and LSWI min In the grids with r' values ​​less than θ2, n / 2 rice sample points are randomly selected, and in the grids with r' values ​​in the interval (0.75-k, 0.75+k), n / 2 non-rice sample points are randomly selected to form a training sample set that can be used for machine learning.

2. The method for extracting samples based on rice phenological characteristics according to claim 1, characterized in that: The calculation formulas for the enhanced vegetation index EVI2 and the land surface water index LSWI in step S12 are as follows: Among them, NIR represents the reflectivity of the near-infrared band, R represents the reflectivity of the visible red band, SWIR represents the reflectivity of the short-wave infrared band, and the value of EVI2 is between -1 and 1, and its value is positively correlated with vegetation coverage; the value of LSWI is also between -1 and 1. The higher the value, the higher the soil moisture and vegetation water content.

3. The method for extracting samples based on rice phenological characteristics according to claim 1, characterized in that: In step S13, the Savitzky-Golay filter is used to reconstruct the remote sensing ecological index time series on each grid, that is, a p-order polynomial fitting is performed on the data points within a certain length window. The formula is as follows: X p，smooth is the result after fitting, w represents the moving window size, and p represents the number of polynomial fitting times.

4. The method for extracting samples based on rice phenological characteristics according to claim 1, characterized in that: In step S15, the reconstructed EVI2 and LSWI time series analysis data are used to calculate the rice phenological characteristic index r on each grid. The formula is as follows: LSWI max with LSWI min is the maximum and minimum value of LSWI index from tillering stage to heading stage, EVI heading With EVI tillering It is the value of EVI2 index at the tillering stage and heading stage; the r value varies from 0 to +∞. The lower the r value, the greater the possibility that the grid is a rice field, and vice versa.

5. The method for extracting samples based on rice phenological characteristics according to claim 1, characterized in that: The histogram stretching in step S22 is as follows: Among them, r i is the rice phenological characteristic index at grid position i, r i ′ is the rice phenological characteristic index after stretching at grid position i, The rice phenological characteristic index is less than r i The number of grids, is the number of grids whose rice phenological characteristic index is less than θ1, is the rice phenological characteristic index between θ1 and r i The number of grids between is the number of grids whose rice phenological characteristic index is greater than θ1.

Images