A Rice Extraction Method Combining Time-Series Optical and SAR Images

Abstract

The present invention relates to a rice extraction method combining time - series optical and SAR images. First, the time - series data of the normalized difference vegetation index (NDVI) and the modified normalized difference water index (MNDWI) are reconstructed. Then, the characteristics of NDVI in the rice flooding - transplanting stage and the growth peak period, the characteristics of MNDWI in the rice flooding - transplanting stage, and the difference characteristics of the VH time - series in the rice flooding - transplanting stage and the growth peak period are quantified to construct a multi - feature rice index (MFRI). The corresponding threshold is determined according to the index statistical results of rice samples to accurately extract rice. When reconstructing the NDVI and MNDWI time - series, spectral features, temporal features, and morphological features are combined, and multi - modal long - time - series remote sensing data are fully utilized. Using this method, the rice planting distribution data of the target year can be obtained quickly and accurately, and the spatio - temporal dynamic changes of the rice planting distribution can be obtained, providing data support and reference for agricultural decision - making, and having important practical application significance.

Description

Technical Field

[0001] The present invention belongs to the technical field of classification and extraction of remote sensing images, and particularly relates to a method for extracting rice by combining time-series optical and SAR images. Background Art

[0002] Rice accounts for more than 9% of the global cultivated land area and is the staple food resource for more than half of the world's population, playing an important role in food security. In addition, rice is a major water consumer and one of the main sources of greenhouse gas emissions. Therefore, quickly and accurately identifying the planting distribution range of rice is of great significance for ensuring food security and evaluating environmental resources.

[0003] Currently, satellite data is the main data source for large-scale rice mapping. Sentinel-1 and Sentinel-2 data have high spatio-temporal resolution and are the main data sources for rice mapping research in recent years. In rice mapping, detection based on key phenological stages is a common method. By analyzing multiple spectral bands and spectral indices, the characteristics corresponding to rice can be screened out, and the rice planting distribution map of a large area can be quickly drawn. However, when classifying using the key phenological periods of rice, it depends on the accuracy of phenological information extraction, and there is a high overlap in the characteristics between rice and wetland vegetation. Therefore, most current studies are based on remote sensing time-series data and rice phenological characteristics, and machine learning or deep learning models are combined to solve the above problems. Although traditional machine learning and deep learning algorithms can quickly and effectively process large-scale data, the models require a large amount of labeled sample data for training, and obtaining rice labeled sample data requires a lot of time and effort. Although many current studies use transfer learning or automatically generate rice samples using phenological characteristics to reduce the demand for ground truth samples, such methods still have problems such as limited model generalization ability, insufficient representativeness of generated samples, and computational complexity.

[0004] The essence of the index method is to explore the commonalities of target objects, which can adapt to target recognition in different scenarios and can conduct mapping research on target objects more flexibly compared to hard classification. Currently, in the mapping research of various ground objects, a variety of indices have been successfully developed and applied. For the monitoring of rice, there are also optical rice indices and SAR-based rice indices, but they cannot avoid their own data defects. Therefore, there is an urgent need for a new method based on rice indices to achieve rapid and large-scale rice mapping. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method for extracting rice by combining time-series optical and SAR images.

[0006] This rice extraction method combining time - series optical and SAR images includes the following steps:

[0007] Step 1: Calculate the Normalized Difference Vegetation Index (NDVI) and the Modified Normalized Difference Water Index (MNDWI) corresponding to the surface reflectance data of the remote - sensing images, and construct the original time series of NDVI and MNDWI;

[0008] Step 2: Reconstruct the NDVI and MNDWI time series for each pixel to characterize the phenological time - series features of the pixel;

[0009] Step 3: Extract the minimum value P1 of NDVI during the rice flooding - transplanting stage, the maximum value P2 during the growth peak period, the maximum value P3 of MNDWI during the rice flooding - transplanting stage, the minimum value P4 of the VH time series during the rice flooding - transplanting stage, and the maximum value P5 during the growth peak period, and calculate the difference h = P5 - P4;

[0010] Step 4: Use f(n) to transform P1 to represent the similarity of each pixel to non - vegetation pixels; use f(v) to transform P2 to represent the similarity of each pixel to vegetation pixels; use f(w) to transform P3 to represent the similarity of each pixel to water pixels; use f(h) to transform the value range of h to 0 - 1, and construct the Multi - Feature Rice Index (MFRI) = f(n)×f(v)×f(w)×f(h);

[0011] Step 5: Determine the corresponding MFRI threshold according to the MFRI statistical results of the samples for extracting the rice planting area.

[0012] Preferably, in Step 2, the harmonic model is used to reconstruct the NDVI and MNDWI time series for each pixel, combining the time - series features, spectral features and spatial texture features.

[0013] Preferably, in Step 4, f(n) is a quadratic function of P1, which is used to transform the distance between the value of P1 and 0 into a value between 0 and 1, amplifying the difference between pixel points and evergreen vegetation.

[0014] Preferably, in Step 4, f(v) is a quadratic function of (1 - P2), which is used to transform the distance between P2 and 1 into a value between - 3 and 1, amplifying the difference between pixel points and non - vegetation.

[0015] Preferably, in Step 4, if P3 is greater than or equal to 0, it is determined that the pixel point is water or a feature with high humidity, and f(w)=1; if P3 is less than 0, f(w)=1 + P3. The closer the value of f(w) is to 1, the greater the possibility that the pixel point is water.

[0016] Preferably, in Step 4, f(h) uses trigonometric functions to transform h into a value between - 1 and 1.

[0017] The beneficial effects of the present invention are as follows:

[0018] 1) First, the present invention reconstructs the time series data of the Normalized Difference Vegetation Index (NDVI) and the Modified Normalized Difference Water Index (MNDWI), then quantifies the characteristics of NDVI in the rice flooding-transplanting stage and the growth peak period, the characteristics of MNDWI in the rice flooding-transplanting stage, and the difference characteristics of the VH time series in the rice flooding-transplanting stage and the growth peak period, constructs the Multi-Feature Rice Index (MFRI), and determines the corresponding threshold according to the index statistical results of the rice samples to accurately extract rice.

[0019] 2) When reconstructing the time series of NDVI and MNDWI, the present invention combines spectral characteristics, temporal characteristics, and morphological characteristics, makes full use of multi-modal long-time series remote sensing data. Using this method, the rice planting distribution data of the target year can be obtained quickly and accurately, and the spatio-temporal dynamic changes of the rice planting distribution can be obtained, providing data support and reference for agricultural decision-making, and having important practical application significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart for rice extraction combining time series optical and SAR images;

[0021] Figure 2 It is the original and reconstructed NDVI time series of rice pixels;

[0022] Figure 3 It is the original and reconstructed MNDWI time series of rice pixels;

[0023] Figure 4 It is a schematic diagram for the design of rice spectral index characteristics;

[0024] Figure 5 It is a schematic diagram of the rice VH backscattering characteristic coefficient;

[0025] Figure 6 It is a box plot of the MFRI values of rice and non-rice samples;

[0026] Figure 7 It is a schematic diagram of the rice extraction result. DETAILED DESCRIPTION OF THE INVENTION

[0027] The following further describes the present invention with reference to the embodiments. The description of the following embodiments is only for helping to understand the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several modifications can still be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0028] Embodiment 1

[0029] As an embodiment, in view of the problems existing in the existing rice extraction methods, such as using a single optical or SAR image, being unable to avoid the quality defects of its own data, and the low calculation efficiency of using fused images, a rice extraction method combining time-series optical and SAR images is proposed. This rice extraction method combining time-series optical and SAR images, as Figure 1 shown, first reconstructs the time-series data of the Normalized Difference Vegetation Index (NDVI) and the Modified Normalized Difference Water Index (MNDWI), then amplifies the differences between rice and other ground objects by quantifying the characteristics of the rice flooding-transplanting stage and the growth peak stage, and determines the corresponding threshold according to the index statistical results of the samples to accurately extract rice, specifically including the following steps:

[0030] Step 1: Use a screening module to screen all available Sentinel-1 and Sentinel-2 remote sensing images of the target area in the target year, and preprocess all remote sensing images;

[0031] Step 2: The spectral index calculation module calculates the Normalized Difference Vegetation Index (NDVI) image corresponding to all Sentinel-2 remote sensing images and the Modified Normalized Difference Water Index (MNDWI), and constructs the original time series of NDVI and MNDWI;

[0032] Step 3: The reconstruction module uses the harmonic model to reconstruct the NDVI and MNDWI time series of each pixel to obtain the reconstructed time-series data, which are respectively used to characterize the phenological time-series characteristics of the pixel.

[0033] Step 4: Extract the minimum value P1 of NDVI in the rice flooding-transplanting stage, the maximum value P2 in the growth peak stage, the maximum value P3 of MNDWI in the rice flooding-transplanting stage, the minimum value P4 of the VH time series in the rice flooding-transplanting stage, and the maximum value P5 in the growth peak stage, and calculate the difference h = P5 - P4;

[0034] Step 5: Use f(n) to standardize the value range expression of P1, f(v) to standardize the value range expression of P2, f(w) to standardize the value range expression of P3, and f(h) to standardize the value range expression of h. The Multi-Feature Rice Index (MFRI) construction module constructs the multi-feature rice index MFRI = f(n) × f(v) × f(w) × f(h);

[0035] Step 6: The extraction module determines the corresponding threshold according to the MFRI statistical results of the samples and extracts the rice planting area, as Figure 7 shown.

[0036] Embodiment 2

[0037] As another embodiment, this second embodiment is proposed based on the first embodiment. A more specific rice extraction method combining time-series optical and SAR images. In step 5, the calculation formulas for f(n), f(v), f(w), f(h), and MFRI are as follows:

[0038] In f(n), the distance between P1 and 0 is transformed into a value range from 0 to 1 using a quadratic function. Square operation is performed on P1 to amplify the difference from evergreen vegetation.

[0039] f(n) = 1 - P1 2

[0040] In f(v), the distance between p2 and 1 is transformed into a value range from -3 to 1 using a quadratic function. Square operation is performed on (1 - P2) to amplify the difference from non-vegetation.

[0041] f(v) = 1 - (1 - P2) 2

[0042] In f(w), if P3 is greater than or equal to 0, it is considered as water body or features with high humidity, and 0 is substituted; if P3 is less than 0, considering the influence of data quality or fitting smoothness, the closer it is to 0, the more likely it is a water body, so the original value is substituted.

[0043]

[0044] In f(h), the dynamic change of VH backscattering is transformed into a value range from -1 to 1 using a trigonometric function.

[0045]

[0046] Each feature meets the requirements in MFRI expressed by multiplication.

[0047] MFRI = f(n) × f(v) × f(w) × f(h)

[0048] In step 6, the expression that needs to be satisfied for rice extraction is:

[0049] MFRI(x,y) ≥ TH MFRI

[0050] Wherein, MFRI(x,y) is the MFRI value at position (x,y), and TH MFRI is the threshold of MFRI. As Figure 6 shown, TH MFRI is 0.57 in this embodiment.

[0051] In summary, the present invention combines time - series optical and SAR images, and by quantifying the key phenological stage characteristics of rice, proposes an index MFRI for rice planting area extraction. Using this index, a large - scale rice distribution dataset for the target year can be quickly obtained, which helps to grasp the spatio - temporal dynamics of rice planting and provides data support and reference for agricultural decision - making.

[0052] It should be noted that the parts that are the same as or similar to those in Embodiment 1 in this embodiment can be referred to each other and will not be elaborated in this application.

[0053] Embodiment 3

[0054] As another embodiment, this Embodiment 3 is proposed on the basis of Embodiments 1 and 2. A more specific method for rice extraction by combining time - series optical and SAR images is implemented based on Sentinel - 1 and Sentinel - 2 data as data sources and relying on the Google Earth Engine (GEE) platform.

[0055] A computer storage medium and a computer program product are provided. The computer storage medium stores a computer program. When the computer program runs on a computer, the computer is made to execute this rice extraction method.

[0056] In step 1, all available Sentinel - 1 and Sentinel - 2 remote sensing images of the target area in the target year are screened, and the remote sensing images are pre - processed. The Sentinel - 1 remote sensing image is radar - band data, and the Lee - sigma filter with a 5×5 window size is used for speckle noise reduction. The Sentinel - 2 remote sensing image is surface reflectance data, and the pre - processing includes atmospheric correction and cloud removal.

[0057] In step 2, the calculation formulas of NDVI and MNDWI are respectively:

[0058]

[0059] where NIR is the near - infrared band value, Red is the red - light band value, Green is the green - light band value, and SWIR1 is the value of the first band of short - wave infrared;

[0060] In this embodiment, NIR, Red, Green, and SWIR1 are respectively B8, B4, B3, and B11 of Sentinel - 2 data.

[0061] In step 3, as Figure 2 and Figure 3 shown, the original time series of NDVI and MNDWI are reconstructed by using 3 - cycle harmonic fitting. The calculation formula for reconstructing the NDVI and MNDWI time series of each pixel is:

[0062]

[0063] Among them, is the fitting value on the t-th day, where t is the Julian day, a is the overall level offset value, b is the linear coefficient, c n and d n are the sine and cosine function coefficients of different frequencies, and T is the number of days in a year.

[0064] Step 4, as Figure Four and Figure Five shown, design the characteristics of the key phenological periods of rice using the reconstructed NDVI and MNDWI original time series and the denoised VH time series.

[0065] It should be noted that the parts that are the same or similar to those in Embodiment 1 and Embodiment 2 in this embodiment can be referred to each other and will not be described in detail in this application.

[0066] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

Claims

1. A rice extraction method combining time series optical and SAR images, characterized in that: The following steps are involved: Step 1: Calculate the normalized vegetation index NDVI and the improved normalized water index MNDWI corresponding to the surface reflectance data of the remote sensing image, and construct the original time series of NDVI and MNDWI; Step 2: Reconstruct the NDVI and MNDWI time series of each pixel to characterize the phenological temporal characteristics of the pixel; Step 3, extract the minimum value P1 of NDVI in the rice flooding-transplanting stage, the maximum value P2 in the growth peak period, and the maximum value P3 of MNDWI in the rice flooding-transplanting stage, extract the minimum value P4 of VH time series in the rice flooding-transplanting stage and the maximum value P5 in the growth peak period, and calculate the difference h=P5-P4; Step 4: transform P1 with f(n) to represent the similarity between each pixel and non-vegetation pixels; transform P2 with f(v) to represent the similarity between each pixel and vegetation pixels; transform P3 with f(w) to represent the similarity between each pixel and water pixels; transform the value range of h into 0-1 with f(h) to construct the multi-feature rice index MFRI = f(n)×f(v)×f(w)×f(h); f(n) is a quadratic function of P1, which is used to convert the distance between the value of P1 and 0 into a value between 0 and 1, thereby amplifying the difference between the pixel and the evergreen vegetation; f(v) is a quadratic function of (1-P2), which is used to convert the distance between P2 and 1 into a value between -3 and 1, thereby amplifying the difference between the pixel and non-vegetation. If P3 is greater than or equal to 0, the pixel is considered to be a water body or a land feature with high humidity, and f(w) = 1; if P3 is less than 0, f(w) = 1 + P3, and the closer the value of f(w) is to 1, the greater the possibility that the pixel is a water body; Step 5: Determine the corresponding MFRI threshold according to the MFRI statistical results of the sample to extract the rice planting range.

2. The rice extraction method combining time series optical and SAR images according to claim 1, characterized in that: In step 2, the harmonic model is used to reconstruct the NDVI and MNDWI time series of each pixel, combining the temporal features, spectral features, and spatial texture features.

3. The rice extraction method combining time series optical and SAR images according to claim 1, characterized in that: In step 4, f(h) uses trigonometric functions to convert h into a value between -1 and 1.

Images