Large-scale soybean and corn mapping method based on agronomic knowledge driving index

Abstract

The invention relates to a large-scale soybean and corn mapping method based on an agronomic knowledge driving index. The invention relates to the large-scale soybean and corn mapping method based on the agronomic knowledge driving index. According to the invention, the problem of lack of efficient and concise crop identification indexes in the prior art is solved. The method comprises the following steps: 1, calculating GWCCI and CARI of a multispectral image; 2, constructing a new corn recognition index GC; and 3, identifying and extracting the pixel values of the multispectral remote sensing image pixels, wherein the pixel values of the multispectral remote sensing image pixels are three defined threshold values, namely # imgabs0 #, # imgabs1 #, # imgabs2 # and # imgabs3 #. According to the method, the blank of corn identification indexes is filled, and meanwhile, the capability of filtering out other crops by the green degree moisture comprehensive index is further improved. The GGC model provided by the invention has strong space-time generalization ability, and does not need to depend on training samples and long-time sequence data. The agricultural knowledge-driven soybean and corn classification method provides a new implementation way for large-scale crop drawing.

Description

Technical Field

[0002] The present invention relates to the field of remote sensing technology, and in particular to a large-scale soybean and corn mapping method based on an agronomy knowledge-driven index. Background Art

[0004] Relying on the rapid development of computer technology, especially classification technology based on image processing, classification algorithms based on remote sensing data have gradually developed. Different from the coarse classification types such as surface cover, more detailed crop distribution information requires more indicative identification features, so efficient crop classification algorithms face more challenges. The huge heterogeneity of global crop distribution leads to the widespread existence of spectral similarities of different crops (same spectrum, different objects) and spectral differences of the same crop (same object, different spectrum), which are technical problems that are focused on solving by remote sensing data classification algorithms. Based on the above problems, classification algorithms usually use high-confidence pure crop pixels and diversified and complex crop identification indicators derived from long-term series data as input data to improve classification accuracy. At present, the mainstream classification algorithms mainly include supervised classification algorithms based on machine learning and unsupervised classification algorithms based on thresholds.

[0005] In the field of large-scale crop mapping, machine learning algorithms have certain classification advantages (such as support vector machine algorithms and random forest algorithms) and are currently the mainstream crop classification algorithms. Based on machine learning methods, classification rules are constructed in a data-driven manner by inputting multidimensional data (original spectral bands, various vegetation indices, terrain data, and environmental data, etc.), and the classification accuracy is relatively high. However, its algorithm belongs to a shallow structural model and is highly dependent on feature engineering (for example, texture features, etc.), making it difficult to extract and utilize deep features of remote sensing images. Although deep machine learning methods (deep learning) can achieve fully automatic deep crop feature mining, they rely on massive sample optimization models that consume a lot of time, money, and manpower costs, making them difficult to apply on a large scale. In addition, the effective identification indicators to which machine learning algorithms are applied have weak physical interpretability and relatively poor spatiotemporal generalization capabilities.

[0006] Unsupervised classification algorithms that do not rely on samples avoid the above problems to a certain extent and are another mainstream trend in large-scale crop classification technology. Based on threshold-based classification algorithms, the key characteristics of crops are captured by quantifying the changes in various vegetation indices during the crop growing season, and effective crop identification indicators are constructed to complete classification. From the perspective of crop identification characteristics, the participation of crop phenology knowledge can effectively improve classification accuracy, indicating that the participation of agricultural knowledge can increase the universality of the algorithm to a certain extent. However, the derivation of its identification indicators often relies on time-series images and complex mathematical formulas, and the application threshold is relatively high. At the same time, the threshold needs to be set manually, which is highly subjective and the classification accuracy is generally low.

[0007] In contrast, the crop index-based method (a type of threshold classification algorithm) achieves the classification of specific crops by directly calculating an index. The method is simple, efficient and highly interpretable. It does not require training data to achieve end-to-end, fully automatic crop mapping and is in a booming stage. Its principle is to enhance the information difference between the target crop and other types of landforms through the combination of specific spectral bands (products), thereby achieving accurate mapping of the target crop. However, the shallow involvement of phenological knowledge cannot meet the needs of multiple crop identification. It is necessary to explore the key physiological and biochemical characteristics of target crops at a deeper level and efficient formulaic expression. It belongs to the process of deep integration of agronomic knowledge and has great development potential. The proposed soybean mapping index has preliminarily verified the advantages of crop mapping based on agronomic knowledge-driven index, but the mining and construction of other major crop indices are still blank. As another staple food crop, the development of corn recognition index based on single-view imagery and the soybean and corn classification method based on agronomic knowledge-driven index have not been reported. Summary of the invention

[0009] The reliance on long time series images and training samples is the main factor limiting the development of large-scale crop classification algorithms, which is attributed to the lack of efficient and concise crop identification indicators. The present invention provides a large-scale soybean and corn mapping method based on an agronomic knowledge-driven index, explores efficient crop identification indicators that do not rely on time series images and sample data, and constructs an efficient and concise crop classification method suitable for large-scale crop classification.

[0010] The large-scale soybean and corn mapping method based on the agronomic knowledge-driven index of the present invention is carried out according to the following steps:

[0011] 1. Select a cloudless multispectral remote sensing image within the calculation time window TW (Time window) and calculate the GWCCI and CARI of the multispectral image;

[0012] Among them, the calculation time window TW is , They represent the time nodes of the start and end of TW respectively; GWCCI is the comprehensive green moisture index; CARI is the carotenoid index;

[0013] GWCCI (Greenness and Water Content Composite Index) is a comprehensive index of greenness and water content. It is a soybean mapping index and is used to identify soybeans (with a good identification effect). The calculation formula is as follows: ; ;

[0014] CARI (Carotenoid index) is a carotenoid index, which is a corn mapping index and can be used to identify corn (with a certain recognition effect). The calculation formula is as follows: ;

[0015] In the formula, , , , and Represents the reflectance of the blue, red, red edge, near infrared and short-wave infrared bands of Sentinel-2 images respectively; is the normalized difference vegetation index;

[0016] 2. Constructing the corn identification index GC, ;

[0017] 3. The pixel value of multispectral remote sensing image pixels is , , and are three defined thresholds; is the GWCCI index graph threshold, , is a threshold based on the GC index; if > If > Then the pixel is soybean or corn. Again with The comparison can determine the crop type of the pixel; if > Then the pixel is corn, Again with The comparison can determine the crop type of the pixel;

[0018] That is, the recognition and extraction of soybean and corn crop pixels are completed.

[0019] Furthermore, the value range of TW is as follows: ;

[0020] in, It is a custom experience threshold; For The rate of change of time; For The rate of change of time; For GCC in daily rate of change; ;

[0021] in, Indicates date; and In the and The GCC value of the day;

[0022] GCC calculation formula: ;

[0023] Among them, R, G, and B represent the red, green, and blue bands respectively.

[0024] Beneficial effects of the present invention:

[0025] 1. The present invention calculates the peak growth period of soybean and corn based on the GCC vegetation index as the optimal time window for algorithm application. Since the large-scale application of crop classification algorithms based on single-view images faces the problem of image loss due to cloud pollution, the present invention adopts the calculation of the optimal time window to largely alleviate the target crop omission error caused by the loss, and the median synthetic image within the time window can be used instead of the single date image to increase the robustness of the algorithm.

[0026] 2. The present invention constructs a new corn identification index GC, which fills the gap in the efficient corn identification technology based on single-view multispectral images, and further enriches the main crop identification and classification index system of single-view images. At present, the efficient crop identification indicators used by crop classification algorithms mostly come from the special phenological characteristics extracted from the entire life cycle of crops and the geometric characteristics of the curve itself. This process is highly dependent on long-term series images. The massive data generated by large-scale multi-year applications and the complex calculation of characteristic indicators limit the rapid and efficient promotion and application of the method. The crop identification index GC constructed by the present invention is based on single-view images. Compared with traditional algorithms, it can greatly reduce the amount of data and simplify the algorithm process when applied at the national / continental / global scale. In addition, GC is based on crop physiological and biochemical characteristics (greenness, canopy water content and pigments), deeply integrates agricultural knowledge, has a sufficient biophysical foundation, and has enhanced spatiotemporal generalization capabilities, and has great application potential.

[0027] 3. The soybean and corn decision tree model GGC constructed by combining the greenness and moisture comprehensive index (GWCCI) and the corn efficient identification index (GC) in this invention is the first efficient soybean and corn classification algorithm based on single-view images in the field. The recognition advantage of the GGC model is based on the grid search method, and finally a single-level decision tree model is constructed by determining three optimal thresholds (one for soybeans and two for corn), avoiding the error accumulation problem of multi-layer decision trees.

[0028] 4. The present invention is based on the pixel value of the multispectral remote sensing image pixel is , , and There are three defined thresholds, which enable soybean and corn identification indicators to be independently applied to the target crop classification, but the collaborative application needs to consider the problem of repeated classification of the same pixel, that is, the possibility that the target pixel is classified as soybean and also classified as corn. Based on the above problems, a binary tree is constructed to solve the problem of repeated classification of a single pixel. This process is the key to the GGC model's ability to effectively distinguish soybeans from corn. In addition, the proposed GGC model has strong robustness in different regions, different years, and different geographical environments and farmland management measures (such as irrigation and rain-fed, one crop / two crops per year, etc.), which effectively expands the scope of application of the algorithm, which is one of the important advantages that distinguish it from machine learning / deep learning algorithms (with weak generalization ability).

[0029] The present invention does not rely on long-term sequence images and training samples, and can identify and efficiently extract soybean and corn crop pixels from remote sensing images based on concise crop identification indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a dynamic spectral characteristic diagram of the three indicators of GWCCI, CARI and GC in different crop growing seasons in Example 1;

[0032] Figure 2 This is the GGC classification process in Example 1;

[0033] Figure 3 This is the classification result of soybean and corn based on the GGC model in Example 1. DETAILED DESCRIPTION

[0035] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0036] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0037] Specific implementation method 1: The large-scale soybean and corn mapping method based on the agricultural knowledge-driven index in this implementation method is carried out according to the following steps:

[0038] 1. Select a cloudless multispectral remote sensing image within the calculation time window TW (Time window, TW), and calculate the GWCCI and CARI of the multispectral image;

[0039] Among them, the calculation time window TW is , They represent the time nodes of the start and end of TW respectively; GWCCI is the comprehensive green moisture index; CARI is the carotenoid index;

[0040] GWCCI is a comprehensive green moisture index, which is used to identify soybeans (with a good identification effect). The calculation formula is as follows: ; ;

[0041] CARI is the carotenoid index, which is used to identify corn (with a certain identification effect). The calculation formula is as follows: ;

[0042] In the formula, , , , and They represent the reflectance of the blue, red, red-edge, near-infrared and short-wave infrared bands of the Sentinel-2 image, respectively.

[0043] 2. Constructing the corn identification index GC, ;

[0044] 3. The pixel value of multispectral remote sensing image pixels is , , and are three defined thresholds; is the GWCCI index graph threshold, , is a threshold based on the GC index; if > If > Then the pixel is soybean or corn. Again with The comparison can determine the plant in the pixel; if > Then the pixel is corn, Again with The comparison can determine the plant in the pixel;

[0045] That is, the recognition and extraction of soybean and corn crop pixels are completed.

[0046] In this embodiment, the GWCCI value of the soybean growth peak period (PGS) is significantly different from that of the other two periods, LGS and EGS; while the CARI value of the corn growth peak period (PGS) is slightly different from that of the other two periods, LGS and EGS.

[0047] This implementation method takes soybean and corn, two global staple food / cash crops, and typical row-planted dryland crops as target crops, and proposes an end-to-end soybean and corn classification method based on single-view images. Based on physiological and biochemical indicators such as canopy moisture content, greenness and pigment, this method uses the differential characteristics of GWCCI (Greenness and watercontent composite index) and CARI (Carotenoid index) at different growth stages of crops to construct a new ratio crop index GC (GWCCI / CARI) for identifying corn, and then constructs a decision tree model based on crop index (GWCCIand GC combined model, GGC) to achieve soybean and corn classification based on single-view images. This method fills the gap in corn recognition index, and further improves the ability of greenness and water comprehensive index to filter out other crops. In addition, the GGC model has strong spatiotemporal generalization ability and does not need to rely on training samples and long time series data. This agricultural knowledge-driven soybean and corn classification method provides a new way to achieve large-scale crop mapping.

[0048] This implementation achieves the following goals:

[0049] (1) Analyze the temporal spectral characteristics of different land covers and explore the unique spectral characteristics of target crops at key growth stages;

[0050] (2) Deeply integrate agricultural knowledge, construct efficient identification indicators for target crops, and develop classification methods based on crop identification index;

[0051] (3) Evaluate the spatiotemporal generalization ability of the proposed method and explain the biophysical basis and limiting factors of crop identification indicators.

[0052] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that:

[0053] The value range of TW is as follows: ;

[0054] in, It is a custom experience threshold; For The rate of change of time; For The rate of change of time; For GCC in daily rate of change; ;

[0055] in, Indicates date; and In the and The GCC value of the day;

[0056] GCC calculation formula: ;

[0057] Wherein, R, G, and B represent red, green, and blue bands respectively. Other steps and parameters are the same as those in the first specific implementation mode.

[0058] The GCC described in this embodiment gradually increases with the growth of the crop, and then tends to be stable during the growth peak period. The GCC during the growth peak period is accompanied by a relatively small change rate. .

[0059] The GCC described in this embodiment is mainly used for extracting vegetation phenological characteristics and characterizing vegetation growth status. Compared with the normalized difference vegetation index, the calculation result is more accurate.

[0060] Specific implementation method three: This implementation method is different from the specific implementation method one in that: , and The three definition thresholds are obtained based on the grid search method. The other steps and parameters are the same as those in the first embodiment.

[0061] Example 1 Large-scale soybean and corn mapping method based on agronomic knowledge-driven index (GGC)

[0062] The specific method is as follows:

[0063] Determine the best phenological period for soybean and corn (calculation time window, Time window, TW). In order to accurately obtain TW in different regions, we apply the GCC vegetation index (Green Chromatic Coordinate, GCC). GCC is mainly used to extract vegetation phenological characteristics and characterize vegetation growth status. Compared with the normalized vegetation index, the calculation results are more accurate. In addition, the calculation of GCC only requires red, green and blue bands, which is easy to promote and apply. The calculation method of GCC is as follows: ;

[0064] In the formula, R, G, and B represent the red, green, and blue bands, respectively. GCC gradually increases with the growth of crops, and then tends to stabilize in the PGS interval, that is, with a relatively small change rate ( ), calculated as follows: ;

[0065] In the formula, Indicates date; and In the and GCC value of the day; Indicates that GCC is in The rate of change of the day. Based on the above formula, TW is defined as follows: ;

[0066] In the formula, , Respectively represent the time nodes of the start and end of TW, is a user-defined empirical threshold.

[0067] Construct a new corn identification index GC. First, introduce a pigment index CARI to identify corn. Studies have shown that there are differences in carotenoid content between corn and other crops, which is conducive to corn identification and classification. Combined with the Green Water Comprehensive Index (GWCCI), within a given time window Choose any cloudless multispectral remote sensing image, calculate the GWCCI and CARI values ​​of the image, and analyze the index performance characteristics of crops at different growth stages. Figure 1 The dynamic spectral characteristic diagrams of the three indicators GWCCI, CARI and GC in different crop growing seasons in Example 1 (EGS, PGS and LGS represent the early growth period, peak growth period and late growth period of crops, respectively); Figure 1 It was found that the GWCCI value during the peak growth period of corn (i.e., within the time window) was significantly different from the other two periods (early growth period and late growth period), while the CARI value during the peak growth period of corn was slightly different from the other two periods. Based on the above differential characteristics, a ratio vegetation index was constructed to identify corn. The calculation formula of the index is as follows: ;

[0068] Formulate classification rules and build a decision tree model GGC. Combine the time series curve characteristics of the main crop index (GWCCI and GC) and the performance of the soybean and corn index based on single-view images, analyze the index to identify the advantages and disadvantages of soybeans and corn, and formulate classification rules. The flow chart is as follows Figure 2 As shown, Figure 2 This is a GGC classification flow chart in Example 1.

[0069] Figure 2 middle, , and are three user-defined thresholds obtained based on a grid search method. is a threshold value of the GWCCI index map for soybean classification. When , it is identified as a soybean pixel, otherwise it is a non-soybean pixel; , is a threshold determined based on the GC index graph. is a looser threshold. When a pixel value is greater than When , the pixel is identified as soybean or corn. The purpose of this threshold is to filter out other crops from the GWCCI map; Is a stricter threshold. When a pixel value is greater than , we conclude that the pixel is definitely corn.

[0070] Based on the above classification process, we formulated classification rules for corn and soybeans respectively, and named the method as the GGC model. The classification rules are as follows: ; ;

[0071] Obtain any multispectral image within the given study area time window, calculate the GWCCI and GC index map, and finally complete the identification and extraction of soybean and corn crop pixels from remote sensing images based on the above classification rules. The results show that the spatial distribution pattern of soybean and corn based on the GGC model is consistent with the actual distribution ( Figure 3 is the classification result of soybean and corn based on the GGC model). Compared with the comparison method, there are fewer omission errors. When applied in different regions, the average overall classification accuracy exceeds 85% (88.90%) (Table 1 Classification accuracy of soybean and corn based on the GGC model), and it has good spatiotemporal generalization ability.

[0072] Table 1 Classification accuracy of soybean and corn based on GGC model

[0073]

[0074] The present invention takes soybean and corn, two global staple food / economic crops, and typical row-planted dryland crops as target crops, and proposes an end-to-end soybean and corn classification method based on single-view images. Based on physiological and biochemical indicators such as canopy moisture content, greenness and pigment, this method uses the differential characteristics of GWCCI (Greenness and water content composite index) and CARI (Carotenoid index) of corn at different growth stages to construct a new ratio crop index (GWCCI / CARI) for identifying corn, and then constructs a decision tree model based on crop index (GWCCI and GCcombined model, GGC) to achieve soybean and corn classification based on single-view images. This method fills the gap in corn recognition index and further improves the ability of greenness and water composite index to filter out other crops. In addition, the GGC model has strong spatiotemporal generalization ability and does not need to rely on training samples and long time series data. This agronomic knowledge-driven soybean and corn classification method provides a new way to achieve large-scale crop mapping.

Claims

1. A large-scale soybean and corn mapping method based on agronomic knowledge-driven index, characterized in that: The large-scale soybean and corn mapping method based on the agronomic knowledge-driven index is carried out in the following steps:

1. Select a cloudless multispectral remote sensing image within the calculation time window TW and calculate the GWCCI and CARI of the multispectral image; Among them, the calculation time window TW is , They represent the time nodes of the start and end of TW respectively; GWCCI is the comprehensive green moisture index; CARI is the carotenoid index; ； ； ； In the formula, , , , and Represents the reflectance of the blue, red, red edge, near infrared and short-wave infrared bands of Sentinel-2 images respectively; is the normalized difference vegetation index; 2. Constructing the corn identification index GC, ； 3. The pixel value of multispectral remote sensing image pixels is , , and are three defined thresholds; is the GWCCI index graph threshold, , is a threshold based on the GC index; if > If > Then the pixel is soybean or corn. Again with The comparison can determine the crop type of the pixel; if > Then the pixel is corn, Again with The comparison can determine the crop type of the pixel; That is, the recognition and extraction of soybean and corn crop pixels are completed.

2. A large-scale soybean and corn mapping method based on agronomic knowledge-driven index according to claim 1, characterized in that: ； in, It is a custom experience threshold; For The rate of change of time; For The rate of change of time; For GCC in daily rate of change; ； in, Indicates date; and In the and The GCC value of the day; GCC calculation formula: ； Among them, R, G, and B represent the red, green, and blue bands respectively.

3. The large-scale soybean and corn mapping method based on agronomic knowledge-driven index according to claim 1, characterized in that: , and The three defined thresholds were obtained based on the grid search method.

Images