Rice remote sensing mapping method based on unique spectral characteristics in rice transplanting period

Abstract

The invention discloses a rice remote sensing mapping method based on unique spectral characteristics in a rice transplanting period, and relates to a rice remote sensing identification method. The method aims at solving the problems that an existing rice drawing method is poor in mobility, a large number of parameters or threshold values need to be set, and long-time-sequence or year-round remote sensing data are relied on. The method comprises the following steps: 1, determining a wave band combination reflecting a unique color in a rice transplanting period; 2, mapping a spectral band to a CIE color space; and 3, constructing a rice identification method in a CIE color space. The method is clear in principle, simple, rapid, high in automation degree and particularly suitable for large-scale (national / intercontinental / global) rice drawing, and has huge application potential and wide application space. The invention belongs to the field of remote sensing mapping.

Description

Technical Field

[0001] The invention relates to a rice remote sensing identification method. Background Art

[0002] Crop classification based on remote sensing data relies on the unique phenological and spectral characteristics of crops. Due to the diversity of planting systems (including crop types, crop varieties, management methods, and planting scales) and the huge differences in climate, environment, and natural conditions in different countries and regions, the phenological and spectral characteristics of many crops overlap, resulting in low classification accuracy for single crops. Rice, as an important food and cash crop, is usually planted at the same time as other food crops such as corn and soybeans, and its spectral characteristics are similar. When identifying a single crop (rice), it is very easy to be confused with other crops (especially corn). At present, scholars have developed a variety of rice mapping methods, mainly including classifier-based methods, phenological-based methods, and index-based methods. However, these methods currently face three main problems: (1) the method has weak transferability; (2) a large number of parameters or thresholds need to be set; and (3) they rely on long-term or full-year remote sensing data. Summary of the invention

[0003] The purpose of the present invention is to solve the technical problems that the existing rice mapping methods have weak mobility, need to set a large number of parameters or thresholds, and rely on long-term or annual remote sensing data, and provide a rice remote sensing mapping method based on the spectral characteristics of the rice transplanting period.

[0004] The rice remote sensing mapping method (RiceTColour) based on the unique spectral characteristics of rice transplanting period is as follows:

[0005] 1. Determine the band combination that reflects the unique color of rice transplanting period:

[0006] Download multispectral remote sensing images of rice transplanting period, select and use SWIR band, NIR band and RED band as three primary color values, namely red, green and blue channels to display images;

[0007] 2. Map the spectral bands to the CIE color space:

[0008] The three primary color values ​​are converted into the X and Y coordinates of the two-dimensional CIE chromaticity using the following formula:

[0009] x=X / (X+Y+Z) (1)

[0010] y=Y / (X+Y+Z) (2)

[0011] X=2.7689R+1.7517G+1.1302B (3)

[0012] Y=1.0000R+4.5907G+0.0601B (4)

[0013] Z=0.0000R+0.0565G+5.5943B (5)

[0014] In the formula, x and y represent the X coordinate and Y coordinate of the CIE color space respectively; the SWIR band, NIR band and RED band are used as the R band, G band and B band respectively, and the observed color is converted into the CIE color space;

[0015] 3. Constructing rice recognition method in CIE color space:

[0016] In order to distinguish rice from other land objects in CIE color space, a lower boundary was established for the scattered points of rice, which corresponds to the lightest green in the selected rice samples. Points above the lower boundary correspond to rice pixels showing darker and greener colors. A polynomial regression boundary was generated and fitted by fitting the rice sample points. First, all convex hull vertices (points) of the rice samples were identified. Second, the vertices with the minimum and maximum x-coordinate values ​​were selected to form a straight line, and the vertices above the straight line were deleted. Third, a third-order polynomial regression was generated by fitting the remaining vertices. The established polynomial regression model is as follows:

[0017] y lower =282.82119x 3 -227.05549x 2 +60.62184x-5.03751 (6)

[0018] The sample pixels of rice are distributed in a specific range in the CIE color space: the X-axis range is 0.235 to 0.346, and the Y-axis range is less than 0.5; therefore, it is stipulated that if the x-coordinate value and y-coordinate value (x, y) of each candidate pixel in the remote sensing image meet the following conditions in the CIE color space:

[0019] y lower <y<0.5 and 0.235<x<0.346 (7)

[0020] It is classified as rice, thus completing the construction of rice recognition rules in the CIE color space, that is, completing rice remote sensing mapping based on the spectral characteristics of the rice transplanting period.

[0021] RiceTColour overall idea:

[0022] Determine the band combination that reflects the unique color of rice transplanting period

[0023] Download multispectral remote sensing images of the rice transplanting period and analyze the spectral characteristics of rice fields, dry land, woodland, grassland, water bodies, construction land and bare land in the visible light (VIS), near infrared (NIR) and short wave infrared (SWIR) bands. Through analysis, rice shows unique spectral characteristics in the SWIR (reflecting moisture) and NIR (reflecting greenness) bands. Due to the humid background, the reflectance values ​​of rice and water bodies in the SWIR band are relatively low (mostly below 0.05), and are generally lower than the reflectance of other objects, especially in the SWIR1 band. Although most of the incident infrared light is transmitted and reflected by the top leaves of vegetation, the NIR value of rice is relatively low (mainly between 0.10 and 0.15), compared with other objects due to its small amount of vegetation (rice seedlings). However, the NIR value of rice is higher than the reflectance value of smooth water surface.

[0024] In view of the unique spectral characteristics of rice during transplanting in the SWIR and NIR bands mentioned above, SWIR, NIR and RED (red band) were selected and used as red, green and blue channels to display the image, respectively. The red band was selected because it has been widely shown to be sensitive to vegetation biomass and leaf area index. Transplanted rice fields show a unique dark green color in the false color composite image composed of SWIR (R), NIR (G) and red (B) bands. Based on these specific colors, rice fields can be easily distinguished from other landforms in the image.

[0025] Mapping spectral bands to CIE color space

[0026] Based on the unique spectrum (color) displayed by transplanted rice fields, this paper proposes a new rice mapping method, which is carried out in the color space created by the International Commission on Illumination (CIE, French name Commission Internationale del'Eclairage) in 1931. The main idea of ​​this method is to establish a classification standard to distinguish rice from other ground objects by analyzing the spatial distribution pattern of training samples of rice and other ground objects in the CIE color space. The biggest advantage of the CIE color space is that it can convert any color composed of traditional three primary color values ​​(RGB: red, green, blue) into a quantitative two-dimensional (2D) color space. In other words, any color observed by the human eye in remote sensing images has a specific numerical position in the CIE color space, which lays the foundation for quantitatively characterizing color (category) similarity. The three primary color values ​​are converted into two-dimensional CIE chromaticity coordinates (x, y) using standard formulas.

[0027] Constructing a rice recognition method in CIE color space (RiceTColour)

[0028] In order to distinguish rice from other features in CIE space, a lower boundary was established for the scattered points of rice. This boundary corresponds to the lightest green in the selected rice samples, and the points above the boundary correspond to rice pixels that show darker and greener colors. A polynomial regression boundary was generated and fitted by fitting the rice sample points. First, all the convex hull vertices (points) of the rice samples were identified. Second, the vertices with the minimum and maximum x-coordinate values ​​were selected to form a straight line, and the vertices above the straight line were deleted. Third, a third-order polynomial regression was generated by fitting the remaining vertices. The established polynomial regression model is as follows:

[0029] y lower =282.82119x 3 -227.05549x 2 +60.62184x-5.03751

[0030] After analysis, the sample pixels of rice are distributed in a specific range in CIE space: the X-axis range is 0.235 to 0.346, and the Y-axis range is below 0.5. Therefore, it is stipulated that if each candidate pixel (x, y) in the remote sensing image meets the following rule conditions in CIE space, it will be classified as rice:

[0031] y lower <y<0.5 and 0.235<x<0.346

[0032] Thus, the construction of rice recognition rules in CIE color space is completed.

[0033] The present invention proposes a new rice remote sensing mapping method (RiceTColour) based on the spectral characteristics of rice during transplanting period. RiceTColour pioneered the rice identification mapping in CIE color space based on the unique spectral characteristics of rice during transplanting period, thus solving the difficulties currently faced by rice remote sensing mapping. The present invention discovered for the first time that during the rice transplanting period, rice presents a unique blue-green color in images composed of SWIR, NIR and RED bands. By converting the three bands of SWIR, NIR and RED into the CIE color space, rice can be completely separated from other land covers, thereby constructing a unique rice identification rule.

[0034] The present invention does not need to rely on a large number of long-term remote sensing images (avoiding complex and time-consuming image preprocessing processes), and only needs to obtain any multispectral remote sensing image during the rice transplanting period to achieve accurate rice mapping. This method is a parameter-free method (does not contain any parameters and thresholds), which avoids the tedious and lengthy parameter debugging process, and thus has good generalizability. This method is an end-to-end, fully automatic rice mapping method with clear principles, simplicity and speed, and a high degree of automation. It does not require training samples to build a model, and the calculation is simple, convenient, and fast. It is particularly suitable for large-scale rice mapping application needs, and has huge application potential and a broad application space.

[0035] The advantages of the present invention include three aspects: (1) it has strong migration ability; (2) it is a method that does not require parameters and thresholds; and (3) it only uses a single image or a composite image collected during the transplantation period. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the basic flow chart of the present invention;

[0037] Figure 2 It is the CIE chromaticity diagram in Experiment 1, the dot density diagram of rice training samples in S1 in the CIE color space, and the scatter diagram of training samples of rice in S1 and other five types of land covers in the CIE color space; (a) in the figure is a diagram of the CIE chromaticity diagram (i.e., color space); (b) is a dot density diagram of rice training samples in S1 in the CIE color space; (c) is a scatter diagram of training samples of rice in S1 and other five types of land covers (dry land, woodland, water area, built-up area, and bare land) in the CIE color space;

[0038] Figure 3 This is a comparison of the rice mapping results of the RiceColour method proposed in Experiment 1 and the corresponding satellite sensor images (RGB: SWIR, NIR, Red) at the five study locations;

[0039] Figure 4 Comparison of rice mapping results obtained by the RiceColour method and the baseline method at the five study sites in Experiment 1; the columns from left to right represent the Landsat 8 satellite sensor imagery, and the results of ARM-SARFS, NDWI, RFC, OCSVM and the proposed RiceColour, respectively. The produced rice maps (dark red) are superimposed on the satellite images for easy visual interpretation. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is not limited to the specific implementation modes listed below, but also includes any combination of the specific implementation modes.

[0041] Specific implementation method 1: The rice remote sensing mapping method based on the spectral characteristics of rice transplanting period in this implementation method is as follows:

[0042] 1. Determine the band combination that reflects the unique color of rice transplanting period:

[0043] Download multispectral remote sensing images of rice transplanting period, select and use SWIR band, NIR band and RED band as three primary color values, namely red, green and blue channels to display images;

[0044] 2. Map the spectral bands to the CIE color space:

[0045] The three primary color values ​​are converted into the X and Y coordinates of the two-dimensional CIE chromaticity using the following formula:

[0046] x=X / (X+Y+Z) (1)

[0047] y=Y / (X+Y+Z) (2)

[0048] X=2.7689R+1.7517G+1.1302B (3)

[0049] Y=1.0000R+4.5907G+0.0601B (4)

[0050] Z=0.0000R+0.0565G+5.5943B (5)

[0051] In the formula, x and y represent the X coordinate and Y coordinate of the CIE color space respectively; the SWIR band, NIR band and RED band are used as the R band, G band and B band respectively, and the observed color is converted into the CIE color space;

[0052] 3. Constructing rice recognition method in CIE color space:

[0053] In order to distinguish rice from other land objects in CIE color space, a lower boundary was established for the scattered points of rice, which corresponds to the lightest green in the selected rice samples. Points above the lower boundary correspond to rice pixels showing darker and greener colors. A polynomial regression boundary was generated and fitted by fitting the rice sample points. First, all convex hull vertices (points) of the rice samples were identified. Second, the vertices with the minimum and maximum x-coordinate values ​​were selected to form a straight line, and the vertices above the straight line were deleted. Third, a third-order polynomial regression was generated by fitting the remaining vertices. The established polynomial regression model is as follows:

[0054] y lower =282.82119x 3 -227.05549x 2+60.62184x-5.03751 (6)

[0055] The sample pixels of rice are distributed in a specific range in the CIE color space: the X-axis range is 0.235 to 0.346, and the Y-axis range is less than 0.5; therefore, it is stipulated that if the x-coordinate value and y-coordinate value (x, y) of each candidate pixel in the remote sensing image meet the following conditions in the CIE color space:

[0056] y lower <y<0.5 and 0.235<x<0.346 (7)

[0057] It is classified as rice, thus completing the construction of rice recognition rules in the CIE color space, that is, completing rice remote sensing mapping based on the spectral characteristics of the rice transplanting period.

[0058] The following experiments were used to verify the effect of the present invention:

[0059] Experiment 1:

[0060] 1. Region and data of the present invention

[0061] 1.1. Region of the present invention

[0062] Five typical rice-growing regions under different environmental conditions were carefully selected (geographic locations of the five study regions: S1, Chiayi and Tainan counties on Taiwan Island, my country; S2, Fujin County, Heilongjiang Province; S3, Yinchuan and Shizuishan cities, Ningxia Hui Autonomous Region; S4, Gushi County, Henan Province; S5, Binyang County, Guangxi Zhuang Autonomous Region; detailed information of each study region is shown by Landsat 8 true color images.) to comprehensively and thoroughly test the effectiveness and generalizability of the developed RiceColour rice mapping method. These regions vary greatly in climate (tropical, cold, arid, temperate, and subtropical), rice crop density (monocropping and multiple cropping), irrigation schemes (surface water irrigation and groundwater irrigation), and crop rotations (rice, other rice, and fallow rice) (Table 1).

[0063] The first study area (S1) of the present invention is located in the southwest of Taiwan Island, my country, consisting of Chiayi and Tainan counties, with a total area of ​​4,155 square kilometers. Rice is the main crop in the region, accounting for about a quarter of the total farmland area. Transplanting is generally carried out from January to March in the rainy season, harvesting in June and July, and transplanting in July and August in the dry season, and harvesting in November and December. The main irrigation water source in the region comes from reservoirs, supplemented by rivers and groundwater. Because it is managed by small farmers, the farmland here is relatively small, with an average rice field area of ​​about 0.2 hectares, scattered in the Jianan Plain. The second to fifth (S2 to S5) study areas of the present invention are located in mainland my country, specifically in the Sanjiang Plain (S2) in northeastern my country, the Yinchuan Plain (S3) in northwest my country, the middle and lower reaches of the Yangtze River Plain (S4) in central and eastern my country, and the Nanning Basin (S5) in southern my country. The agricultural landscapes of S2 and S4 are quite simple, dominated by rice. In contrast, according to the field survey, the landscapes of the remaining two sites were relatively complex, with four main crops: rice, corn, spring wheat, and fodder crops in S3, and rice, sugarcane, corn, and sweet potatoes in S5. The proportion of rice fields in S2 to S5 to the total cultivated land area was 89.54%, 20.06%, 76.65%, and 20.91%, respectively. Regarding irrigation schemes, the main irrigation water sources for S2 came from the Songhua River and groundwater extracted by water pumps. According to the irrigation network, the rice fields of S3 were mainly irrigated from the Yellow River, while the rice fields of S4 and S5 relied mainly on reservoir water for irrigation (Table 1).

[0064] Table 1 Summary description of the five study areas

[0065]

[0066]

[0067] 1.2. Data

[0068] 1.2.1. Ground reference data

[0069] The ground reference data (training and validation plots) of the proposed study area were collected by referring to rice maps, field surveys, and VFR images (such as Google Earth (GE) VFR data). The training samples were collected only from S1 to establish the rice mapping model, while the validation samples were collected in each of the five proposed regions to evaluate the validity of the model. A stratified random sampling scheme was used to obtain the training samples. Specifically, a total of 378 training polygons were determined, including 188 rice patches and 190 other land cover patches. The training polygons were superimposed with Landsat8 data, and the pixels within the training polygons were used as training samples, resulting in 3503 rice plots and 7842 non-rice plots (including 1216 drylands, 2303 forests, 1509 waters, 1203 built-up areas, and 1611 bare land). In order to extensively verify the validity of the proposed method, 3393 validation samples were collected in each study area, and the number of samples for rice and non-rice classes was strictly proportional to the total area of ​​each class.

[0070] 1.2.2. Satellite remote sensing images

[0071] The Landsat-8 surface reflectance level 1 product (T1_SR) provided by the Google Earth Engine (GEE) cloud computing platform was used for rice mapping. To collect data from Landsat 8, the availability of cloud-free T1_SR products for each study area collected during the transplanting period was checked on GEE. Both S3 and S5 have single images without clouds, and the weather usually remains clear in spring. However, it is extremely difficult to obtain cloud-free images in the remaining three areas, especially in S1 and S4 where there are few rainy and cloudy days. Therefore, single images were used in S3 and S5, and composite images were used in S1, S2, and S4 for rice mapping. In order to test the robustness of the rice mapping method, selected images of each study area were collected in different years (2015 to 2020). The detailed information on the Landsat 8 image acquisition used in the present invention is shown in Table 2.

[0072] In addition to Landsat-8, Sentinel-1 SAR data were collected for comparison in the present region. Full-year Sentinel-1 ground-based sounding (GRD) SAR data covering S1 to S4 were obtained from Google Earth Engine (GEE), with approximately 30 valid observations for each region (Table 2). GRD data were calibrated and orthogonally corrected dual-polarization (VV / VH) products with a spatial resolution of 10 m. Since S5 had no SAR data from January to May, the region was excluded from SAR-based rice mapping.

[0073] Table 2 Detailed information of rice mapping images acquired by Landsat 8 and Sentinel-1

[0074]

[0075] 2. Method comparison

[0076] In this invention, a new method named RiceTColour is proposed to detect rice fields based on the unique spectrum of rice fields during transplanting using CIE color space. The workflow of RiceColour is as follows Figure 4 As shown, the details are as follows.

[0077] The effectiveness of the proposed rice mapping method (RiceColour) was evaluated using four other benchmark methods: a SAR-based method, an index-based method, and two supervised classifier-based methods. For the SAR-based methods, the recently proposed ARM-SARFS automatic rice mapping method was adopted because it outperformed other SAR-based methods. ARM-SARFS includes four predefined thresholds, namely T1, T2, T3, and T4. In the experiments, a grid search from -30 to -10 with a step size of 1 was used to adjust the optimal values ​​of T1 and T2 for each region of the present invention. The optimal combinations of T1 and T2 for S1 to S4 were -20 and -23, -25 and -25, -25 and -25, and -20 and -20, respectively. T3 and T4 are fixed thresholds depending on the rice type (early rice, mid-season rice, and late rice) and are directly used for rice mapping.

[0078] For the index-based approach, the Normalized Difference Water Index (NDWI) was chosen, which is calculated by calculating the normalized difference between the green band (i.e., band 3 of Landsat 8) and the SWIR band (i.e., band 6 of Landsat 8). Through cross-validation, the threshold of NDWI was adjusted from -0.30 to -0.10 with a step size of 0.002, and the optimal threshold was found to be -0.228. For the supervised classifier-based approach, the Random Forest Classifier (RFC) and One-Class Support Vector Machine (OCSVM) were used as references. To ensure a fair comparison, the SWIR, NIR, and Red bands of Landsat 8 images were used as inputs to both classifiers, consistent with the inputs of the proposed RiceColour method.

[0079] Two important parameters of OCSVM, and, which control the kernel width and the proportion of abnormal samples of the target class respectively, need to be carefully determined. Using a grid search method, the best and were optimized to 10 and 0.01 from (0.01, 0.05, 0.1, 0.5, 1, 2.5, 5, 10, 25) and (0.01, 0.05%, 0.1, 0.5, 1, 2.5, 5, 10, 25), respectively. In order to quantitatively evaluate the accuracy of the produced rice map, four indicators were used, including overall accuracy (OA), producer accuracy (PA), user accuracy (UA), and F1 score. The F1 score is the harmonic mean of PA and UA. Compared with OA, it can better indicate the ability of the model to identify imbalanced classes.

[0080] 3. Results and Analysis

[0081] 3.1. Rice mapping results of RiceColour

[0082] The pixel-based minimum SWIR synthesis method was applied in S1, S2, and S4 to generate composite images for rice mapping, while in S3 and S5, single-date images were used for rice classification (Table 2). Figure 3 The mapping results generated by the proposed RiceColour method and the corresponding satellite sensor images of all five study areas of the present invention are shown. As can be seen from the figure, the rice map produced is highly consistent with the spatial pattern of rice fields observed in the satellite sensor images of the five study areas (identifiable by their dark green color). This demonstrates the effectiveness and significant generalization ability of the proposed method for rice detection in rice-growing areas under various natural conditions and management practices. The accuracy of the rice map was evaluated using validation samples from the five study areas, and the confusion matrix and corresponding classification accuracy are listed in Table 3. As shown in the table, the RiceColour method achieved very high mapping accuracy, with OAs above 95% for all study areas of the present invention. Importantly, the RiceColour method was very stable in the five study areas, producing OAs of approximately 97% (ranging from 95.76% to 98.50%), indicating its significant generalization ability in rice detection. In addition, it produced relatively balanced PA and UA for the target class of rice of the present invention. The PA of all the present points (except S3) was above 92%, indicating that almost all (more than 92%) rice fields in each present point were successfully detected by the proposed method. Rice UA was also >90% for most sites (S2, S4, and S5), indicating that only a small fraction of non-rice pixels were incorrectly labeled as rice. Balanced rice mapping accuracy can also be observed from the F1 score, which ranged from 0.87 to 0.97 across the five sites.

[0083] Table 3 Confusion matrix of classification maps generated by RiceTColour method and classification accuracy of the five study sites

[0084]

[0085] 3.2. Rice mapping benchmark comparison

[0086] The rice mapping results of the proposed RiceColour method were further tested with four benchmark methods (ARM-SARFS, NDWI, RFC, and OCSVM) in five study areas. The rice mapping results of the five methods were validated by visual inspection and quantitative accuracy assessment. To facilitate visual inspection, one or two typical rice-growing areas were selected in the five proposed sites for comparison. The original Landsat 8 satellite sensor images and the mapping results of various methods in these areas are shown in Figure 2. Figure 4 As shown in the figure, the proposed RiceColour method consistently achieves the most accurate and ideal rice maps during S1 to S5. Figure 4 (a) and (b)), it is clear that ARM-SARFS fails to identify a large number of rice fields, while NDWI and RFC both misclassify a large portion of built-up areas and fish ponds as rice. In addition, NDWI, RFC, and OCSVM often misclassify roads in and around farmland as rice. In contrast, the RiceColour method accurately distinguishes rice from built-up areas, fish ponds, and road networks (clearly visible on the map) ( Figure 4 (a) and (b)). For S2 to S5, the SAR-based ARM-SARFS produced inaccurate rice classification maps, missing a large portion of rice fields, especially in S3. Although the remaining three baseline methods (NDWI, RFC, and OCSVM) performed more accurately than ARM-SARFS, NDWI tended to overestimate rice, while both RFC and OCSVM tended to underestimate rice. In contrast, OCSVM missed a relatively large proportion of rice in S2, S3, and S5, and missed the edges of most rice fields in S4. Although RFC outperformed OCSVM in mitigating underestimation, it still failed to detect many small areas of rice fields in S2 to S4. Surprisingly, as shown in Figure 4 As shown in (c) to (g), the proposed RiceColour method corrects almost all of the above misclassifications and accurately and completely detects rice.

[0087] Quantitative accuracy evaluation further demonstrates the effectiveness and superiority of the proposed method. As shown in Table 4, the RiceColour method consistently achieves the highest OA compared to the baseline method, which is consistent with the previous visual inspection results. Although ARM-SARFS achieved moderate overall accuracy (75%-90%) during S1 to S4, it achieved very low producer accuracy (<71%), especially in S1 and S3 (only 42.26% and 37.04%, respectively), where the proportion of rice fields was relatively low (Table 1). NDWI achieved highly accurate PA (91%-99%) in all areas of the present invention; however, this was obtained at the expense of high misclassification (i.e., low UA). For example, the UA in S3 was only about 33%, indicating that more than two-thirds (67%) of the pixels classified as rice were misclassified (i.e., non-rice pixels were incorrectly identified as rice pixels). In other words, the rice area was seriously overestimated, especially in areas with a relatively small proportion of rice (S1, S3, and S5). Both supervised classifiers (RFC and OCSVM) achieved promising accuracies (OA and PA) in S1, comparable to that of RiceColour. However, their producer accuracies decreased significantly when generalized to S2 to S5. For example, when OCSVM was generalized to S2 to S5, its PA was very poor, ranging between 55.97% (S5) and 79.37% (S4). Similarly, RFC’s PA also decreased significantly when generalized to S2 to S5, especially in S2 (89.46%) and S3 (86.77%).

[0088] Table 4. Accuracy of rice field maps generated using five methods (ARM-SARFS, NDWI, RFC, OCSVM, and RiceTColour) at five study sites. The largest (best) value in each row is shown in bold.

[0089]

Claims

1. A rice remote sensing mapping method based on the unique spectral characteristics of rice transplanting period, characterized by The rice remote sensing mapping method based on the spectral characteristics of rice transplanting period is as follows:

1. Determine the band combination that reflects the unique color of rice transplanting period: Download multispectral remote sensing images of rice transplanting period, select and use SWIR band, NIR band and RED band as three primary color values, namely red, green and blue channels to display images; 2. Map the spectral bands to the CIE color space: The three primary color values ​​are converted into the X and Y coordinates of the two-dimensional CIE chromaticity using the following formula: x=X / (X+Y+Z) (1) y=Y / (X+Y+Z) (2) X=2.7689R+1.7517G+1.1302B (3) Y=1.0000R+4.5907G+0.0601B (4) Z=0.0000R+0.0565G+5.5943B (5) In the formula, x and y represent the X coordinate and Y coordinate of the CIE color space respectively; the SWIR band, NIR band and RED band are used as the R band, G band and B band respectively, and the observed color is converted into the CIE color space; 3. Constructing rice recognition method in CIE color space: In order to distinguish rice from other land objects in the CIE color space, a lower boundary was established for the scattered points of rice, which corresponds to the lightest green in the selected rice samples. The points above the lower boundary correspond to rice pixels showing darker and greener colors. A polynomial regression boundary was generated and fitted by fitting the rice sample points. First, all convex hull vertices of the rice samples were identified. Second, the vertices with the minimum and maximum x-coordinate values ​​were selected to form a straight line, and the vertices above the straight line were deleted. Third, a third-order polynomial regression was generated by fitting the remaining vertices. The established polynomial regression model is as follows: y lower =282.82119x 3 -227.05549x 2 +60.62184x-5.03751 (6) The sample pixels of rice are distributed in a specific range in the CIE color space: the X-axis range is 0.235 to 0.346, and the Y-axis range is less than 0.5; therefore, it is stipulated that if the x-coordinate value and y-coordinate value of each candidate pixel in the remote sensing image meet the following conditions in the CIE color space: y lower <y<0.5 and 0.235<x<0.346 (7) It is classified as rice, thus completing the construction of rice recognition rules in the CIE color space, that is, completing rice remote sensing mapping based on the spectral characteristics of the rice transplanting period.

Images