Regional farmland water-saving irrigation and conventional irrigation classification and high-precision mapping method based on machine learning

Abstract

The invention discloses a regional farmland water-saving irrigation and conventional irrigation classification and high-precision mapping method based on machine learning. The method comprises the following steps: S1, acquiring climate, geography, crops and remote sensing image information data related to a farmland in a research area; s2, selecting seven indexes including ELE, LS, ET, LAI, CI, DLST and LTH to represent water-saving irrigation (WI) and conventional irrigation (FI), performing binary segmentation by using a threshold method, obtaining classification of WI and FI under each index, and generating WI candidate samples; s3, WI and FI typical pixels are determined, and a training sample library is formed; s4, synthesizing a characteristic wave band as a classifier input characteristic; s5, based on the training sample library and classifier input features, training a classifier through a machine learning algorithm; and S6, applying the classifier to the characteristic wave band in the research area to obtain the classification and classification graph of the farmland WI and FI in the research area.

Description

Technical Field

[0001] The present invention relates to the field of agricultural information technology, and more specifically to a method for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning and high-precision mapping. Background Art

[0002] High-precision regional farmland irrigation information, such as which areas use water-saving irrigation and which areas use traditional flooding, is crucial for farmland management planning, irrigation system design, irrigation water conservation, scientific water resource allocation, farmland greenhouse gas emission estimation, and carbon sequestration and emission reduction program design. However, this information currently relies primarily on field surveys and statistics, which are costly in terms of manpower, material resources, and financial resources. Furthermore, this method only applies to official administrative participation and a single administrative unit and cannot automatically obtain data over large areas. Therefore, statistical information currently only exists for a single administrative unit. A "single map" of farmland irrigation information with high spatial precision (e.g., 500-meter resolution) at county, municipal, provincial, and national scales is currently lacking. This lack of fundamental information hinders the optimization of farmland irrigation system design, water conservation management, and the accurate estimation of farmland greenhouse gas emissions.

[0003] Therefore, new technologies are needed to at least partially address the above limitations of the existing technologies. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention proposes a method for regional farmland water-saving irrigation and conventional irrigation classification and high-precision mapping based on machine learning, aiming to solve the technical problem of how to quickly and automatically classify farmland irrigation systems (water-saving irrigation and conventional irrigation) over a large area (such as a prefecture-level city, province or country) and map them with high precision (resolution 500 meters).

[0005] More specifically, according to one aspect of the present invention, a method for classifying and mapping regional farmland water-saving irrigation and conventional irrigation with high precision based on machine learning is provided, comprising the following steps:

[0006] S1: Obtain climate, geography, crops and remote sensing image information related to farmland in the study area;

[0007] S2: Based on the data obtained in S1, seven indicators, including altitude (ELE), surface slope (LS), actual evaporation (ET), leaf area index (LAI), planting intensity (CI), daily surface temperature difference (DLST), and the interval between transplanting and heading (LTH), were selected to characterize water-saving irrigation (WI) and conventional irrigation (FI). The threshold method was used for binary segmentation to obtain the classification of water-saving irrigation (WI) and conventional irrigation (FI) under each indicator, and multiple groups of corresponding water-saving irrigation (WI) candidate samples were generated;

[0008] S3: Further screen the candidate samples to determine the typical pixels of WI and FI, thereby forming a training sample library;

[0009] S4: Using remote sensing image data to synthesize characteristic bands as input features, the input features include VH band, NDVI (normalized difference vegetation index), EVI (enhanced vegetation index), GI (greenness index), LSWI1 (surface water index 1) and LSWI2 (surface water index 2);

[0010] S5: Based on the training sample library in step S3, train a classifier using a machine learning algorithm; and

[0011] S6: Apply the classifier to the characteristic bands in the study area to obtain the classification of conventional irrigation / water-saving irrigation of farmland in the study area and the classification map.

[0012] According to an embodiment of the present invention, the resolution of the classification map is below 500 meters, and the farmland is selected from wheat fields, corn fields, soybean fields and rice fields.

[0013] According to the implementation scheme of the present invention, in S2, the statistical values ​​of farmland irrigation water consumption at the municipal level in the study area are used to determine the areas of water-saving irrigation (WI) and conventional irrigation (FI) at each municipal level, and each indicator (a total of 7) is used to try different thresholds for classification until the area of ​​water-saving irrigation (WI) in the classification result is consistent with the statistical data at the municipal level. The threshold at this time is the optimal threshold for the indicator at the municipal level classification.

[0014] According to an embodiment of the present invention, S3 further includes filtering the farmland range in the study area with a 3×3 average value filter to remove the edge of the farmland and avoid errors caused by pixel mixing.

[0015] According to an embodiment of the present invention, in S4, the VH band is synthesized using Sentinel-1SAR data, including preprocessing the synthetic aperture radar data according to the steps of boundary noise removal, thermal noise removal, radiation calibration, and terrain correction to obtain the median backscatter coefficient; and MODIS data is used to calculate NDVI, EVI, GI, LSWI1 and LSWI2.

[0016] According to an embodiment of the present invention, in S5, the machine learning algorithm is a random forest (RF) method, and the number of decision trees is 150 to 250, preferably 200.

[0017] According to an embodiment of the present invention, step S6 includes random sampling, RF training, and irrigation schedule classification. To reduce the impact of a single random sampling on classifier training, the continuous process of random sampling, RF training, and irrigation schedule classification is repeated multiple times, for example, 10 times. The probability of each grid being identified as WI in these multiple classifications is calculated, and grids with a probability of being identified as WI exceeding 50% are determined as WI farmland, and the other grids are determined as FI farmland.

[0018] According to an embodiment of the present invention, the method for regional farmland water-saving irrigation and conventional irrigation classification and high-precision mapping based on machine learning further includes step S7, evaluating the accuracy of irrigation system classification.

[0019] According to another aspect of the present invention, there is also provided a device for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning, comprising:

[0020] The information acquisition module is used to obtain climate, geography, crops and remote sensing image information related to farmland in the study area;

[0021] The water-saving irrigation (WI) candidate sample production module selects seven indicators, including altitude (ELE), surface slope (LS), actual evaporation (ET), leaf area index (LAI), planting intensity (CI), daily surface temperature difference (DLST), and the interval between transplanting and heading (LTH), to characterize water-saving irrigation (WI) and conventional irrigation (FI). For each indicator, a threshold method is used for binary segmentation to obtain the classification of water-saving irrigation (WI) and conventional irrigation (FI) under each indicator, and generate multiple groups of corresponding water-saving irrigation (WI) candidate samples;

[0022] The training sample library formation module further screens the candidate samples and determines the typical pixels of WI and FI, thereby forming a training sample library;

[0023] Feature band synthesis module: uses remote sensing image data to synthesize feature bands as input features, including VH band, NDVI (normalized vegetation index), EVI (enhanced vegetation index), GI (greenness index), LSWI1 (surface water index 1) and LSWI2 (surface water index 2);

[0024] Classifier training module: Based on the training sample library and input features, the classifier is trained through machine learning algorithms; and

[0025] Classification and mapping module: The classifier is applied to the characteristic bands in the study area to obtain the classification and classification map of conventional irrigation / water-saving irrigation of farmland in the study area.

[0026] According to an embodiment of the present invention, the apparatus further comprises an irrigation schedule classification accuracy assessment module.

[0027] According to another aspect of the present invention, there is also provided an electronic device, characterized in that it includes: a memory and one or more processors;

[0028] The memory is used to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in the present invention.

[0029] This method uses proxy indicators and their thresholds for irrigation schedules to quickly and automatically acquire a sufficient number of Wi-Fi training sample points. For example, seven environmental variables are selected as proxy indicators for irrigation schedules. For each proxy indicator, a binary segmentation method is applied using a thresholding method to obtain Wi-Fi classifications for each proxy indicator. The classification results are then combined, and the most common denominator across all classifications is used as the Wi-Fi training sample point.

[0030] This method achieves automatic classification and high-precision spatial mapping of farmland irrigation regimes. For example, it selects six bands or remote sensing indices as input features: the Sentienl-1VV band, the HH band, NDVI, EVI, GI, LSWI1, and LSWI2. It then uses a random forest (RF) method to train a classifier, which is then applied to the characteristic bands of rice fields across China to create a classification map of rice field irrigation regimes (water-saving irrigation and conventional irrigation).

[0031] Compared with the existing technology, the present invention realizes automatic classification of large-scale farmland irrigation systems and high-precision spatial mapping based on available data resources, which is of great significance to farmland irrigation information monitoring and supervision, agricultural irrigation system planning, agricultural water-saving irrigation improvement, and farmland greenhouse gas estimation and emission reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The same reference numerals in the accompanying drawings designate the same or similar components or parts. The objects and features of the present invention will become more apparent from the following description taken in conjunction with the accompanying drawings, in which:

[0033] Figure 1 1 is a flow chart of a method for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning and high-precision mapping according to an embodiment of the present invention;

[0034] Figure 2 2. It is a schematic structural diagram of a device for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning according to one embodiment of the present invention;

[0035] Figure 3is a schematic structural diagram of an electronic device according to an embodiment of the present invention; and

[0036] Figure 4 1 is a graph comparing the estimated and statistical areas of WI rice fields at the municipal level (a) and provincial level (b) according to one embodiment of the present invention. DETAILED DESCRIPTION

[0037] To clearly illustrate the solutions of the present invention, preferred embodiments are given below and described in detail with reference to the accompanying drawings. The following description is merely illustrative in nature and is not intended to limit the application or use of the present invention.

[0038] It should be understood that the machine learning algorithm (model) cited in the present invention is itself known, such as the various sub-modules, various parameters, operating mechanisms, training processes, etc. of the model. Therefore, the present invention focuses on how to use machine learning to determine the classification and mapping process of regional farmland water-saving irrigation and conventional irrigation in the study area.

[0039] Figure 1 This is a flow chart of a method for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning and high-precision mapping according to one embodiment of the present invention. Figure 1 As shown, the regional farmland water-saving irrigation and conventional irrigation classification and high-precision mapping method based on machine learning of the implementation plan mainly includes four parts: forming a training sample library, synthesizing features, classification, and accuracy evaluation.

[0040] More specifically, the method for regional farmland water-saving irrigation and conventional irrigation classification and high-precision mapping based on machine learning according to an embodiment of the present invention may include the following steps:

[0041] First, collect information related to farmland within the study area, including climate, geography, crops, and remote sensing imagery. For example, municipal-level agricultural irrigation water statistics, topography, elevation, soil data, agricultural production data, and relevant meteorological and remote sensing data can be collected. Geographic data can include topography, elevation, slope, and crop distribution. Soil data can include soil type, albedo, soil evaporation limit, runoff curve number, soil drainage rate, soil water limit or wilting point, field capacity, and saturated water content. Meteorological data can include daily solar radiation, daily maximum and minimum temperatures, daily rainfall, daily relative humidity, and daily average wind speed. Agricultural production data can include irrigation, variety, planting density, and sowing methods. Remote sensing data can include Sentinel-1 SAR and MODIS remote sensing data.

[0042] Based on the data obtained, seven proxy indicators were selected: elevation (ELE), surface slope (LS), actual evaporation (ET), leaf area index (LAI), planting intensity (CI), diurnal land surface temperature difference (DLST), and the interval between transplanting and heading (LTH) to characterize water-saving irrigation (WI) and conventional irrigation (long-term flood irrigation (FI). In other words, these seven proxy indicators of irrigation schedules were used to generate sample points for water-saving irrigation and conventional irrigation. The likelihood of water-saving irrigation on a farmland was negatively correlated with ELE, LS, ET, and LAI, and positively correlated with CI, DLST, and LTH.

[0043] For each of the above indicators, binary segmentation is performed using the threshold method to obtain the WI / FI classification under each indicator, that is, to generate corresponding multiple groups of water-saving irrigation (WI) candidate samples.

[0044] More specifically, using a thresholding method, pixels with lower ELE, LS, ET, and LAI, or higher CI, DLST, and LTH, are identified as WI candidates. For each proxy indicator, the binary threshold at each prefecture-level can be determined, for example, based on prefecture-level statistics on water-saving irrigation area and irrigation water consumption. For example, using ELE as an example, the pixel with the lowest elevation within the rice field is first identified as a WI candidate. The total area of ​​the WI candidate pixels is then compared to the target hectares provided by the statistical data. If the total area is less than the target area, the pixel with the next lowest elevation is identified as a WI candidate. This step is repeated until the total area equals or exceeds the target area for the first time. Using seven variables, seven sets of WI candidates are generated.

[0045] Afterward, candidate samples are further screened to identify representative pixels. For example, the consistency of the classification results for each of the seven proxy indicators is used, and the samples with the highest consistency are selected as WI / FI training sample points. More specifically, if a pixel is classified as WI using all seven proxy indicators, it is determined to be the final WI representative pixel; if a pixel is classified as non-WI farmland using all seven proxy indicators, it is determined to be a FI representative pixel. In some provinces, there may not be enough pixels to be classified as both WI and non-WI farmland in the initial identification. In this case, the restriction can be gradually relaxed from seven candidate samples to six, five, and four. In rare cases, if the number of pixels that meet all four candidate samples is still insufficient, pixels from the nearest neighboring province can be used to supplement the sample. In addition, a 3×3 average filter can be used to filter the farmland area to remove the edge of the farmland and avoid errors caused by pixel mixing. In each province, at least 2000 WI typical pixels and 2000 FI typical pixels can be generated.

[0046] Input features are another key factor influencing the classification of farmland irrigation schedules. Research has found that selecting six bands or remote sensing indices as input features, including the VH band, NDVI (Normalized Difference Vegetation Index), EVI (Enhanced Vegetation Index), GI (Greenness Index), LSWI1 (Land Water Index 1), and LSWI2 (Land Water Index 2), can achieve good classification results. Therefore, we synthesize characteristic bands from remote sensing imagery data as input features for the machine learning model.

[0047] More specifically, the VH band can be synthesized using Sentinel-1 SAR data, including preprocessing the synthetic aperture radar data according to steps such as boundary noise removal, thermal noise removal, radiometric calibration, and terrain correction to obtain the backscatter coefficient (i.e., the synthesized VH band value) as an input feature for classification. It should be understood that the above-mentioned operation processing is well known in the art (for example, see: Singha, M., Dong, J., Zhang, G., Xiao, X., 2019. High resolution paddy rice maps in cloud-prone Bangladesh and Northeast India using Sentinel-1 data. Scientific Data 6. https: / / doi.org / https: / / doi.org / 10.1038 / s41597-019-0036-3). Similar processing can be performed on multiple years of data to obtain the median value. For example, in this study, taking rice as an example, we synthesized the median backscatter coefficient every 20 days from March to August 2018, which is the main rice growing season in my country.

[0048] The other five indices, NDVI, EVI, GI, LSWI1, and LSWI2, can be calculated using MODIS data. Before synthesizing features, the "StateQA" field can be used to remove cloud cover. Afterwards, the masked MODIS data is used to calculate the above six indices, NDVI, EVI, GI, LSWI1, and LSWI2, using the following formulas (1 to 5): nir , ρ red , ρ blue , ρ green , ρ swir1 and ρ swir2 These are the surface reflectance values ​​for the near-infrared, red, blue, green, shortwave infrared 1, and shortwave infrared 2 bands received by the MODIS sensor. Finally, we synthesized the median values ​​for the same periods from 2016 to 2021 into feature bands. As shown in the figure, taking rice as an example, there are 23 MODIS product periods during the rice growing season (March to August) each year, resulting in a total of 115 MODIS composite features (23 periods per index x 5 indices).

[0049]

[0050] Next, based on the training sample library in step S3, a classifier is trained using a machine learning algorithm. Here, we use the random forest (RF) method to train the classifier. Of course, other suitable machine learning models can also be used.

[0051] It's important to note that the number of training samples significantly impacts classification results. Too few samples may not adequately reflect the overall characteristics; too large a sample size requires a significant amount of computational time. To balance classification accuracy and computational efficiency, for most provinces, we randomly sampled 2,000 typical pixels (1,000 each of WI and FI) from each province's sample library. For provinces with smaller paddy fields, such as Ningxia, Xinjiang, and Inner Mongolia, we randomly sampled 1,000 WI and 1,000 FI typical pixels from the province's agricultural region. Preliminary experiments found that a range of 150 to 2,500 decision trees yielded acceptable classification results at a relatively fast rate. Therefore, for example, the number of decision trees for RF can be set to 200.

[0052] After generating the RF classifier, we applied it to characteristic bands across farmland nationwide to create a classification map for irrigation mapping. To reduce the impact of single random sampling on classifier training, we repeated the process of random sampling, RF training, and irrigation schedule classification 10 times, calculating the probability of each grid being identified as WI across these 10 classifications. Finally, grids with a probability of WI exceeding 50% were designated as WI rice fields, while all other grids were designated as FI rice fields. This resulted in a classification map for conventional irrigation and water-saving irrigation for the farmland in the study area.

[0053] Finally, the accuracy of the irrigation regime classification map was evaluated.

[0054] The accuracy of the classification results can be evaluated in the following two ways. First, the estimated WI area in the classification map can be compared with the water-saving irrigation area in the statistical yearbook. For example, a regression model between the estimated WI area and the actual area can be established at the provincial and municipal levels. The coefficient of determination (R 2 ) to evaluate the accuracy of its estimated area. Secondly, the accuracy of the classified map can be evaluated using ground sample points.

[0055] Figure 2 Schematic diagram of a device for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning according to an embodiment of the present invention. Figure 2As shown, the device includes: an information acquisition module 210 for acquiring climate, geography, crops and remote sensing image information related to farmland in the study area; a water-saving irrigation (WI) candidate sample production module 220, which selects seven indicators, including altitude (ELE), surface slope (LS), actual evaporation (ET), leaf area index (LAI), planting intensity (CI), surface daily temperature difference (DLST) and transplanting period to heading period interval (LTH), to characterize water-saving irrigation (WI) and conventional irrigation (FI) based on the data obtained by S1, and uses the threshold method to perform binary segmentation to obtain the water-saving irrigation (WI) and conventional irrigation (FI) under each indicator. (FI) classification, and generate corresponding multiple groups of water-saving irrigation (WI) candidate samples; training sample library formation module 230, further screening the candidate samples, determining WI and FI typical pixels, thereby forming a training sample library; feature band synthesis module 240: using remote sensing image data to synthesize feature bands as input features; classifier training module 250, based on the training sample library and input features, training the classifier through machine learning algorithm; classification and mapping module 260: applying the classifier to the feature bands in the study area, thereby obtaining the classification and classification map of conventional irrigation / water-saving irrigation of farmland in the study area; and irrigation system classification accuracy assessment module 270.

[0056] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention is shown in FIG. Figure 3 As shown, the electronic device includes a processor 310, a memory 320, an input device 330 and an output device 340; the number of processors 310 in the electronic device can be one or more. Figure 3 In the figure, a processor 310 is used as an example; the processor 310, memory 320, input device 330 and output device 340 in the electronic device can be connected via a bus or other means. Figure 3 The bus connection is taken as an example.

[0057] Memory 320, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the building electrical energy efficiency evaluation method in the embodiments of the present invention (e.g., information acquisition module 210; water-saving irrigation candidate sample generation module 220; training sample library formation module 230; characteristic band synthesis module 240; classifier training module 250; classification and mapping module 260; and irrigation schedule classification accuracy assessment module 270). By running the software programs, instructions, and modules stored in memory 320, processor 310 executes various functional applications and data processing of the electronic device, thereby implementing the aforementioned building electrical energy efficiency evaluation method.

[0058] The memory 320 may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function; the data storage area may store data created based on the use of the terminal, etc. Furthermore, the memory 320 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 320 may further include a memory remotely located relative to the processor 310, and such remote memory may be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof. The input device 330 may be used to receive input digital or character information and to generate key signal inputs related to user settings and function control of the electronic device. The output device 340 may include a display device such as a display screen.

[0059] Example

[0060] This case uses rice as an example to conduct a verification experiment on my country's rice irrigation system using the method of the present invention, as follows:

[0061] The validation data for this experiment comes from the Global Geo-Referenced Field Photo Library, a website where volunteers from around the world share field photos with geographic coordinates. We collected photos of all rice fields in China from this website and determined their irrigation regimes based on the flooding status of the rice fields in the photos. Specifically, the water levels in the rice fields were divided into three categories: non-flooded, shallow water, and deep water, depending on whether there was a continuous body of water in the rice field and whether the continuous water body submerged the rice roots. Several photos were collected for each rice field, so there would be multiple observations of the water level. If a field was mainly shallow water or non-flooded, we classified it as a WI rice field, otherwise it was classified as a FI rice field. A total of 4,828 field photos were identified, and the irrigation regimes of 376 fields were collected. These fields will serve as ground reference points to verify the accuracy of the classification results.

[0062] Elevation data were obtained from the National Tibetan Plateau Data Center, and surface slope data were obtained from the National Earth System Science Data Center. Both are generated based on the NASA ASTER global digital elevation model. LAI data are from the MCD15A3H product, which provides LAI values ​​at 500 m every four days. 30 m resolution CI data are from Liu et al. (Liu, C., Zhang, Q., Tao, S., Qi, J., Ding, M., Guan, Q., Wu, B., 2020. A new framework to map fine-resolution cropping intensity across the globe: Algorithm, validation, and implication. Remote Sensing of Environment 251.) https: / / doi.org / https: / / doi.org / 10.1016 / j.rse.2020.112095 .). LTH data are generated based on the MCD12Q1 product, and rice LTH is estimated using the date of each phenological period. LST data are generated based on the MOD11A2 product, and the diurnal temperature difference is calculated by subtracting the nighttime land surface temperature from the daytime land surface temperature. ET data are from the study of Zheng et al. (Zheng, C., Jia, L., Hu, G., 2022. Global land surface evapotranspiration monitoring by ET Monitor model driven by multi-source satellite earth observations. Journal of Hydrology 613. https: / / doi.org / https: / / doi.org / 10.1016 / j.jhydrol.2022.128444), and the average ET is calculated. In this study, all data were resampled to the GCS_WGS_1984 coordinate system with a resolution of 500 m by averaging.

[0063] Using a threshold segmentation method, we used seven environmental variables—elevation, surface slope, actual evaporation, leaf area index, planting intensity, daily surface temperature difference, and the interval between transplanting and heading—as proxy indicators to generate seven candidate samples for conventional irrigation / water-saving irrigation classification. Further screening and filtering were then performed to identify representative pixels for conventional irrigation / water-saving irrigation in each province or major rice-producing region. A classifier was then trained using a machine learning algorithm (Random Forest) and applied to characteristic bands across China, resulting in a 500-meter-resolution classification map of conventional / water-saving irrigation rice in China. The accuracy of the irrigation mapping classification results was also evaluated using statistical yearbook data.

[0064] The results of the WI and FI classifications of rice paddies in China show that WI irrigation is predominant in North and Northwest China, where precipitation is limited; FI irrigation is predominant in East China, where precipitation is abundant; and a coexistence of WI and FI irrigation methods in other regions. Nationwide, the proportion of FI rice paddies decreases from the southeast to the northwest, mirroring the spatial trend in irrigation water use. These findings suggest that the choice of rice paddy irrigation system depends on local water availability, with a higher proportion of WI rice paddies in regions with limited irrigation water.

[0065] According to the map, the area of ​​WI rice fields in China in 2019 was approximately 1.23×10 7 ha, accounting for 49.3% of the total rice area. Our estimated area of ​​WI rice fields is slightly higher than the WI area calculated in the statistical yearbook (1.12×10 7 The estimated rice areas in Northeast China, North China, Northwest China, Southwest China, South China, and East China are 2.59 Mha, 1.14 Mha, 0.85 Mha, 2.15 Mha, 3.47 Mha, and 2.04 Mha, respectively. WI is more prevalent in double-season rice than in single-season rice. In single-season rice, WI accounts for approximately 8.25 Mha, or 46.6% of the total single-season rice area, while in double-season rice, WI accounts for 4.02 Mha, or 56.1% of the total double-season rice area.

[0066] Figure 4 The regression relationship between the estimated WI rice paddy area and the statistical area at the municipal or provincial level is shown. The provincial regression analysis shows that the R 2 The city-level regression analysis shows that the R 2 It is about 0.88, and the regression slope is 0.99, indicating that the estimated WI area is close to the actual area, which proves the effectiveness and applicability of the method of the present invention.

[0067] Based on available data resources, the present invention realizes automatic classification of large-scale farmland irrigation systems and spatial high-precision mapping, which is of great significance to farmland irrigation information monitoring and supervision, agricultural irrigation system planning, agricultural water-saving irrigation improvement, and farmland greenhouse gas estimation and emission reduction.

[0068] Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the specific implementation methods of this specification should not be understood as limiting the present invention.

Claims

1. A method for classifying and mapping regional farmland water-saving irrigation and conventional irrigation with high precision based on machine learning, comprising the following steps: S1: Obtain climate, geography, crops and remote sensing image information related to farmland in the study area; S2: Based on the data obtained in S1, seven indicators, including altitude (ELE), surface slope (LS), actual evaporation (ET), leaf area index (LAI), planting intensity (CI), daily surface temperature difference (DLST), and the interval between transplanting and heading (LTH), were selected to characterize water-saving irrigation (WI) and conventional irrigation (FI). For each indicator, a threshold method was used for binary segmentation to obtain the classification of water-saving irrigation (WI) and conventional irrigation (FI) under each indicator, and generate multiple groups of corresponding water-saving irrigation (WI) candidate samples; S3: Further screen the candidate samples to determine the typical pixels of WI and FI, thereby forming a training sample library; S4: Using remote sensing image data to synthesize characteristic bands as input features, the input features include VH band, NDVI (normalized difference vegetation index), EVI (enhanced vegetation index), GI (greenness index), LSWI 1 (surface water index 1) and LSWI2 (surface water index 2); S5: Based on the training sample library in step S3 and the input features in step S4, a classifier is trained using a machine learning algorithm; and S6: Apply the classifier to the characteristic bands in the study area to obtain the classification of conventional irrigation / water-saving irrigation of farmland in the study area and the classification map.

2. The method according to claim 1, characterized in that The resolution of the classification map is below 500 meters, and the farmland is selected from wheat fields, corn fields, soybean fields and rice fields.

3. The method according to claim 1, characterized in that In S2, the water-saving irrigation area or irrigation water consumption calculated at the municipal level within the study area was used as the dichotomy threshold.

4. The method according to claim 1, wherein S3 also includes filtering the farmland range in the study area with a 3×3 average filter to remove the edge of the farmland and avoid errors caused by pixel mixing.

5. The method according to claim 1, wherein In S4, the VH band is synthesized using Sentinel-1SAR data, which includes preprocessing the synthetic aperture radar data according to the steps of boundary noise removal, thermal noise removal, radiometric calibration, and terrain correction to obtain the median backscatter coefficient; Calculate NDVI, EVI, GI, LSWI 1, and LSWI2 using MODIS data.

6. The method according to claim 1, characterized in that In S5, the machine learning algorithm is a random forest (RF) method, and the number of decision trees is 150 to 250, preferably 200.

7. The method according to claim 1, characterized in that Step S6 includes random sampling, RF training, and irrigation schedule classification. To reduce the impact of a single random sampling on classifier training, this continuous process of random sampling, RF training, and irrigation schedule classification is repeated multiple times, for example, 10 times. The probability of each grid being identified as WI in these multiple classifications is calculated. Grids with a probability of being identified as WI exceeding 50% are identified as WI farmland, and the remaining grids are identified as FI farmland.

8. The method according to claim 1, characterized in that The method further includes step S7, evaluating the accuracy of the irrigation schedule classification.

9. A device for classifying and mapping regional farmland water-saving irrigation and conventional irrigation based on machine learning, characterized in that: include: The information acquisition module is used to obtain climate, geography, crops and remote sensing image information related to farmland in the study area; The water-saving irrigation (WI) candidate sample production module selects seven indicators, including altitude (ELE), surface slope (LS), actual evaporation (ET), leaf area index (LAI), planting intensity (CI), daily surface temperature difference (DLST), and the interval between transplanting and heading (LTH), to characterize water-saving irrigation (WI) and conventional irrigation (FI). For each indicator, a threshold method is used for binary segmentation to obtain the classification of water-saving irrigation (WI) and conventional irrigation (FI) under each indicator, and generate multiple groups of corresponding water-saving irrigation (WI) candidate samples. The training sample library formation module further screens the candidate samples and determines the typical pixels of WI and FI, thereby forming a training sample library; Feature band synthesis module: uses remote sensing image data to synthesize feature bands as input features, including VH band, NDVI (normalized vegetation index), EVI (enhanced vegetation index), GI (greenness index), LSWI 1 (surface water index 1) and LSWI2 (surface water index 2); Classifier training module: Based on the training sample library and input features, the classifier is trained through machine learning algorithms; as well as Classification and mapping module: applying the classifier to the characteristic bands in the study area, thereby obtaining the classification and classification map of conventional irrigation / water-saving irrigation of farmland in the study area; wherein, the device also preferably includes an irrigation system classification accuracy assessment module.

10. An electronic device, characterized in that: include: memory and one or more processors; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 10.

Images