Winter wheat drought-resistant water-saving variety identification and selection method, system and device

Abstract

The invention provides a method, a system and a device for identifying and selecting drought-resistant and water-saving winter wheat varieties. The method comprises the following steps: acquiring multi-source and multi-time-sequence wheat remote sensing data; non-image data is preprocessed; dividing the preprocessed non-image data into a non-image training set and a non-image test set which are used for biomass prediction, LAI and yield prediction, and dividing the image data into an image training set and an image test set which are used for lodging detection; training a biomass prediction model, a yield prediction model and an LAI prediction model by using the non-image training set and the non-image test set; training a lodging detection model by using the image training set and the image test set; the trained biomass prediction model, yield prediction model and lodging detection model are utilized to predict the biomass, LAI, yield and lodging conditions of the selected varieties; and determining dominant varieties according to biomass, plant height, LAI, yield and lodging conditions. According to the invention, high-throughput and high-precision variety identification and selection can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of variety selection, and in particular relates to a method, system and device for selecting drought-resistant and water-saving winter wheat varieties. Background Art

[0002] As one of the world's most important food crops, wheat is of undeniable importance. Due to the uneven temporal and spatial distribution of water resources and precipitation in some regions, selecting drought-resistant and water-saving varieties is an effective way to increase wheat yields and ensure food security. However, traditional methods for selecting drought-resistant and water-saving varieties have significant limitations: At the data acquisition level, manual selection is inefficient and time-consuming, severely restricting the selection and promotion of superior varieties. Existing methods rely primarily on manual measurement of phenotypic parameters such as biomass and leaf area index, which is not only time-consuming and labor-intensive, but also destructive sampling that results in loss of test materials. In large-scale, multi-variety comparative trials, manual measurement errors are large, standards are difficult to unify, and the workload is enormous. Furthermore, dynamic monitoring during the growth period is impossible. Although satellite remote sensing technology can provide large-scale monitoring, its insufficient spatial resolution makes it difficult to distinguish subtle differences between varieties.

[0003] At the variety selection level: First, the selection cycle is long. The current drought resistance index evaluation method relies on the measured yield data after harvest. It takes an entire growing season from sowing to the output of selection results, which seriously restricts the promotion process of excellent varieties; second, the evaluation dimension is relatively single. The existing selection and evaluation methods are mostly based on yield indicators, ignoring key factors such as crop water consumption, resulting in high-yield varieties being selected but water-saving performance cannot be guaranteed; third, there is little research on the selection method of winter wheat drought-resistant and water-saving varieties based on UAV remote sensing, which restricts the development of high-throughput drought-resistant and water-saving selection of winter wheat. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, system and device for selecting drought-resistant and water-saving winter wheat varieties, which can realize high-throughput variety selection through multi-source and multi-time series data.

[0005] The present invention discloses a method for selecting drought-resistant and water-saving winter wheat varieties, comprising the following steps: S1. Acquire multi-source and multi-time series wheat remote sensing data, wherein the wheat remote sensing data includes image data and non-image data; S2. preprocessing the non-image data; S3. Dividing the preprocessed non-image data into non-image training sets and non-image test sets for biomass prediction, LAI, and yield prediction, and dividing the image data into image training sets and image test sets for lodging detection; S4, using the non-image training set and the non-image test set to train a biomass prediction model, a yield prediction model, and a LAI prediction model; S5. Training a lodging detection model using the image training set and the image test set; S6. Predicting the biomass, LAI, yield, and lodging of the candidate varieties using the trained biomass prediction model, yield prediction model, and lodging detection model; S7. Determine the dominant variety among the candidate varieties based on the biomass, plant height, LAI, yield and lodging condition.

[0006] Furthermore, in S1, the non-image data includes vegetation index data, image texture data, vegetation coverage data, variety data, canopy temperature data, meteorological data, and agronomic data; The image data includes plot images of wheat during its critical growth period.

[0007] Furthermore, S2 includes the following steps: S201, performing dummy variable processing on the text data in the non-image data; S202: Perform normalization processing on the numerical data in the non-image data.

[0008] Furthermore, S3 includes the following steps: S301. For biomass prediction and LAI prediction, n-1 replicates of each variety are used as training sets, and the remaining 1 replicate is used as a test set; S302. For yield prediction, data is divided into two methods: fixed-period multi-source data and fixed-data source multi-period data. N-1 replicates of each variety are used as training sets, and the remaining replicate is used as a test set. S303. For lodging detection, the training set and the test set are divided into a ratio of 7:3.

[0009] Furthermore, S4 includes the following steps: S401, setting hyperparameter ranges, and using the non-image training set and non-image test set for biomass prediction and LAI prediction and the random forest model to train the biomass prediction model and the LAI prediction model respectively; S402 , setting a parameter range, and using the non-image training set and non-image test set for yield prediction and the random forest model to train a crop yield prediction model.

[0010] Furthermore, S5 includes the following steps: S501, performing color and direction data enhancement processing on the image training set and the image test set through the transforms module in pytorch; S502 , setting hyperparameters, and using a convolutional neural network and the enhanced image training set and image test set to train a lodging detection model.

[0011] Furthermore, S7 includes the following steps: S701. Calculating the drought resistance index of the candidate varieties based on the predicted yield value; S702, evaluating the water-saving performance of the candidate varieties based on the predicted values ​​of plant height, biomass, and LAI; S703, screening candidate varieties using drought resistance index, water conservation and lodging as screening indicators, and obtaining three groups of screening results respectively; S704. The candidate varieties that are present in the three groups of screening results are regarded as dominant varieties.

[0012] Furthermore, in S701, the calculation formula for the drought resistance index is: , , in, D GYI,p represents the drought resistance index under dry farming treatment, W GYI,p represents the drought resistance index under two irrigation treatments, GY D,p It represents the predicted yield of the candidate varieties under dry farming treatment, GY W,p It represents the predicted yield of the selected varieties under two irrigation treatments. MGY D,p represents the average predicted yield of the remaining varieties under dry farming treatment, MGY W,p represents the average predicted yield of the remaining varieties under the two irrigation treatments; In S702, the evaluation formula for water conservation is: , , , , in, PH D and PH D,mean They represent the predicted plant height of the selected varieties under dry farming treatment and the average predicted plant height of all varieties under dry farming treatment; Biomass D and Biomass D,meanThey represent the predicted biomass of the selected varieties under dryland farming treatment and the average predicted biomass of all varieties under dryland farming treatment; LAI D and LAI D,mean They represent the predicted LAI of the candidate varieties under dryland farming treatment and the average predicted LAI of all varieties under dryland farming treatment; LAI W and LAI W,mean represent the predicted LAI of the selected varieties under the two irrigation treatments and the average predicted LAI of all varieties under the two irrigation treatments, respectively.

[0013] The present invention also provides a winter wheat drought-resistant and water-saving variety selection system, comprising: An acquisition module, configured to acquire multi-source and multi-time series wheat remote sensing data, wherein the wheat remote sensing data includes image data and non-image data; A preprocessing module, configured to preprocess the non-image data; a data set partitioning module for partitioning the preprocessed non-image data into non-image training sets and non-image test sets for biomass prediction, LAI, and yield prediction, and for partitioning the image data into image training sets and image test sets for lodging detection; The first training module is used to train the biomass prediction model, the yield prediction model and the LAI prediction model using the non-image training set and the non-image test set, A second training module is used to train a lodging detection model using the image training set and the image test set; The prediction module is used to predict the biomass, LAI, yield and lodging of the candidate varieties using the trained biomass prediction model, yield prediction model and lodging detection model; a plant height calculation module, configured to calculate the plant height using the image data; The selection module is used to determine the dominant varieties of the candidate varieties based on the biomass, plant height, LAI, yield and lodging conditions.

[0014] The present invention also provides a device for selecting drought-resistant and water-saving winter wheat varieties, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the above method.

[0015] The beneficial effects brought by the present invention are: 1. This invention achieves precise estimation of phenotypic traits through the fusion of multi-source data. Specifically, it overcomes the limitations of traditional manual measurement methods and innovatively constructs a crop phenotypic trait estimation system based on multi-source remote sensing data. By integrating multidimensional remote sensing information such as visible light, multispectral, and thermal infrared, combined with advanced machine learning algorithms, comprehensive monitoring of key crop growth parameters is achieved. This not only significantly improves measurement efficiency but also, through the complementary advantages of multi-source data, significantly enhances the accuracy and reliability of estimation results, providing a solid phenotypic data foundation for variety evaluation.

[0016] 2. This invention constructs an intelligent yield prediction model system. Addressing the long processing times and low accuracy of traditional yield prediction methods, this invention establishes a new paradigm for yield prediction based on machine learning. By systematically optimizing a combination of machine learning algorithms and deeply exploring temporal growth characteristics, a prediction model with strong generalization capabilities is constructed. This model system can grasp the patterns of crop yield formation and achieve dynamic predictions within the growth period, providing a technical path for early variety evaluation.

[0017] 3. This invention has developed a comprehensive selection system for drought-resistant and water-saving winter wheat varieties based on remote sensing technology, innovatively proposing a new approach to variety selection based on the trinity of "phenotype, yield, and resistance." By deeply integrating remote sensing monitoring technology with agronomic evaluation standards, a comprehensive evaluation system encompassing multidimensional indicators such as drought resistance, water conservation, and stress tolerance has been constructed. This framework shifts from single-yield evaluation to comprehensive trait evaluation, significantly enhancing the scientific and comprehensive nature of variety selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention; Figure 2 Schematic diagram of data correction in the present invention (where a represents the correction diagram of canopy temperature obtained by multispectral and visible light remote sensing images, and b represents the correction diagram of predicted plant height and measured plant height); Figure 3 Schematic diagram of the biomass estimation results based on ensemble learning in the present invention (where a represents the schematic diagram for 2023, and b represents the schematic diagram for 2024); Figure 4 Schematic diagram of the LAI estimation results based on ensemble learning in the present invention (where a represents the schematic diagram for 2023, and b represents the schematic diagram for 2024); Figure 5 Schematic diagram of yield prediction results based on ensemble learning for fixed-period multi-source data in the present invention (where a represents a schematic diagram of the greening period, b represents a schematic diagram of the jointing period, c represents a schematic diagram of the heading period, d represents a schematic diagram of the flowering period, e represents a schematic diagram of the early grain filling period, and f represents a schematic diagram of the middle grain filling period); Figure 6 Schematic diagram of the yield prediction results of fixed data multi-time series based on ensemble learning in the present invention (wherein, a is a schematic diagram of remote sensing data, b is a schematic diagram of the combination of remote sensing data and canopy temperature data, c is a schematic diagram of the combination of remote sensing data, canopy temperature data, and variety data, d is a schematic diagram of the combination of remote sensing data, canopy temperature data, variety data, and meteorological data, e is a schematic diagram of the combination of remote sensing data, canopy temperature data, variety data, meteorological data, and texture data, and f is a schematic diagram of the combination of remote sensing data, canopy temperature data, variety data, meteorological data, texture data, and agronomic data); Figure 7 This is a schematic diagram of the results of the selection of drought-resistant and water-saving winter wheat varieties based on UAV remote sensing in 2023 in the present invention; Figure 8 This is a schematic diagram of the results of the selection of drought-resistant and water-saving winter wheat varieties based on drone remote sensing in 2024 in the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] like Figures 1 to 8 As shown, the present invention provides a method for selecting drought-resistant and water-saving winter wheat varieties, comprising the following steps: S1. Acquire multi-source, multi-time series wheat remote sensing data. This data comes from 48 winter wheat varieties in 2023 and 54 winter wheat varieties in 2024. Wheat remote sensing data includes image data and non-image data. Specifically, in S1, non-image data includes vegetation index data, image texture data, vegetation cover data, variety data, canopy temperature data, meteorological data, and agronomic data.

[0021] Vegetation index data includes multispectral vegetation index data and visible light vegetation index data. Image texture data includes six types of texture data extracted from seven bands of imagery, totaling 42 texture features. Variety data includes growth period length, lodging resistance, number of grains per ear, number of ears per mu, 1000-grain weight, and plant height. Canopy temperature data includes canopy temperature data and its derived indices. Meteorological data includes growing degree days and precipitation data. Agronomic data includes biomass data and LAI data.

[0022] The vegetation index data in this embodiment is obtained by using a drone to obtain multispectral and visible light remote sensing images at noon when the weather is clear and cloudless during the key growth period. Figure 2Band operations were then performed using GDAL in Python to obtain multispectral vegetation index and visible light vegetation index data. Image texture data was based on visible and multispectral imagery acquired by drones, and six image texture features were obtained using the Skimage library in Python. Vegetation cover data was obtained using the OTSU threshold segmentation method in OpenCV to segment soil and vegetation pixels. The proportion of vegetation pixels to the total number of pixels was then calculated to obtain vegetation cover data. Variety data was collected from an online wheat variety database, including data on the length of the growing season, lodging resistance, number of grains per ear, number of ears per mu, thousand-grain weight, and plant height. Plant height was calculated based on DSM (digital surface model) images acquired by drone remote sensing, by subtracting the DSM acquired during the growing season from the DSM of bare soil. Canopy temperature data was obtained by acquiring canopy temperature images from drones and batch processing and temperature measurement using the DJI Thermal SDK. A FLIR camera was used to capture canopy temperature during flight and collated with the batch temperature measurements to obtain actual canopy temperature data. Meteorological data is the actual meteorological data from the weather station, from which precipitation data and growing degree day data are calculated.

[0023] The image data includes images of wheat plots during key growth stages. Specifically, this includes image data for greening, jointing, heading, flowering, pre-grain filling, and grain filling. After stitching the field remote sensing images of key growth stages, ArcMap was used to create a .shp file containing the boundaries of each plot. This .shp file was applied to the remote sensing imagery to crop out images of each plot. Images from 2023 and 2024 were integrated to construct an image dataset, and the images were labeled, with lodging images marked as 1 and non-lodging images as 0.

[0024] S2. Preprocess the non-image data. Specifically, S2 includes the following steps: S201. Variety data is text data (winter, semi-winter, etc.), so it needs to be processed as a dummy variable in Python to obtain binary numerical data in the form of 0 and 1.

[0025] S202, vegetation index data, image texture data, vegetation coverage data, canopy temperature data, meteorological data and agronomic data are numerical data, so they need to be normalized using Min-max to ensure dimensional consistency.

[0026] The normalization formula is: .

[0027] Among them, x Min and x MaxRespectively represent the minimum and maximum values ​​of the feature in the dataset. Through normalization, the original data is mapped to the interval [0,1], thus achieving dimensional consistency between different features.

[0028] S3, dividing the pre-processed non-image data into a non-image training set and a non-image test set for biomass prediction, LAI, and yield prediction, and dividing the image data into an image training set and an image test set for lodging detection. Specifically, S3 includes the following steps: S301. For biomass and LAI prediction, n-1 replicates of each variety are used as the training set, and the remaining replicate is used as the test set. In this example, each variety has three water treatments and three replicates. First, the measured biomass and LAI values ​​are combined to match each data set. Then, two replicates of each variety are used as the training set, and the remaining replicate is used as the test set. This ensures that the model does not use all three replicates during training, thereby ensuring the model's extrapolation performance.

[0029] S302. For yield prediction, data is divided using fixed-period multi-source data and fixed-data source multi-period methods. n-1 sets of repetitions for each variety are used as the training set, and the remaining 1 set of repetitions is used as the test set. In this example, multi-source and multi-time series data can produce many combinations, so the data is divided using fixed-period multi-source data and fixed-data source multi-period methods. That is, in the first fixed-period combination method using different data sources, multi-source data for greening, jointing, heading, flowering, early grain filling, and mid-grain filling are combined. The data combinations are remote sensing, remote sensing + canopy temperature, remote sensing + canopy temperature + variety, remote sensing + canopy temperature + variety + meteorology, remote sensing + canopy temperature + variety + meteorology + texture, and remote sensing + canopy temperature + variety + meteorology + texture + agronomy. In the second combination of data from different periods of fixed data sources, a fixed data source was used, that is, the combination of the above six data sources, and different time series data were spliced, namely, greening, greening + jointing, greening + jointing + heading, greening + jointing + heading + flowering, greening + jointing + heading + flowering + early filling stage, greening + jointing + heading + flowering + early filling stage + middle filling stage.

[0030] S303: For lodging detection, the training and test sets were divided into a ratio of 7:3. After the dataset was divided, an image dataset iterator was built using PyTorch. The data and labels were combined and input into the PyTorch Dataloader to load the data. In the Dataloader, batch_size was set to 20, shuffle was set to True, and drop_last was set to True.

[0031] S4, using the non-image training set and the non-image test set to train the biomass prediction model, the yield prediction model, and the LAI prediction model. Specifically, S4 includes the following steps: S401, set the hyperparameter range according to Table 1, and use the non-image training set and non-image test set for biomass prediction and LAI prediction and the random forest model to train the biomass prediction model and LAI prediction model respectively. In this example, firstly, the highly sensitive features were selected by correlation analysis, and then the high sensitivity features were selected by reference. Figure 4 By setting up baseline models, such as linear regression models and four integrated machine learning models, such as XGBoost, RFR, Adaboost, and GBDT, the optimal machine learning method was determined based on two years of data. Finally, the optimal prediction model was determined to be RFR. Figure 3 and Figure 4 .

[0032] Table 1 Hyperparameter search ranges for different machine learning algorithms

[0033] It should be noted that the hyperparameter search range can be reasonably adjusted based on experience.

[0034] S402. Set the parameter range and use the non-image training set and non-image test set for yield prediction and the random forest model to train the crop yield prediction model. In this example, the yield prediction model was trained by two data partitioning methods and the machine learning training method mentioned in S401. It was determined that the optimal machine learning model was RFR, the optimal yield prediction period was the filling period, and the optimal multivariate data combination method was the combination of remote sensing, crown temperature and variety data. Figure 7 and Figure 8 .

[0035] S5. Train the lodging detection model using the image training set and the image test set. In this example, a convolutional neural network was built using the Pytorch module. Four convolutional networks of different depths were constructed: Conv1, Conv2, Conv3, and Conv4. Specifically, S5 also includes the following steps: S501. Use the transforms module in PyTorch to perform color and orientation data augmentation on the image dataset. In this example, the transforms module in PyTorch is used to resize the image. Then, the dataset is subjected to color jitter, a random horizontal flip with a probability of 0.1 using RandomHorizontalFlip, and a random vertical flip with a probability of 0.1 using RandomVerticalFlip. This ensures high generalization capabilities within a limited image dataset.

[0036] S502. Set hyperparameters and train the lodging detection model using a convolutional neural network combined with the augmented dataset. In this example, set batch size to 100, epochs to 100, and learning_rate to 0.01. Load the hyperparameters into the model and train the model.

[0037] After training, the lodging detection results were verified to determine the training accuracy. Accuracy, precision, and recall were used to measure the performance of convolutional neural networks with different convolution depths. The shallow convolution layer achieved the highest accuracy, precision, and recall, making it the optimal convolutional neural network model for lodging detection.

[0038] S6. Use the trained biomass prediction model, yield prediction model and lodging detection model to predict the biomass, LAI, yield and lodging of the candidate varieties.

[0039] S7, determining the dominant varieties of the candidate varieties based on biomass, plant height, LAI, yield and lodging. Further, S7 includes the following steps: S701. Calculate the drought resistance index of the candidate varieties based on the predicted yield. The calculation formula for the drought resistance index is: , , in, D GYI,p represents the drought resistance index under dry farming treatment, W GYI,p represents the drought resistance index under two irrigation treatments, GY D,p It represents the predicted yield of the candidate varieties under dry farming treatment, GY W,p It represents the predicted yield of the selected varieties under two irrigation treatments. MGY D,p represents the average predicted yield of the remaining varieties under dry farming treatment, MGY W,p Represents the average predicted yield of the remaining varieties under the two irrigation treatments.

[0040] S702. Evaluate the water-saving performance of the candidate varieties based on the predicted values ​​of plant height, biomass, and LAI. The water-saving performance evaluation formula is: , , , , in, PH D and PH D,mean They represent the predicted plant height of the participating varieties under drought-resistant farming treatment and the average predicted plant height of all varieties under drought-resistant farming treatment, respectively. Biomass D and Biomass D,mean They represent the predicted biomass of the selected varieties under dry farming treatment and the average predicted biomass of all varieties under dry farming treatment, respectively. LAI D and LAI D,mean They represent the predicted LAI of the candidate varieties under dryland farming treatment and the average predicted LAI of all varieties under dryland farming treatment, respectively. LAI W and LAI W,mean represent the predicted LAI of the selected varieties under the two irrigation treatments and the average predicted LAI of all varieties under the two irrigation treatments, respectively.

[0041] S703. The candidate varieties were screened using drought resistance index, water-saving property and lodging condition as screening indicators, and three groups of screening results were obtained.

[0042] S704, the candidate varieties that are present in the three sets of screening results are regarded as dominant varieties. In this example, the variety selection results were obtained over two years, and the selection results were highly consistent with the selection results based on the measured values. Figure 7 and Figure 8 .

[0043] Based on the same inventive concept, the present invention also provides a winter wheat drought-resistant and water-saving variety selection system, comprising: The acquisition module is used to obtain multi-source and multi-time series wheat remote sensing data, which includes image data and non-image data.

[0044] The preprocessing module is used to preprocess non-image data.

[0045] The dataset partitioning module is used to divide the preprocessed non-image data into non-image training sets and non-image test sets for biomass prediction, LAI and yield prediction, and to divide the image data into image training sets and image test sets for lodging detection.

[0046] The first training module is used to train a biomass prediction model, a yield prediction model and a LAI prediction model using a non-image training set and a non-image test set.

[0047] The second training module is used to train the lodging detection model using the image training set and the image test set.

[0048] The prediction module is used to predict the biomass, LAI, yield and lodging of the candidate varieties using the trained biomass prediction model, yield prediction model and lodging detection model.

[0049] The plant height calculation module is used to calculate the plant height using image data.

[0050] The selection module is used to determine the dominant varieties among the candidate varieties based on biomass, plant height, LAI, yield and lodging.

[0051] The present invention also provides a device for selecting drought-resistant and water-saving winter wheat varieties, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the above method.

[0052] The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for selecting drought-resistant and water-saving winter wheat varieties, characterized in that: The steps include: S1. Acquire multi-source and multi-time series wheat remote sensing data, wherein the wheat remote sensing data includes image data and non-image data; S2. preprocessing the non-image data; S3. Dividing the preprocessed non-image data into non-image training sets and non-image test sets for biomass prediction, LAI, and yield prediction, and dividing the image data into image training sets and image test sets for lodging detection; S4, using the non-image training set and the non-image test set to train a biomass prediction model, a yield prediction model, and a LAI prediction model; S5. Training a lodging detection model using the image training set and the image test set; S6. Predicting the biomass, LAI, yield, and lodging of the candidate varieties using the trained biomass prediction model, yield prediction model, and lodging detection model; S7. Determine the dominant variety among the candidate varieties based on the biomass, plant height, LAI, yield and lodging condition.

2. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 1, characterized in that: In S1, the non-image data includes vegetation index data, image texture data, vegetation coverage data, variety data, canopy temperature data, meteorological data and agronomic data; The image data includes plot images of wheat during its critical growth period.

3. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 1, characterized in that: S2 includes the following steps: S201, performing dummy variable processing on the text data in the non-image data; S202: Perform normalization processing on the numerical data in the non-image data.

4. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 1, characterized in that: S3 includes the following steps: S301. For biomass prediction and LAI prediction, n-1 replicates of each variety are used as training sets, and the remaining 1 replicate is used as a test set; S302. For yield prediction, data is divided into two methods: fixed-period multi-source data and fixed-data source multi-period data. N-1 replicates of each variety are used as training sets, and the remaining replicate is used as a test set. S303. For lodging detection, the training set and the test set are divided into a ratio of 7:

3.

5. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 1, characterized in that: S4 includes the following steps: S401, setting hyperparameter ranges, and using the non-image training set and non-image test set for biomass prediction and LAI prediction and the random forest model to train the biomass prediction model and the LAI prediction model respectively; S402 , setting a parameter range, and using the non-image training set and non-image test set for yield prediction and the random forest model to train a crop yield prediction model.

6. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 1, characterized in that: S5 includes the following steps: S501, performing color and direction data enhancement processing on the image training set and the image test set through the transforms module in pytorch; S502 , setting hyperparameters, and using a convolutional neural network and the enhanced image training set and image test set to train a lodging detection model.

7. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 1, characterized in that: S7 includes the following steps: S701. Calculating the drought resistance index of the candidate varieties based on the predicted yield value; S702, evaluating the water-saving performance of the candidate varieties based on the predicted values ​​of plant height, biomass, and LAI; S703, screening candidate varieties using drought resistance index, water conservation and lodging as screening indicators, and obtaining three groups of screening results respectively; S704. The candidate varieties that are present in the three groups of screening results are regarded as dominant varieties.

8. The method for selecting drought-resistant and water-saving winter wheat varieties according to claim 7, characterized in that: In S701, the calculation formula for the drought resistance index is: ， ， in, D GYI,p represents the drought resistance index under dry farming treatment, W GYI,p represents the drought resistance index under two irrigation treatments, GY D,p It represents the predicted yield of the candidate varieties under dry farming treatment, GY W,p It represents the predicted yield of the selected varieties under two irrigation treatments. MGY D,p represents the average predicted yield of the remaining varieties under dry farming treatment, MGY W,p represents the average predicted yield of the remaining varieties under the two irrigation treatments; In S702, the evaluation formula for water conservation is: ， ， ， ， in, PH D and PH D,mean They represent the predicted plant height of the selected varieties under dry farming treatment and the average predicted plant height of all varieties under dry farming treatment; Biomass D and Biomass D,mean They represent the predicted biomass of the selected varieties under dryland farming treatment and the average predicted biomass of all varieties under dryland farming treatment; LAI D and LAI D,mean They represent the predicted LAI of the candidate varieties under dryland farming treatment and the average predicted LAI of all varieties under dryland farming treatment; LAI W and LAI W,mean represent the predicted LAI of the selected varieties under the two irrigation treatments and the average predicted LAI of all varieties under the two irrigation treatments, respectively.

9. A system for selecting drought-resistant and water-saving winter wheat varieties, characterized in that: include: An acquisition module, configured to acquire multi-source and multi-time series wheat remote sensing data, wherein the wheat remote sensing data includes image data and non-image data; A preprocessing module, configured to preprocess the non-image data; a data set partitioning module for partitioning the preprocessed non-image data into non-image training sets and non-image test sets for biomass prediction, LAI, and yield prediction, and for partitioning the image data into image training sets and image test sets for lodging detection; The first training module is used to train the biomass prediction model, the yield prediction model and the LAI prediction model using the non-image training set and the non-image test set, A second training module is used to train a lodging detection model using the image training set and the image test set; The prediction module is used to predict the biomass, LAI, yield and lodging of the candidate varieties using the trained biomass prediction model, yield prediction model and lodging detection model; a plant height calculation module, configured to calculate the plant height using the image data; The selection module is used to determine the dominant varieties of the candidate varieties based on the biomass, plant height, LAI, yield and lodging conditions.

10. A device for selecting drought-resistant and water-saving winter wheat varieties, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the method according to claim 1.

Images