A rice drought resistance phenotype analysis method based on multi-task deep learning
By integrating multi-view, multi-temporal image acquisition and physiological spatial alignment techniques through multi-task deep learning methods, the problems of feature mixing and insufficient dynamic modeling in rice drought resistance phenotype analysis were solved, achieving accurate analysis of rice drought resistance phenotypes and improving the effectiveness of genotype association analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG AGRI UNIV
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies are insufficient to accurately analyze the complex phenotypic characteristics of rice under drought stress, and deep learning models are sensitive to light fluctuations and environmental noise, lacking dynamic modeling capabilities, resulting in insufficient effectiveness of phenotypic-genotypic association analysis.
By employing a multi-task deep learning approach, integrating multi-view, multi-temporal image acquisition, deep feature decoupling, and physiological spatial alignment techniques, color and morphological physiological branches are constructed. Combined with temporal constraints and genome-wide association analysis, drought level classification and agronomic trait regression are achieved, thereby improving the comprehensiveness of phenotypic analysis and the efficiency of gene mining.
This approach enables precise analysis of drought-resistant phenotypes in rice, improves the comprehensiveness and stability of phenotype analysis, enhances the correlation between genotype and phenotype, and improves the accuracy of drought-resistant genetic loci detection and the reliability of candidate gene screening.
Smart Images
Figure CN122156986A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of crop phenotypic analysis, computer vision, and artificial intelligence, specifically to a method for analyzing rice drought resistance phenotypic patterns based on multi-task deep learning. Background Technology
[0002] Drought stress is one of the major abiotic stresses affecting rice growth, development, and yield, severely restricting the stability and sustainability of agricultural production. Accurately analyzing the physiological responses and phenotypic changes of rice under drought stress is a core prerequisite for studying drought resistance mechanisms and cultivating superior drought-resistant varieties. In recent years, the rapid development of high-throughput crop phenotyping platforms has enabled the provision of multi-view, multi-temporal, and multi-modal image data for monitoring crop growth status, overcoming the limitations of low efficiency and high subjectivity in traditional phenotyping observations, and providing data support for large-scale population drought resistance phenotyping system research. Simultaneously, breakthroughs in deep learning technology in computer vision have promoted the gradual application of methods based on convolutional neural networks, visual Transformers, and other models in plant phenotyping recognition, disease detection, and other fields, providing new technical pathways for the automated extraction of complex phenotypic features.
[0003] Traditional methods for analyzing drought-resistance phenotypes in rice often rely on manually defined image segmentation rules or manual feature extraction. These methods can only capture simple phenotypic indicators such as leaf area and plant height, failing to adequately depict the complex color changes and morphological distortions experienced during drought response. Furthermore, they are highly sensitive to light fluctuations and environmental noise, resulting in insufficient accuracy and stability in phenotypic representation. Existing deep learning methods also have significant limitations, mostly focusing on single classification or regression tasks, making it difficult to simultaneously meet the dual requirements of drought severity identification and agronomic trait quantification. In addition, deep features have weak correlations with actual plant physiological processes, leading to poor model interpretability and a lack of ability to continuously model dynamic phenotypic changes under drought stress. This results in insufficient effectiveness of phenotypic-genotype association analysis, failing to provide reliable support for drought-resistance gene discovery. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for analyzing rice drought resistance phenotypes based on multi-task deep learning. This method integrates multi-view, multi-temporal image acquisition, deep feature decoupling, and physiological space alignment techniques to construct a multi-task deep learning framework for accurate analysis of rice drought resistance phenotypes. It utilizes a visual Transformer model to extract features and independently constructs color and morphological physiological branches to avoid feature confounding; it constructs a drought resistance physiological space through measured agronomic traits, linking deep features with actual physiological traits; it simultaneously performs drought level classification and agronomic trait regression to improve the comprehensiveness of the analysis; and it combines temporal constraints and genome-wide association analysis to strengthen phenotype-genotype associations and improve gene mining efficiency.
[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for analyzing rice drought resistance phenotypes based on multi-task deep learning, the specific steps of which are as follows:
[0006] S1, Image Acquisition: Use a high-throughput crop phenotyping platform to acquire multi-view, multi-temporal RGB images of rice before drought treatment, during drought treatment, and during the recovery period;
[0007] S2, Feature Extraction: The visual Transformer model DINOv2-ViT-B is used to extract global feature identifiers and local image patch feature identifiers from the preprocessed RGB image;
[0008] S3, Branch Construction and Feature Decoupling: Based on the features in S2, color physiological branches and morphological physiological branches are constructed, and mutually independent color physiological features and morphological physiological features are obtained through feature decoupling constraints.
[0009] S4, Physiological Space Construction and Feature Projection: Based on five agronomic traits measured by a high-throughput crop phenotyping platform, a drought-resistant physiological space is constructed. The fused two types of physiological features are projected into this space to obtain drought-resistant depth phenotypic features.
[0010] S5, Multi-task network construction and optimization: Construct a multi-task deep learning network to simultaneously perform drought level classification and agronomic trait regression tasks, and optimize features through a joint loss function;
[0011] S6, Temporal Constraint Feature Optimization: Temporal constraints are applied to drought resistance depth phenotypic features at different time phases to obtain drought resistance depth phenotypic features that can be used for phenotypic analysis and gene mining;
[0012] S7, Genetic Locus Identification: Using the features of S6 as input, genome-wide association analysis is used to identify genetic loci related to drought resistance in rice.
[0013] Furthermore, the high-throughput crop phenotyping platform in S1 includes a rotating stage, a multi-angle RGB imaging module, and an automatic transmission system. The greenhouse cultivation system maintains consistent light, temperature, and humidity. Rice samples are planted individually in pots of uniform size and arranged in a randomized or completely randomized manner. Before drought treatment, the plants are irrigated normally until the set growth stage. By stopping watering or reducing water supply, they enter different drought stages. The imaging perspective includes side and top views, covering key nodes before drought treatment, during treatment, and multiple time points during the recovery period.
[0014] Furthermore, the preprocessing in S2 includes image size unification and color normalization. Based on the local image block feature identification, edge map, skeleton map and texture map are introduced. The edge map is obtained through edge detection, the skeleton map is obtained through skeletonization algorithm, and the texture map is generated by LBP operator or Gabor operator. After the three types of images are encoded by convolutional layer, they are sent together with the local image block feature identification into the morphological branch network to supplement morphological information.
[0015] Furthermore, the feature decoupling constraint in S3 is a feature orthogonal constraint or a mutual information minimization constraint; the color physiology branch integrates RGB color histogram, HSV color space features, ExG index and green ratio GPAR to characterize the yellowing and withering changes caused by drought; the morphology physiology branch integrates convolutional edge feature map, skeleton map and local image block feature identifier to characterize the leaf curling and wrinkling changes caused by drought.
[0016] Furthermore, the five agronomic traits in S4 are green proportion, projected area, green leaf area, leaf water content and plant height. After standardization, the five agronomic traits are used to construct a drought-resistant physiological space using principal component analysis. This space is a low-dimensional feature space that reflects the main physiological changes in drought. The two types of physiological traits after fusion are projected onto this space through linear or nonlinear mapping.
[0017] Furthermore, the multi-task deep learning network in S5 includes a shared DINOv2-ViT-B feature encoder, a drought level classification head, and an agronomic trait regression head; the drought level classification head includes a fully connected layer and a softmax layer, and the output results are four categories: normal, mild drought, moderate drought, and severe drought; the traits output by the agronomic trait regression head include plant projected area, green ratio, green area, plant height, leaf water content, and SPAD value.
[0018] Furthermore, the joint loss function in S5 includes cross-entropy loss for drought classification, mean squared error loss for agronomic trait regression, feature decoupling loss, physiological spatial alignment loss, temporal consistency loss, and temporal smoothing loss, which optimize the deep feature representation of the shared feature encoder through backpropagation.
[0019] Furthermore, the temporal constraints in S6 include temporal consistency constraints and temporal smoothing constraints. Temporal consistency constraints control the magnitude of differences in features between adjacent time phases, while temporal smoothing constraints achieve smoothing of the feature time series by penalizing second-order differences, thus avoiding non-physiological mutations.
[0020] Furthermore, the genome-wide association analysis in S7 includes quality control processing of the sample genotype data, which involves filtering low-frequency SNPs and removing high-deletion-rate sites. The statistical association between SNPs and drought resistance depth phenotypic characteristics is calculated using an MLM model or a FarmCPU model. Candidate gene intervals are determined through linkage disequilibrium analysis, and drought resistance-related candidate genes are screened in conjunction with gene annotation information.
[0021] Compared with existing technologies, this method for analyzing rice drought resistance phenotypes based on multi-task deep learning has the following advantages:
[0022] I. This invention integrates multi-view, multi-temporal image acquisition, deep feature decoupling, and physiological space alignment techniques to construct a multi-task deep learning framework, effectively overcoming the limitations of traditional phenotypic analysis methods. By independently constructing color and morphological physiological branches and using feature decoupling constraints, it achieves accurate separation and representation of drought-induced color and morphological changes, avoiding phenotypic misjudgments caused by feature confounding. Through the construction of a drought-resistant physiological space using measured agronomic traits, it establishes a clear correlation between deep features and real physiological traits, solving the industry pain point of uninterpretable features in deep learning models. The multi-task network simultaneously completes drought level classification and agronomic trait regression, significantly improving the comprehensiveness and accuracy of phenotypic analysis, eliminating dependence on single features or static time points, and providing stable and reliable technical support for the quantitative analysis of rice drought-resistant phenotypes.
[0023] II. This invention optimizes dynamic phenotypic features through temporal consistency and smoothing constraints, and combines this with a closed-loop technology of genome-wide association analysis to significantly improve the effectiveness of drought resistance genetic locus mining. Targeting the dynamic changes in rice drought response, this invention constrains deep phenotypic features at different time phases to ensure that their changes conform to plant physiological patterns, avoid interference from non-physiological mutations, and improve feature stability. The optimized drought resistance deep phenotypic features are then used for gene association analysis to strengthen the association between phenotype and genotype, improve the accuracy of significant SNP detection and the reliability of candidate gene screening, and effectively solve the problems of insufficient dynamic modeling and low gene mining efficiency in existing methods.
[0024] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0026] Figure 1 Here is the overall method flowchart;
[0027] Figure 2 This is a flowchart of the feature extraction and processing.
[0028] Figure 3 Flowchart for multi-task learning and feature optimization. Detailed Implementation
[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0030] Example 1:
[0031] Experimental materials and equipment:
[0032] The experimental material was the indica rice variety Guichao No. 2. Three hundred and sixty healthy, uniform seeds were selected, soaked and germinated, and then planted one seedling in each of three hundred and sixty uniformly sized pots (20 cm in diameter and 25 cm in height). The experimental equipment employed a high-throughput crop phenotyping platform, which integrated a rotating stage, four multi-angle RGB imaging modules, and an automatic transmission system. The accompanying greenhouse cultivation system precisely controlled key environmental parameters such as light intensity, temperature, and humidity, ensuring stable and controllable environmental conditions throughout the experiment.
[0033] Experimental steps:
[0034] S1, Image Acquisition: After planting, the potted plants were neatly arranged in randomized blocks within the greenhouse cultivation system. The greenhouse was maintained at a light intensity of 30,000 lux, a temperature of 28 degrees Celsius, and a humidity of 60%. All plants were irrigated normally until the tillering stage. Watering was then stopped, and a drought treatment was initiated. Imaging was conducted on the day before the drought treatment, the third day of the drought treatment (mild drought stage), the fifth day of the drought treatment (moderate drought stage), the seventh day of the drought treatment (severe drought stage), and the second and fifth days of the recovery period. After the plants entered the imaging area, the rotating platform automatically rotated 360 degrees, simultaneously acquiring RGB images from four side views and one top view using four RGB imaging modules. Twenty-one images were acquired per plant at each time point (twenty side view images and one top view image), comprehensively covering the overall structure and local details of the plant. For example... Figure 1 As shown, the overall methodology begins with the acquisition of multi-view, multi-temporal images.
[0035] S2, Feature Extraction: Based on the RGB images acquired in S1, preprocessing is performed. First, the image size is uniformly adjusted to 512 pixels by 512 pixels, and then color normalization is completed using the Z-score normalization method. The preprocessed image is input into the visual Transformer model DINOv2-ViT-B. The model automatically segments the image into 16-pixel by 16-pixel patches and generates patch feature labels and global feature labels through a self-attention structure. Simultaneously, edge maps are generated using the Canny edge detection algorithm, skeleton maps are generated using the Zhang-Suen skeletonization algorithm, and texture maps are generated using the LBP operator. The three types of images are encoded through three convolutional layers and then fed into the morphological branch network along with the patch feature labels to support subsequent morphological and physiological feature extraction. Figure 2 As shown, the feature extraction and processing flow includes image preprocessing, DINOv2-ViT-B feature extraction, and supplementary encoding of edge, skeleton, and texture information.
[0036] S3, Branch Construction and Feature Decoupling: Based on the features extracted in S2, color physiological and morphological physiological branches are constructed. The color physiological branch integrates RGB color histogram, HSV color space features, ExG index, and green ratio (GPAR), and encodes them through a three-layer multilayer perceptron with 256, 128, and 64 neurons per layer, respectively. After encoding, it is deeply fused with the color information in the global feature identifier to obtain color physiological features. The morphological physiological branch integrates convolutional edge feature maps, skeleton maps, and patch feature identifiers, and extracts morphological physiological features through a hybrid structure of two convolutional layers and one Transformer layer. The convolutional kernel size is 3x3, with 64 and 128 kernels, respectively. Feature orthogonality constraints are introduced to make the color physiological features and morphological physiological features tend to be independent of each other during model training, achieving effective decoupling of the two types of features.
[0037] S4, Physiological Space Construction and Feature Projection: Five agronomic traits—green proportion, projected area, green leaf area, leaf water content, and plant height—were measured at each time point using a high-throughput crop phenotyping platform. The min-max standardization method was used to standardize the measured data, eliminating dimensional differences between traits. The standardized agronomic trait data were then input into a principal component analysis model. The first three principal components were retained to construct a drought-resistance physiological space. This space has a cumulative variance contribution rate of 85%, effectively reflecting the main physiological changes in rice under drought stress. The color and morphological physiological characteristics fused in S3 were projected onto this drought-resistance physiological space using a nonlinear mapping method, ultimately obtaining drought-resistance depth phenotypic features with clear physiological orientation.
[0038] S5, Multi-task Network Construction and Optimization: A multi-task deep learning network is constructed, comprising a shared DINOv2-ViT-B feature encoder, a drought level classification head, and an agronomic trait regression head. The drought level classification head consists of two fully connected layers and a softmax layer, with 128 and 14 neurons in the fully connected layers, respectively, outputting four classification results: normal, mild drought, moderate drought, and severe drought. The agronomic trait regression head also has two fully connected layers, with 128 and 16 neurons, respectively, outputting six agronomic trait data: plant projected area, green ratio, green area, plant height, leaf water content, and SPAD value. The joint loss function consists of cross-entropy loss for drought classification, mean squared error loss for agronomic trait regression, feature decoupling loss, physiological spatial alignment loss, temporal consistency loss, and temporal smoothing loss, with each loss weight set to a ratio of 1:1:0.8:0.8:0.5:0.5. The feature encoder is optimized using a backpropagation algorithm, with 100 iterations and a learning rate of 0.001. Figure 3 As shown, the multi-task learning and feature optimization process includes a shared encoder, dual task heads, joint loss function, and time-series constraint optimization.
[0039] S6, Temporal Constraint Feature Optimization: For drought resistance depth phenotypic features of the same plant at different time phases, temporal consistency constraints and temporal smoothing constraints are applied. The temporal consistency constraint strictly controls the difference in features between adjacent time phases to not exceed 0.1, while the temporal smoothing constraint is achieved through penalized second-order differences, with the penalty coefficient set to 0.01. Through the synergistic effect of these two types of constraints, the continuity and rationality of feature changes over time are ensured, ultimately yielding drought resistance depth phenotypic features that can be directly used for phenotypic analysis and gene mining.
[0040] S7. Genetic Locus Identification: Genotypic data from 360 plants were extracted. Strict quality control was performed on the genotypic data, filtering out SNPs with minor allele frequencies less than 0.05 and removing loci with deletion rates greater than 0.1 to ensure data quality. An MLM model was used to calculate the statistical association between SNPs and drought resistance phenotypic characteristics, setting a significance threshold of P < 10 to the power of -6. Candidate gene intervals were determined through linkage disequilibrium analysis, and candidate genes related to drought resistance were systematically screened using a rice gene annotation database.
[0041] This embodiment uses the indica rice variety Guichao 2 as the experimental material. Multi-view, multi-temporal RGB images were acquired using a high-throughput crop phenotyping platform, covering various stages of drought treatment, before treatment, and during the recovery period. After preprocessing with size adjustment and color normalization, features were extracted using the DINOv2-ViT-B model, and edge maps, skeleton maps, and texture maps were fused to supplement morphological information. Feature decoupling was achieved by constructing color and morphological physiological branches and applying orthogonal constraints. A drought-resistance physiological space was constructed based on five measured agronomic traits, and feature projection was completed. Further optimization using a multi-task network and temporal constraints were applied, and finally, a MLM model was used to screen drought-resistance-related genetic loci. The entire process strictly followed the technical solution of this invention, with reasonable parameter settings for each step, achieving a complete closed loop from image acquisition to gene mining, and verifying the feasibility and standardization of the technical solution on indica rice varieties.
[0042] Example 2:
[0043] Experimental materials and equipment:
[0044] The experimental material was the Nipponbare japonica rice variety. Two hundred and forty healthy and uniform seeds were selected, soaked and germinated, and then evenly planted in pots of uniform size with a diameter of 18 cm and a height of 22 cm. One plant was planted in each pot, for a total of 240 pots. The experimental equipment was consistent with that in Example 1, using a high-throughput crop phenotyping platform and a matching greenhouse cultivation system. Its composition and function were exactly the same, enabling precise control of environmental parameters and automated imaging of plants, ensuring the standardization and comparability of the experiment.
[0045] Experimental steps:
[0046] S1, Image Acquisition: Pots were placed randomly within the greenhouse cultivation system, maintaining a light intensity of 28,000 lux, a temperature of 26 degrees Celsius, and a humidity of 55%. All plants were irrigated normally until the jointing stage. Watering was then stopped, and a drought treatment was implemented. Imaging was conducted on the day before the drought treatment, on the fourth day of the drought treatment (mild drought stage), the sixth day of the drought treatment (moderate drought stage), the eighth day of the drought treatment (severe drought stage), and on the third and sixth days of the recovery period. After the plants entered the imaging area, the rotating platform automatically rotated 360 degrees, and three RGB imaging modules acquired RGB images from three side views and one top view. Twenty-one images were acquired per plant at each time point (twenty side view images and one top view image), comprehensively capturing the phenotypic information of the plants at different drought stages.
[0047] S2, Feature Extraction: The RGB images acquired in S1 are preprocessed, with the image size uniformly adjusted to 384 pixels by 384 pixels. Color normalization is performed using the min-max normalization method. The preprocessed images are then input into the visual Transformer model DINOv2-ViT-B. The model segments the image into 16-pixel by 16-pixel patches, automatically generating patch feature labels and global feature labels. Simultaneously, an edge map is generated using the Sobel edge detection algorithm, a skeleton map is generated using a skeleton extraction algorithm, and a texture map is generated using the Gabor operator. These three types of images are encoded through two convolutional layers and then fed into the morphological branch network along with the patch feature labels, laying the foundation for accurate extraction of morphological and physiological features.
[0048] S3, Branch Construction and Feature Decoupling: Based on the global feature identifiers and patch feature identifiers extracted in S2, color physiological branches and morphological physiological branches are constructed. The color physiological branch integrates RGB color histograms, HSV color space features, ExG index, and green ratio GPAR, and encodes them through two multilayer perceptrons with 128 and 64 neurons per layer, respectively. After encoding, it is fused with the color information in the global feature identifiers to obtain color physiological features. The morphological physiological branch integrates convolutional edge feature maps, skeleton maps, and patch feature identifiers, and extracts morphological physiological features through a hybrid structure of three convolutional layers and one Transformer layer. The convolutional kernel size is 3x3, with 32, 64, and 128 kernels, respectively. A mutual information minimization constraint is introduced to make the color physiological features and morphological physiological features tend to be independent during training, achieving effective separation of the two types of features.
[0049] S4, Physiological Space Construction and Feature Projection: Five agronomic traits—green proportion, projected area, green leaf area, leaf water content, and plant height—were measured at various time points using a high-throughput crop phenotyping platform. Z-score standardization was used to process the measured data to eliminate the influence of dimensions. The standardized agronomic trait data were input into a principal component analysis model, retaining the first three principal components to construct a drought-resistance physiological space. This space has a cumulative variance contribution rate of 82%, accurately reflecting the core physiological changes in rice under drought stress. The two physiological traits fused in S3 were projected onto this drought-resistance physiological space through linear mapping to obtain physiologically significant drought-resistance depth phenotypic features.
[0050] S5, Multi-task Network Construction and Optimization: The constructed multi-task deep learning network includes a shared DINOv2-ViT-B feature encoder, a drought level classification head, and an agronomic trait regression head. The drought level classification head consists of two fully connected layers and a softmax layer, with 64 and 4 neurons respectively; the agronomic trait regression head consists of two fully connected layers, with 64 and 6 neurons respectively. The weight ratios of the loss functions in the joint loss function are set to 1:1:0.7:0.7:0.6:0.6, the number of iterations is set to 80, and the learning rate is 0.0008. The feature encoder is continuously optimized through backpropagation to ensure that the network has both accurate drought level classification and agronomic trait regression capabilities.
[0051] S6, Temporal Constraint Feature Optimization: Temporal consistency constraints and temporal smoothing constraints are applied to the drought resistance depth phenotypic features of the same plant at different time phases. The temporal consistency constraint controls the difference in features between adjacent time phases to not exceed 0.12, and the penalty coefficient of the temporal smoothing constraint is set to 0.02. Through the synergistic effect of the two types of constraints, non-physiological mutations in features are avoided, ensuring that the trend of feature change is continuous and reasonable, and finally, stable and reliable drought resistance depth phenotypic features are obtained.
[0052] S7. Genetic Locus Identification: Genotypic data from 240 plants underwent quality control processing, filtering out SNPs with minor allele frequencies less than 0.05 and removing loci with deletion rates greater than 0.1. The FarmCPU model was used to calculate the statistical association between SNPs and drought resistance phenotypic characteristics, with a significance threshold set at P < 10 to the power of -6. Candidate gene intervals were determined through linkage disequilibrium analysis. Combined with rice gene annotation information, a comprehensive screening of drought-related candidate genes was conducted, providing gene-level support for research on rice drought resistance mechanisms.
[0053] This embodiment uses the japonica rice variety Nipponbare and employs the same high-throughput crop phenotyping platform and equipment as in Embodiment 1. Plants were placed in a completely randomized manner and irrigated until the jointing stage to initiate drought treatment. Multi-view RGB images were acquired at key time points. By adjusting key conditions such as image size, standardization methods, and network parameters, core steps including feature extraction, branch construction and feature decoupling, physiological space construction and projection, multi-task network optimization, and time-series constraints were completed. Finally, the FarmCPU model was used to conduct genetic locus association analysis. This embodiment, through differentiated material selection, placement methods, and parameter settings, further verifies the flexibility and adaptability of the technical solution of this invention. Its operational process is standardized and orderly, covering the experimental needs of different japonica rice varieties, and provides a valid reference for the large-scale application of the technical solution.
[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for analyzing rice drought resistance phenotypes based on multi-task deep learning, characterized in that, The specific steps of this method are as follows: S1, Image Acquisition: Use a high-throughput crop phenotyping platform to acquire multi-view, multi-temporal RGB images of rice before drought treatment, during drought treatment, and during the recovery period; S2, Feature Extraction: The visual Transformer model DINOv2-ViT-B is used to extract global feature identifiers and local image patch feature identifiers from the preprocessed RGB image; S3, Branch Construction and Feature Decoupling: Based on the features in S2, color physiological branches and morphological physiological branches are constructed, and mutually independent color physiological features and morphological physiological features are obtained through feature decoupling constraints. S4, Physiological Space Construction and Feature Projection: Based on five agronomic traits measured by a high-throughput crop phenotyping platform, a drought-resistant physiological space is constructed. The fused two types of physiological features are projected into this space to obtain drought-resistant depth phenotypic features. S5, Multi-task network construction and optimization: Construct a multi-task deep learning network to simultaneously perform drought level classification and agronomic trait regression tasks, and optimize features through a joint loss function; S6, Temporal Constraint Feature Optimization: Temporal constraints are applied to drought resistance depth phenotypic features at different time phases to obtain drought resistance depth phenotypic features that can be used for phenotypic analysis and gene mining; S7, Genetic Locus Identification: Using the features of S6 as input, genome-wide association analysis is used to identify genetic loci related to drought resistance in rice.
2. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The high-throughput crop phenotyping platform in S1 includes a rotating stage, a multi-angle RGB imaging module, and an automatic transmission system. The greenhouse cultivation system maintains consistent light, temperature, and humidity. Rice samples are planted individually in uniform-sized pots and arranged in a randomized or completely randomized manner. Before drought treatment, the plants are irrigated normally until the set growth stage. By stopping watering or reducing water supply, they enter different drought stages. The imaging perspective includes side and top views, covering key nodes before drought treatment, during treatment, and multiple time points during the recovery period.
3. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The preprocessing in S2 includes image size unification and color normalization. Based on the local image block feature identification, edge map, skeleton map and texture map are introduced. The edge map is obtained by edge detection, the skeleton map is obtained by skeletonization algorithm, and the texture map is generated by LBP operator or Gabor operator. After the three types of images are encoded by convolutional layer, they are sent to the morphological branch network along with the local image block feature identification to supplement morphological information.
4. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The feature decoupling constraint in S3 is a feature orthogonal constraint or a mutual information minimization constraint; the color physiology branch integrates RGB color histogram, HSV color space features, ExG index and green ratio GPAR to characterize the yellowing and withering changes caused by drought; the morphology physiology branch integrates convolutional edge feature map, skeleton map and local image block feature identifier to characterize the leaf curling and wrinkling changes caused by drought.
5. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The five agronomic traits in S4 are green proportion, projected area, green leaf area, leaf water content and plant height. After standardization, the five agronomic traits are used to construct a drought-resistant physiological space using principal component analysis. This space is a low-dimensional feature space that reflects the main physiological changes in drought. The two types of physiological traits after fusion are projected onto this space through linear or nonlinear mapping.
6. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The multi-task deep learning network in S5 includes a shared DINOv2-ViT-B feature encoder, a drought level classification head, and an agronomic trait regression head. The drought level classification head includes a fully connected layer and a softmax layer, and the output results are four categories: normal, mild drought, moderate drought, and severe drought. The traits output by the agronomic trait regression head include plant projected area, green ratio, green area, plant height, leaf water content, and SPAD value.
7. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The joint loss function in S5 includes cross-entropy loss for drought classification, mean squared error loss for agronomic trait regression, feature decoupling loss, physiological spatial alignment loss, temporal consistency loss, and temporal smoothing loss. It optimizes the deep feature representation of the shared feature encoder through backpropagation.
8. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The temporal constraints in S6 include temporal consistency constraints and temporal smoothing constraints. Temporal consistency constraints control the magnitude of differences in features between adjacent time phases, while temporal smoothing constraints achieve smoothing of the feature time series by penalizing second-order differences, thus avoiding non-physiological mutations.
9. The method for analyzing rice drought resistance phenotypes based on multi-task deep learning according to claim 1, characterized in that, The genome-wide association analysis in S7 includes quality control processing of sample genotype data, such as filtering low-frequency SNPs and removing high-deletion-rate sites, using MLM or FarmCPU models to calculate the statistical association between SNPs and drought resistance depth phenotypic characteristics, determining candidate gene intervals through linkage disequilibrium analysis, and screening drought resistance-related candidate genes in combination with gene annotation information.