A method and system for pre-harvest identification of wheat waterlogging tolerance based on unmanned aerial vehicle dynamic phenotype

Abstract

The application discloses a wheat waterlogging tolerance pre-harvest identification method and system based on unmanned aerial vehicle dynamic phenotype, and relates to agricultural remote sensing and crop high-throughput phenotype. Multi-time phase unmanned aerial vehicle canopy images of the same batch of wheat genotypes under normal water control and waterlogging treatment, covering stress accumulation and recovery stages after drainage are obtained; plot-level structure and spectral variables are extracted, paired responses of the same genotype on the same date are constructed, and baseline centralization of the starting time point is carried out; stress accumulation, damage correction condition recovery and residual effect scores are calculated from the comprehensive canopy response trajectory, a dynamic stress recovery function index is integrated, and material ranking, grading or pre-harvest candidate list is output. The method can reduce initial difference interference and provide an interpretable process evaluation.

Description

Technical Field

[0001] This application relates to the fields of agricultural remote sensing, crop high-throughput phenotyping and digital breeding technologies, specifically to a method and system for pre-harvest identification of wheat waterlogging tolerance based on UAV dynamic phenotyping. Background Technology

[0002] Waterlogging restricts gas exchange in the wheat root zone and affects nutrient absorption, leading to leaf yellowing, premature canopy senescence, and yield loss. During breeding, it is necessary to identify from a large population of materials those that have experienced less waterlogging damage, recover well after drainage, and have minimal late-stage residual effects.

[0003] Current methods for evaluating wheat waterlogging tolerance typically involve visual inspection, destructive sampling, or yield measurement at harvest. Visual inspection is susceptible to the influence of the observation time and the experience of the evaluators; destructive sampling makes it difficult to continuously observe the same material; and yield measurement at harvest only yields results after the growth period has ended. These methods are insufficient for continuous, non-destructive priority screening of large-scale breeding populations before harvest.

[0004] Unmanned aerial vehicle (UAV) remote sensing can repeatedly acquire canopy spectral and structural information of breeding plots. Existing methods often use vegetation indices from a single date or absolute canopy status under waterlogging treatment for evaluation, which easily contaminates genotypic baseline growth, inherent leaf color, phenology, and differences in observation environment on the same day. Even when using multi-date data, the variables for each date are often directly expanded and input into the prediction model, leading to cross-date information redundancy, variable collinearity, and difficulty in explaining the reasons for material sorting.

[0005] Furthermore, the apparent recovery rate after drainage is related to the degree of damage during drainage. More severely damaged materials have greater potential for recovery, and directly comparing recovery slopes may misinterpret differences in damage initiation as differences in recovery capacity. Furthermore, the canopy state on a single date may correspond to different process pathways, making it impossible to distinguish between lightly damaged but slowly recovering materials and heavily damaged but rapidly recovering materials. Summary of the Invention

[0006] This invention addresses the challenges of existing technologies in eliminating initial material differences, fairly comparing recovery capabilities at different damage initiation points, and converting multi-temporal information into interpretable pre-harvest ordination. It proposes a technical solution for wheat waterlogging scenarios, separating relative stress responses from multi-temporal UAV canopy data of the same genotype (treatment / control), converting continuous response trajectories into stress accumulation, damage-corrected recovery, and late residual states, and generating interpretable material ordination before harvest.

[0007] Technical solution:

[0008] This invention first discloses a method for pre-harvest identification of wheat waterlogging tolerance based on UAV dynamic phenotypes, comprising:

[0009] Step S101: Obtain multi-temporal UAV canopy images of the same batch of wheat genotypes under normal moisture control and waterlogging treatment, wherein the multi-temporal UAV canopy images cover the waterlogging stress accumulation stage and the recovery stage after drainage.

[0010] Step S102: Perform image correction, cross-date spatial registration, region of interest determination and canopy segmentation on the multi-temporal UAV canopy images, and extract the cell-level canopy structure variables and canopy spectral variables of each genotype at each observation time point;

[0011] Step S103: Construct paired responses for the corresponding variables of the same genotype and the same observation time point under the waterlogging treatment and the normal water control, and use the paired responses at the treatment start time point to base-center the subsequent paired responses;

[0012] Step S104: Perform orientation unification and aggregation on at least one canopy structure variable and at least one canopy spectral variable after baseline centering to obtain the comprehensive canopy response trajectory of each genotype;

[0013] Step S105: Generate a stress accumulation score based on the maximum damage and cumulative loss of the integrated canopy response trajectory during the waterlogging stress accumulation stage; generate a conditional recovery score based on the early recovery slope of the integrated canopy response trajectory after drainage and by regression correction of the early recovery slope with the damage gap during drainage; generate a residual effect score based on the canopy response of the integrated canopy response trajectory in the later recovery stage, which together constitute the three process scores.

[0014] Step S106: The stress cumulative score, conditional recovery score and residual effect score are unified and integrated to obtain the dynamic stress recovery function index of each genotype.

[0015] Step S107: Output at least one of the following genotypes from the list of waterlogging tolerance, waterlogging tolerance level, and pre-harvest candidate materials, based on the dynamic stress recovery function index.

[0016] Preferably, the multi-temporal UAV canopy images are multispectral images, including at least red, green, red-edge, and near-infrared bands; the normal moisture control and the waterlogging treatment are set up in one-to-one corresponding plots for the same genotype, and UAV observations are conducted at the start of treatment, the waterlogging stress accumulation stage, the drainage time point, and the recovery stage after drainage.

[0017] Preferably, processing the multi-temporal UAV canopy images includes:

[0018] Radiometric correction is performed based on standard reflector data and / or incident light sensor data to generate multispectral orthophotos;

[0019] The multispectral orthophotos at each observation time point are transformed to a unified projection coordinate system, and cross-date spatial registration is performed with the image at the processing start time point as a reference.

[0020] The region of interest for each cell is determined on the image at the starting time of the processing, and the region of interest is applied to the registered images at each observation time.

[0021] Canopy pixels are distinguished from background pixels based on the normalized vegetation index threshold, and morphological opening and closing operations are performed on the segmentation results to generate a canopy mask.

[0022] Preferably, the canopy structure variables include canopy coverage, and the canopy spectral variables include soil-regulated vegetation index, green normalized vegetation index, and normalized difference red edge index; the comprehensive canopy response trajectory is obtained by baseline-centered paired response weighted aggregation of the canopy coverage, soil-regulated vegetation index, green normalized vegetation index, and normalized difference red edge index.

[0023] Preferably, the paired response of the g-th genotype to the m-th variable at the t-th observation time point. and baseline-centered paired response Calculate according to the following formulas respectively:

[0024]

[0025]

[0026] in, and These are the variable values ​​under the waterlogging treatment and the normal moisture control, respectively. To handle paired responses at the start time.

[0027] Preferably, the generation of the stress cumulative score includes: determining the maximum damage based on the maximum negative deviation of the integrated canopy response trajectory relative to the baseline within a stress time window from the start of treatment to the drainage time window; integrating the negative deviation of the integrated canopy response trajectory within the stress time window to obtain the cumulative loss; and aggregating the maximum damage and the cumulative loss after unifying and standardizing them in a favorable direction to obtain the stress cumulative score; wherein, in one embodiment, the stress time window is from day 0 to day 11 after treatment.

[0028] Preferably, the generation of the conditional recovery score includes: fitting the slope of the comprehensive canopy response trajectory within the early recovery time window starting from the self-drainage point to obtain the early recovery slope of the g-th genotype. ; according to Damage gap during drainage calculation ,in For the g-th genotype at the drainage time point The comprehensive canopy response was obtained; Huber loss was used for fitting. and the residual The favorable direction is unified and standardized to obtain the conditional recovery score, where α0 and α1 represent robust regression parameters, respectively; wherein, in one embodiment, the early recovery time window is from day 11 to day 14 after treatment.

[0029] Preferably, the generation of the residual effect score includes: calculating the median of the integrated canopy response trajectory within the late recovery time window. ,according to Residual damage was obtained The residual damage is then standardized and oriented in a favorable direction to obtain the residual effect score; wherein, in one embodiment, the recovery period window is from day 18 to day 20 after treatment.

[0030] Preferably, the Dynamic Stress Recovery Function Index (D-SRFI) g Calculate, where S g R g and F g These represent the cumulative stress score, conditional recovery score, and residual effect score for the g-th genotype, respectively. S w R and w F The weights are non-negative and sum to 1; in the equal-weight implementation, wS=wR=wF=1 / 3.

[0031] Preferably, step S107 outputs at least one of the following: maximum damage, cumulative loss, early recovery slope, damage gap during drainage, recovery residual after damage correction, canopy response in the later recovery stage, and process score, to provide a process explanation and data audit basis for material sorting.

[0032] This invention also discloses a pre-harvest identification system for wheat waterlogging tolerance based on UAV dynamic phenotypes, comprising:

[0033] The data acquisition module is used to acquire multi-temporal UAV canopy images of the same batch of wheat genotypes under normal water control and waterlogging treatment, covering the waterlogging stress accumulation stage and the recovery stage after drainage.

[0034] The image processing module is used to perform image correction, cross-date spatial registration, cell region determination and canopy segmentation on the multi-temporal UAV canopy images, and extract cell-level canopy structure variables and canopy spectral variables;

[0035] The paired response construction module is used to construct paired responses between waterlogging treatment and normal water control with the same genotype and the same observation time point, and to center the baseline at the treatment start time point to form a comprehensive canopy response trajectory.

[0036] The process feature calculation module is used to generate stress cumulative score, conditional recovery score after drainage damage gap correction, and residual effect score based on the comprehensive canopy response trajectory.

[0037] The comprehensive evaluation module is used to unify and integrate the scores of the three processes to obtain the dynamic stress recovery function index.

[0038] The results output module is used to output at least one of the following for each genotype: waterlogging tolerance ranking, waterlogging tolerance level, and pre-harvest candidate material list, based on the dynamic stress recovery function index.

[0039] The present invention also discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the aforementioned method for pre-harvest identification of wheat waterlogging tolerance based on UAV dynamic phenotypes.

[0040] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for pre-harvest identification of wheat waterlogging tolerance based on UAV dynamic phenotypes.

[0041] Beneficial effects of the present invention

[0042] Compared with the scheme using absolute canopy values ​​under waterlogging treatment, this application can reduce the impact of material basal growth, inherent leaf color, phenology and initial plot differences on waterlogging response evaluation by matching treatments / controls of the same genotype and date and centering the baseline at the start of treatment.

[0043] This application converts the comprehensive canopy trajectory into three process scores: stress accumulation, conditional recovery, and residual effects. It also corrects the early recovery slope with the damage gap during drainage, which can distinguish different damage depths, recovery rates, and late recovery states, providing a process explanation for material ranking.

[0044] This application uses limited canopy structure and spectral variables, along with process-level aggregation, to compress multi-temporal data and outputs underlying trajectory descriptors and process scores, facilitating pre-harvest prioritization and result auditing of large-scale breeding materials. In a validation example containing 902 genotypes with complete UAV phenotypes and measured harvest traits, the three process scores significantly compressed the input dimensions while maintaining predictive power for yield retention and thousand-grain weight retention. The D-SRFI was positively correlated with both measured retention rates, indicating that the process decomposition and comprehensive ranking can provide endpoint trait support for pre-harvest candidate reduction. Attached Figure Description

[0045] Figure 1 The overall flowchart of the method for identifying wheat waterlogging resistance before harvest provided in the embodiments of this application is shown.

[0046] Figure 2 This is a schematic diagram of the processing / control pairing response and the baseline centering of the processing start point provided in the embodiments of this application.

[0047] Figure 3 A schematic diagram illustrating the construction of the three-process features and dynamic stress recovery function index provided in the embodiments of this application.

[0048] Figure 4 A block diagram of a wheat waterlogging resistance pre-harvest identification system provided in an embodiment of this application.

[0049] Figure 5 The UAV dynamic phenotypic features provided in the embodiments of this application represent the prediction results of wheat yield retention rate and thousand-grain weight retention rate ((a, b) multi-date original variables; (c, d) dynamic functional traits; (e, f) D-SRFI). Detailed Implementation

[0050] The embodiments of this application are described below with reference to the accompanying drawings. These embodiments are used to explain the technical solutions of this application and do not constitute a limitation on the scope of protection. Unless otherwise specified, normal moisture control is denoted as CK, waterlogging treatment as WL, genotype as g, observation time point as t, and UAV phenotypic variables as m.

[0051] Example 1: Multi-temporal UAV data acquisition and cell phenotypic extraction

[0052] A field trial involving multiple wheat genotypes was established. Each genotype had corresponding plots in both the control (CK) and wheat low-level (WL) zones, with consistent sowing, fertilization, and field management practices except for water treatment. Genotypes were arranged separately in the two treatment zones, and a one-to-one correspondence between CK and WL was established using consistent material identification.

[0053] In one specific implementation, floodwater treatment begins on day 0 after treatment, and drainage is carried out after observation on day 11. UAVs conduct observations on days 0, 4, 8, 9, 11, 12, 13, 14, 16, 18, 19, and 20, respectively, thus covering the treatment initiation, floodwater accumulation, and post-drainage recovery phases. In practical applications, other relative observation points can be set based on the floodwater initiation, drainage time, and recovery duration, as long as at least the treatment initiation data, stress period data, early post-drainage recovery data, and late recovery data are obtained.

[0054] In one specific implementation, a drone equipped with a multispectral camera is used to acquire images in the blue, green, red, red-edge, and near-infrared bands. The drone follows a preset flight path, performing each mission at a fixed flight altitude, speed, forward overlap rate, and lateral overlap rate, and strives to conduct observations during similar daytime hours and under low wind and no-rain conditions.

[0055] Standard reflector images were acquired before and after each flight, and incident light conditions during the flight were recorded. The original band signals were corrected based on reflector data and / or incident light sensor data to generate orthorectified images for each band. Posterior relative normalization of the band signals could also be performed based on the cross-date statistical characteristics of the CK cell to reduce differences in irradiance conditions and sensor responses between flights. In this paper, the band signal ρ represents the corrected or posteriorly normalized comparable band signal.

[0056] Orthophotos of each date are transformed to a unified projection coordinate system. Using the image at the processing start time as a reference, cross-date spatial registration is performed in conjunction with ground control points or stable image features. Regions of Interest (ROIs) for each cell are determined on the image at the processing start time, and the same ROIs are applied to the registered images of each date. Records with missing images, abnormal stitching, boundaries exceeding the effective image range, or unreliable cross-date matching are marked or removed.

[0057] Within each ROI, the Normalized Difference Vegetation Index (NDVI) is calculated, and a threshold is used to distinguish wheat canopy from bare soil, shade, and other background elements. In one specific implementation, pixels with NDVI > 0.20 are defined as canopy pixels, and morphological opening and closing operations are performed sequentially using 3×3 pixel structuring elements to remove isolated noise and fill small pores within the canopy. Canopy cover (CC) is calculated according to the following formula:

[0058]

[0059] Where, N vegetation N represents the number of canopy pixels within the canopy mask. valid The number of effective pixels within the ROI is the output of the canopy structure variable used for dynamic response construction.

[0060] Cell-level spectral variables are calculated within the canopy mask. In one specific implementation, core variables include canopy cover (CC), soil-modified vegetation index (OSAVI), normalized green vegetation index (GNDVI), and normalized difference red edge index (NDRE), calculated as follows:

[0061]

[0062]

[0063]

[0064] Where, ρ NIR ρ Red ρ Green and ρ RE The corrected or normalized signals for near-infrared, red, green, and red-edge bands are respectively used. The soil-regulated vegetation index OSAVI is used to extract spectral phenotypes sensitive to canopy vitality, the green normalized vegetation index GNDVI is used to extract spectral phenotypes related to canopy chlorophyll and vitality, and the normalized differential red-edge index NDRE is used to extract spectral phenotypes related to canopy red-edge vitality, chlorophyll status, and senescence process.

[0065] Example 2: Paired Response, Baseline Centering, and Synthetic Trajectory

[0066] Candidate variables were subjected to a quality screening process relying solely on UAV data, including missing value rate, number of unique values, proportion of non-positive values, CK / WL pair integrity, and cross-date availability. For variables requiring log ratio calculation, only data that met the positive value criterion and had paired CK / WL records of the same genotype and date were retained. Representative variables could be further selected based on variable correlation, cross-date stability, and phenotypic coverage; endpoint yield or thousand-grain weight was excluded from this variable selection.

[0067] Please refer to Figure 2. Treatment / control paired response of the m-th variable for the g-th genotype at the t-th observation time point. Calculate according to the following formula:

[0068]

[0069] in, and These are the variable values ​​for the genotype under WL and CK conditions, respectively. This formula performs a logarithmic ratio transformation on the WL and CK values ​​for the same genotype, the same date, and the same variable, and outputs a paired response.

[0070] To mitigate paired differences before processing begins, baseline centering is performed on the paired responses at the processing start point t=0: Baseline-centered paired responses are used when higher variable values ​​indicate better canopy condition. , A negative value indicates that WL suppresses CK relative to WL, with lower values ​​indicating stronger suppression. For variables with opposite directions, their directions are unified before aggregation.

[0071] The baseline-centered paired responses of the selected M core variables are aggregated to obtain the comprehensive canopy response trajectory C. g (t):

[0072]

[0073] This formula aggregates core structural and spectral variables that are uniform in M ​​directions. In the equal-weighted implementation: In this embodiment, M=4, and the four variables are canopy coverage, OSAVI, GNDVI, and NDRE. Alternatively, preset non-negative normalized weights can be used to aggregate the selected structural and spectral variables, but the calculation rules should remain consistent within the same batch of materials.

[0074] Example 3: Three-Process Feature Calculation

[0075] Please refer to Figure 3. Let the start point of the waterlogging treatment be t0, and the drainage point be t. d The time window T under duress s =[t0,t d [Inner C] g (t) The maximum negative deviation from the baseline yields the maximum damage I. g :

[0076]

[0077] This formula extracts the maximum negative deviation of the overall canopy response within the stress time window Ts, which is used to quantify the maximum depth of relative canopy suppression during the waterlogging period.

[0078] The cumulative loss A is obtained by integrating the negative deviation within the stress time window. g :

[0079]

[0080] This formula is used to quantify the combined load of damage intensity and duration during the flooding period, A g The larger the value, the more severe the cumulative damage.

[0081] For discrete observation points, the area under the curve between adjacent observations can be calculated using the trapezoidal integral. In a specific implementation, T sThis refers to days 0 to 11 after treatment.

[0082] For I respectively g and A g After robust standardization and conversion to a direction where higher values ​​indicate less damage, the stress cumulative score S is obtained by equal-weighted or pre-weighted aggregation. g For example, S can be set to... g =[-Z(I g )-Z(A g )] / 2, where Z(·) is a robust normalization transformation predetermined based on the training group or the current group to be evaluated.

[0083] Conditional recovery scores are used to characterize the relative recovery efficiency at the same level of damage during drainage. First, within the early recovery time window T after drainage... r Inner Fit C g The slope of (t) as a function of time yields the early recovery slope β. 1g In a specific implementation, T r This refers to days 11 to 14 after treatment.

[0084] Calculate the damage gap D during drainage g To characterize the depth of damage at the point of recovery:

[0085]

[0086] The non-negative damage gap at the point of output drainage is used as a covariate for recovery slope correction.

[0087] The Huber loss was used to fit the relationship between the early recovery slope and the damage gap during drainage in the genotypic population:

[0088]

[0089] Where α0 and α1 are robust regression parameters, e g The regression residual for the g-th genotype is used to fit the relationship using Huber loss. The residual e is used as a conditional recovery phenotype to reduce the impact of differences in the damage initiation point and extreme observations.

[0090] For residual e g By unifying favorable directions and performing robust standardization, the conditional recovery score R is obtained. g In the stated direction, R g A higher value indicates that the genotype recovers faster than expected under given drainage gap conditions. Using Huber loss can reduce the impact of extreme observations on the damage correction relationship.

[0091] During the recovery period, time window T l The median L of the integrated canopy response trajectory was calculated internally. g And calculate residual damage Used to quantify negative canopy deviation that has not recovered after drainage:

[0092]

[0093] This formula converts the median response in the later recovery phase into non-negative residual damage, and then uses directional unification to form a residual effect score. For Q... g By unifying the favorable directions and performing robust standardization, the residual effect score F is obtained. g In a specific implementation, T l This refers to days 18 to 20 post-treatment. F g A higher value indicates a smaller residual negative deviation in the later stages of recovery.

[0094] The three process scores reflect the damage load during the stress period, the recovery efficiency at a given damage initiation point, and the residual state in the later stages of recovery, respectively. These scores are generated by outputting S... g R g and F g It can provide a process explanation for a single comprehensive ranking.

[0095] Example 4: D-SRFI and Pre-harvest Ordination

[0096] The scores from the three processes are integrated to obtain the dynamic stress recovery function index (D-SRFI) for the g-th genotype. g Used to integrate the three process dimensions into a pre-harvest material ranking score:

[0097] D-SRFI g

[0098] Among them, w S w R and w F All are non-negative and w S +w R +w F =1. The higher the D-SRFI, the more the material has milder stress-induced cumulative damage, higher damage correction and recovery efficiency, and smaller residual effects in the later stages of recovery.

[0099] In the equal weighting implementation, w S =w R =w F =1 / 3.

[0100] Genotypes are sorted from highest to lowest according to D-SRFI, and a list of waterlogging tolerance grades or pre-harvest candidate materials is generated based on preset quantity, preset proportion, or preset score threshold. The output results include at least the material identifier and D-SRFI; preferably, the maximum damage, cumulative loss, early recovery slope, damage gap during drainage, damage-corrected recovery residual, canopy response in the later recovery stage, and scores for the three processes are output simultaneously.

[0101] The pre-harvest candidate material list is used to narrow down the range of materials for subsequent yield measurement, artificial observation, physiological testing, or multi-environment rescreening. D-SRFI characterizes the dynamic canopy tolerance and recovery process relative to the corresponding CK, and is not equivalent to absolute high yield, nor is it used alone to determine whether the material has completed variety finalization.

[0102] Yield, thousand-grain weight, or corresponding retention rate are used as independent measured harvest-period validation information and are not included in the selection of core UAV variables, the construction of three-process features, or the D-SRFI calculation. Therefore, the method can output material ordination before harvest; after harvest, the consistency between the ordination and the final agronomic performance can be evaluated using measured endpoint traits.

[0103] Example 5: Systems, Devices, and Storage Media

[0104] Please refer to Figure 4. The data acquisition module 100 receives genotype identifiers, processing identifiers, observation time points, and multi-temporal UAV canopy images. The image processing module 200 performs image correction, orthorectification, cross-date registration, ROI management, canopy segmentation, and cell-level variable extraction.

[0105] The paired response construction module 300 matches CK / WL records according to the same genotype and the same observation time point, calculates the log-paired response, processes the baseline-centered response at the starting time point, and integrates the canopy response trajectory.

[0106] The process characteristic calculation module 400 generates the maximum damage, cumulative loss, early recovery slope, damage gap during drainage, damage-corrected recovery residual, and late-recovery canopy response from the comprehensive canopy response trajectory, and further generates S g R g and F g .

[0107] The comprehensive evaluation module 500 calculates the D-SRFI according to preset weights. The results output module 600 generates at least one of the following: a material score table, a ranking, a grade, and a candidate material list, and saves the results as an electronic results record.

[0108] The system can be implemented by one or more electronic devices. The memory of the electronic device stores program instructions and processing parameters, and the processor executes the program instructions to complete the aforementioned image data processing, dynamic phenotypic calculation, and result output. A computer-readable storage medium can store a program for causing the processor to execute the above methods.

[0109] Other implementation methods

[0110] While maintaining syngeneic CK / WL pairing, baseline centering at the treatment start point, and three-process decomposition, other structural or spectral variables that can stably generate positive paired responses can be selected based on sensor band configuration and canopy feature availability. Variable selection should rely solely on UAV data quality, pairing integrity, cross-date availability, redundancy, and phenotypic category coverage, without using back-selection variables based on the endpoint to be validated.

[0111] The specific time windows for the stress period, early recovery period, and late recovery period can be set relative to the actual start and drainage time of the flooding. The time windows should respectively cover the damage accumulation, the initial recovery after drainage, and the late recovery state, and the same definition should be used for the same evaluation batch.

[0112] The weights for D-SRFI can be equal or pre-defined non-negative normalized weights. If an unequal weighting scheme is used, the weights or weight determination rules should be fixed before processing the batch to be evaluated to avoid using the endpoint results to reverse the weighting of the same batch of materials.

[0113] Example 6: Verification of the End Point of Actual Harvest Date

[0114] After wheat matured, the CK and WL plots were harvested, threshed, and cleaned separately. After the grains were dried to a stable quality and the moisture content was uniformly corrected, the grain mass of each plot was weighed and the yield per unit area was calculated. The thousand-grain weight was repeatedly measured for thoroughly mixed grain samples, and the average value was taken. The yield retention rate YR for the g-th genotype was calculated. g TGWR (Ten Thousand Grain Weight Retention Rate) g Calculate according to the following formulas:

[0115]

[0116]

[0117] Among them, Y WL,g and Y CK,g TGW represents the measured yield per unit area for the g-th genotype under WL and CK conditions, respectively. WL,g and TGW CK,g These are the measured thousand-grain weights under the corresponding conditions.

[0118] In this embodiment, 902 wheat genotypes with complete CK / WL pairings, UAV dynamic phenotypes, and measured harvest traits were included in the validation. The endpoint trait was not used for UAV core variable screening, Huber regression relationship determination, three-process score construction, or D-SRFI weight setting.

[0119] A five-fold outer layer and a four-fold inner layer nested cross-validation, grouped by genotype, were employed, with measured yield retention rate and measured thousand-grain weight retention rate as prediction targets, respectively. Missing value imputation, standardization, Huber regression, principal component analysis, exponent calculation, and model parameter selection were all fitted only within the corresponding outer training fold before being applied to the outer test fold, to avoid leakage of the same genotype or endpoint information from the training process to the testing process.

[0120] When partial least squares regression was used, the outer test reduced-mean coefficients of determination (R²) for yield retention rate and thousand-grain weight retention rate of the 48-dimensional multi-date original core variables were 0.729 and 0.626, respectively; after compressing the original variables into three process scores of stress accumulation, conditional recovery, and residual effects, the corresponding R² values ​​were 0.754 and 0.651, respectively; after further integration into a one-dimensional D-SRFI, the corresponding R² values ​​were 0.610 and 0.531, respectively. Figure 5 The results of the one-dimensional D-SRFI were higher than those of the full-time principal component index (0.276 and 0.245) and the single-date baseline (0.189 and 0.167) on day 11 after processing.

[0121] Among the 902 genotypes, the Spearman correlation coefficients between D-SRFI and measured yield retention and measured thousand-grain weight retention were 0.823 and 0.640, respectively. When the top 20% of materials in terms of D-SRFI were used as the pre-harvest candidate set, this candidate set covered 80.2% of the top 10% of materials in terms of measured yield retention and 60.4% of the top 10% of materials in terms of measured thousand-grain weight retention, respectively.

[0122] Based on the pre-screening results of the top 20% of D-SRFI results before harvest, further confirmation was achieved using measured multi-trait data at harvest: yield retention rate not less than 0.7660, measured yield per unit area under WL conditions not less than 4839.5 kg / ha, and thousand-grain weight retention rate not less than 0.8655. A total of 50 genotypes simultaneously met the above conditions, forming a priority re-screening set supported by measured multi-trait data at harvest.

[0123] The thresholds and performance values ​​mentioned above are verification results obtained under specific experimental populations, locations, seasons, and data segmentation conditions in this embodiment. They are used to illustrate the feasibility and screening effect of the technical solution and do not constitute a limitation on the scope of protection of the claims, nor do they indicate that the 50 genotypes have been finalized. In practical applications, the candidate ratio or agronomic thresholds can be predetermined based on breeding objectives, population distribution, and rescreening resources.

[0124] This embodiment demonstrates that, without using the measured harvest endpoint in reverse for core UAV variable selection and D-SRFI construction, the three-process dynamic phenotypes are consistent with the measured yield and thousand-grain weight retention capacity; D-SRFI can narrow down the candidate range before harvest, while the measured endpoint after harvest is used for independent verification and multi-trait confirmation.

[0125] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any equivalent substitutions or improvements made within the scope of the technical concept and claims disclosed in this application shall fall within the protection scope of this application.

Claims

1. A method for pre-harvest identification of wheat waterlogging tolerance based on UAV dynamic phenotypes, characterized in that, include: Step S101: Obtain multi-temporal UAV canopy images of the same batch of wheat genotypes under normal moisture control and waterlogging treatment, wherein the multi-temporal UAV canopy images cover the waterlogging stress accumulation stage and the recovery stage after drainage. Step S102: Perform image correction, cross-date spatial registration, region of interest determination and canopy segmentation on the multi-temporal UAV canopy images, and extract the cell-level canopy structure variables and canopy spectral variables of each genotype at each observation time point; Step S103: Construct paired responses for the corresponding variables of the same genotype and the same observation time point under the waterlogging treatment and the normal water control, and use the paired responses at the treatment start time point to base-center the subsequent paired responses; Step S104: Perform orientation unification and aggregation on at least one canopy structure variable and at least one canopy spectral variable after baseline centering to obtain the comprehensive canopy response trajectory of each genotype; Step S105: Generate a stress accumulation score based on the maximum damage and cumulative loss of the integrated canopy response trajectory during the waterlogging stress accumulation stage; Based on the early recovery slope of the integrated canopy response trajectory after drainage, and by regressing the early recovery slope with the damage gap during drainage, a conditional recovery score is generated; based on the canopy response of the integrated canopy response trajectory in the later recovery stage, a residual effect score is generated, which together constitute the three process scores; Step S106: The stress cumulative score, conditional recovery score and residual effect score are unified and integrated to obtain the dynamic stress recovery function index of each genotype. Step S107: Output at least one of the following genotypes from the list of waterlogging tolerance, waterlogging tolerance level, and pre-harvest candidate materials, based on the dynamic stress recovery function index.

2. The method according to claim 1, characterized in that, The multi-temporal UAV canopy images are multispectral images, including at least red, green, red-edge, and near-infrared bands; the normal moisture control and the waterlogging treatment are set up in one-to-one corresponding plots for the same genotype, and UAV observations are carried out at the treatment start time, the waterlogging stress accumulation stage, the drainage time, and the recovery stage after drainage.

3. The method according to claim 1, characterized in that, Processing the multi-temporal UAV canopy images includes: Radiometric correction is performed based on standard reflector data and / or incident light sensor data to generate multispectral orthophotos; The multispectral orthophotos at each observation time point are transformed to a unified projection coordinate system, and cross-date spatial registration is performed with the image at the processing start time point as a reference. The region of interest for each cell is determined on the image at the starting time of the processing, and the region of interest is applied to the registered images at each observation time. Canopy pixels are distinguished from background pixels based on the normalized vegetation index threshold, and morphological opening and closing operations are performed on the segmentation results to generate a canopy mask.

4. The method according to claim 1, characterized in that, The canopy structure variables include canopy coverage, and the canopy spectral variables include soil-modified vegetation index, green normalized vegetation index, and normalized difference red edge index; the comprehensive canopy response trajectory is obtained by baseline-centered paired response weighted aggregation of the canopy coverage, soil-modified vegetation index, green normalized vegetation index, and normalized difference red edge index.

5. The method according to claim 1, characterized in that, The paired response of the g-th genotype to the m-th variable at the t-th observation time point. and baseline-centered paired response Calculate according to the following formulas respectively: in, and These are the variable values ​​under the waterlogging treatment and the normal moisture control, respectively. To handle paired responses at the start time.

6. The method according to claim 1, characterized in that, The generation of the stress cumulative score includes: determining the maximum damage based on the maximum negative deviation of the integrated canopy response trajectory relative to the baseline within the stress time window from the start of treatment to the drainage time window; performing time integration on the negative deviation of the integrated canopy response trajectory within the stress time window to obtain the cumulative loss; and aggregating the maximum damage and the cumulative loss after unifying and standardizing them in a favorable direction to obtain the stress cumulative score. The generation of the conditional recovery score includes: fitting the slope of the comprehensive canopy response trajectory within the early recovery time window starting from the self-drainage point to obtain the early recovery slope of the g-th genotype. ; according to Damage gap during drainage calculation ,in For the g-th genotype at the drainage time point The comprehensive canopy response was obtained; Huber loss was used for fitting. and the residual The favorable direction is unified and standardized to obtain the conditional recovery score, where α0 and α1 represent robust regression parameters, respectively; The generation of the residual effect score includes: calculating the median of the comprehensive canopy response trajectory within the late recovery time window. ,according to Residual damage was obtained The residual damage is then standardized and oriented in a favorable direction to obtain the residual effect score.

7. The method according to claim 1, characterized in that, The Dynamic Stress Recovery Function Index (D-SRFI) g Calculate, where S g R g and F g These represent the cumulative stress score, conditional recovery score, and residual effect score for the g-th genotype, respectively. S w R and w F The weights are non-negative and sum to 1.

8. The method according to claim 1, characterized in that, Step S107 outputs at least one of the following: maximum damage, cumulative loss, early recovery slope, damage gap during drainage, recovery residual after damage correction, canopy response in the later recovery stage, and process score, to provide a process explanation and data audit basis for material sorting.

9. A pre-harvest identification system for wheat waterlogging tolerance based on UAV dynamic phenotypes, characterized in that, include: The data acquisition module is used to acquire multi-temporal UAV canopy images of the same batch of wheat genotypes under normal water control and waterlogging treatment, covering the waterlogging stress accumulation stage and the recovery stage after drainage. The image processing module is used to perform image correction, cross-date spatial registration, cell region determination and canopy segmentation on the multi-temporal UAV canopy images, and extract cell-level canopy structure variables and canopy spectral variables; The paired response construction module is used to construct paired responses between waterlogging treatment and normal water control with the same genotype and the same observation time point, and to center the baseline at the treatment start time point to form a comprehensive canopy response trajectory. The process feature calculation module is used to generate stress cumulative score, conditional recovery score after drainage damage gap correction, and residual effect score based on the comprehensive canopy response trajectory. The comprehensive evaluation module is used to unify and integrate the scores of the three processes to obtain the dynamic stress recovery function index. The results output module is used to output at least one of the following for each genotype: waterlogging tolerance ranking, waterlogging tolerance level, and pre-harvest candidate material list, based on the dynamic stress recovery function index.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 9.

Images