CCO imaging (Fig. 1 A and B) represents an advanced photography technique facilitating rapid imaging in field conditions. The route of CCO imaging consists of multiple single-circle routes (Fig. 1 A). In this route, the overlapping parts can be divided into two categories: intra-circle overlap and inter-circle overlap. The intra-circle overlap refers to the overlapping areas between images within a single circular path; while inter-circle overlap refers to the overlapping parts between two adjacent single circular reconstruction areas. As shown in Fig. 1 A, an example of a CCO imaging strategy is illustrated, featuring a sequence of four distinct circles, each designed with a 50 ​% overlap with its neighboring circle. The area that is most conducive to image reconstruction is marked in blue, whereas the red marking delineates the square inscribed within this optimal reconstruction zone. The condition of a 50 ​% overlap between circles is met when the edges of the inscribed squares of two adjacent circles are just touching. It is stipulated that both the overlaps within the CCO imaging must not fall below a 50 ​% threshold; otherwise, the imaging process may fail to reconstruct the image accurately. Several key parameters, including the elevation (H) of the CCO imaging path, the angle at which the camera is positioned (θ), the extent of overlap within individual circles, and the overlap between different circles, are crucial and must be tailored to fit the specific demands of the mission. The term ‘route height’ describes the vertical distance in relation to the DSM and is dependent on the GSD required for the mission objectives. This can be calculated using the following formula [16]:

where H represents the relative height, f signifies the focal length, a denotes the pixel size, and θ stands for the lens tilt angle. For the DJI Phantom 4 RTK drone, the values of f and a are 8.8 ​mm and 0.0024 mm/pixel, respectively. The radius (R) of the CCO imaging route is calculated using the following formula:

The intra-circle overlap (ICO) can be computed based on the number of images captured within a single circle, using the following equation:

where FOV represents the field of view of the lens, and LA signifies the count of images captured within a single circle. Taking into account a chosen inter-circle overlap (ITO), one can compute the distance D between the centers of neighboring circles within the intersecting circular pattern. This computation is visually elucidated in Fig. 1 C, and the calculation is as follows:

where y denotes the physical dimension of the sensor in the y-direction and. ∂ represents half of the FOV of the lens in the y-direction.

A collection of 262 F5 recombinant inbred lines (RILs) was generated by iteratively self-pollinating a genetic cross between two wheat cultivars, Zhongmai 578 (ZM578) and Jimai 22 (JM22). These RILs were used for phenotypic evaluations and QTL mapping. ZM578, a leading cultivar in 2020 and 2021, is widely cultivated across the southern and northern regions of the Winter Wheat Region of the Yellow and Huai Rivers (YHRVWWR) [6]. Known for its high yield, superior bread-making quality, robust disease resistance, and resilience to adverse conditions, it covers approximately 0.4 million hectares. JM22, introduced in the northern YHRVWWR in 2006, was released in Anhui in 2010 and Henan in 2011, catering to the southern YHRVWWR [6]. The parents and 262 RILs were grown in Xinxiang (113°46′ E, 35°8′ N) during the 2021–2022 and 2022–2023 growing seasons. They were also planted in Dezhou (116°20′ E, 37°17′ N) for the same two seasons and in Gaoyi (114°36′ E, 37°38′ N) during the 2021–2022 season. In Xinxiang, experiments included two irrigation treatments: full irrigation and limited irrigation, with approximately 250 ​mm of water applied during the tillering phase. The full irrigation treatment received additional water during the jointing, heading, and grain filling stages. Experiments in Dezhou and Gaoyi used full irrigation. All experiments followed a randomized complete block design with three replications, resulting in 786 plots per experiment for the RIL population. Each plot, representing a cultivar, covered 3.6 ​m² (3 ​m ​× ​1.2 ​m) and was arranged in six rows with 0.2 ​m spacing. The numbering of each environment and seeding dates are detailed in Table 1. Fertilization and pest management were optimized based on local soil and climate conditions. Additionally, a diverse panel of 120 wheat varieties was cultivated in Xinxiang and Zhoukou (114°39′ E, 33°37′ N) during the 2018–2019 growing seasons under full and limited irrigation, respectively. The PH data from this panel were used to validate the efficacy of kompetitive allele-specific PCR (KASP) markers.

CCO imaging was performed using the DJI Phantom 4 RTK drone (SZ DJI Technology Co., Shenzhen, China) (Fig. 2 A), equipped with a 20-megapixel RGB lens. This drone integrates real-time kinematic (RTK) technology to ensure centimeter-level precision positioning. The flight routes were automatically planned using Waypoint Master software (Weber Intelligent Control Technology Co., Ltd., Beijing, China), with input from KML and DSM files relevant to the target area (Fig. 2 B). The camera viewing angle for CCO imaging was set to 45°. To ensure precise point cloud data acquisition, the flight altitude for CCO imaging routes across all growth environments was maintained at 4 ​m above the canopy (Fig. 2 C), resulting in a ground sampling distance (GSD) of 1.5 ​mm per pixel. To increase data collection efficiency, two drones were operated simultaneously. Specific flight dates for each environment are provided in Table 1.

The UAV flights and manual PH (field-measured PH) measurements were conducted during the grain filling stage. For manual PH measurements, three representative plants were selected from consistent growth areas within each plot. The height of each plant was measured using a ruler, and the average of these three measurements was used to determine the PH for that plot.

The acquired field image sequences (Fig. 2 D) were imported into DJI Terra v3.7.6, and 3D models were generated using the SfM-MVS algorithm [32] (Fig. 2 E–F). The resulting point clouds were exported in LAS format. The detailed processing flow of the 3D point clouds data is shown in Fig. 2 G-J. First, a DSM was created using the rasterize_canopy function from the “lidR” package (available at lidR GitHub) in R v4.2.2. Polygons were then outlined on the DSM in ArcMap v10.5 to differentiate between different plots. The “clip_roi” function from the “lidR” package was used to segment the point clouds using LAS files and a shapefile containing polygonal elements. Finally, the point clouds were denoised using the “classify_noise” function with the statistical outlier removal algorithm. In the 3D point clouds data analysis, 3D-PH was determined by measuring the vertical distance between the ground level and the upper quantile of the Z-coordinates of the points. In some growing environments (E1, E2, E4, and E5), where many ridges were present above ground level, using the low quantile as ground level could lead to erroneous results. Therefore, a manual method was used to select the ground level of the plots in the point clouds (Fig. 2 K). Various height quantiles were utilized to define multi-level 3D-PH. To minimize errors, PH was averaged across three replicates of each line in each environment. The Pearson correlation coefficient (r), root mean square error (RMSE), and relative RMSE (RRMSE) were used as evaluation metrics.

where n is the number of samples; yi and yˆi represent the measured and the extracted PH of sample i, respectively. yi¯ is the mean value of the actual PH, respectively. Higher values of r and lower values of RMSE and RRMSE indicate better accuracy in extracting PH.

Statistical analysis encompassed the application of analysis of variance (ANOVA) to evaluate the statistical significance linked with various sources of variation. To mitigate the influence of genotypic factors on the field-measured PH (FM-PH) and 3D-PH from point clouds, the best linear unbiased estimation (BLUE) was computed across seven different environments. The aim was to eliminate any potential bias introduced by the genotypic design. ANOVA and BLUE estimation were performed using the linear model in the QTL IciMapping software [33]:

where i=1,2,…,n, j=1,2,…,e, k=1,2,…,r, μ represents the overall mean of the population Rk/j is the k th replication effect in the j th environment, Gi represents the genotypic effect of the i th genotype, Ej represents the environmental effect of the j th environment, GEij represents the interaction effect between the i th genotype and j th environment, and εijk represents the random error effect which was assumed to be normally distributed with a mean of zero. After establishing the linear model for ANOVA, the collective degrees of freedom and the sum of squares can be partitioned according to the elements outlined within the linear model [34]. This division allows for the determination of the mean square value for each distinct source of variation, subsequently facilitating significance testing. Heritability indicates the proportion of genetic variation relative to the overall phenotypic variation. The calculation for broad-sense heritability is outlined below [35]:

where σG2,σGE2, and σε2 represent genotype, genotype × environment interaction, and residual error variances, respectively, e and r were the numbers of environments and replicates, respectively. Correlation analysis and data visualization were performed using R v4.2.2.

Genotyping of the RILs and two parental lines was conducted using the wheat 50K SNP array, available through CapitalBio Corporation (Beijing, China; http://www.capitalbio.com). This array comprises 55,224 SNPs, evenly distributed across the 21 wheat chromosomes. To construct a genetic map, markers that were monomorphic between the parents and those with high rates of missing data (exceeding 20.0 ​%) or a minor allele frequency (MAF) below 0.3 were excluded. This filtering process resulted in a set of 9661 high-quality polymorphic markers, which were then utilized for subsequent analysis. Duplicate markers were removed using the BIN function in lciMapping v4.2 [36], leaving 1501 non-redundant markers for linkage analysis using the regression mapping algorithm in JoinMap 4.0 Software. A genetic map consisting of 1501 bin markers was constructed across all chromosomes except for 6B. This linkage map covered a total length of 2384.95 ​cM, with an average genetic distance of 1.59 ​cM per bin marker. Among the three genomes, the A genome had the highest number of markers (562), followed by the B genome (545) and the D genome (394), as outlined in Table S1.

QTL analysis was performed using the inclusive composite interval mapping (ICIM) function within the IciMapping v4.2 software. Specific mapping parameters were set, including step ​= ​0.1 ​cM and PIN ​= ​0.001, while the logarithm of odds (LOD) threshold was established at 2.5. QTLs identified across at least three distinct environments were considered stable. Favorable alleles were defined as those resulting in decreased values. Following the International Rules of Genetic Nomenclature [37], the nomenclature of all QTLs was assigned.

Marker physical positions were determined utilizing wheat genome sequences obtained from the International Wheat Genome Sequencing Consortium (IWGSC RefSeq 1.0, accessible at http://www.wheatgenome.org/) [38]. Allele-specific primers for KASP assays were meticulously designed using the PolyMarker tool (available at http://polymarker.tgac.ac.uk/). The primer mixture for the assays consisted of 46 ​μL of ddH2O, 30 ​μL of a common primer (100 ​μmol L^-1), and 12 ​μL of each tailed primer (100 ​μmol L^-1). Assays were prepared in 384-well formats, resulting in approximately 3 ​μL reaction volumes each. These reactions included 10–20 ​ng μL^-1 of DNA, 3 ​μL of a 1 ​× ​KASP master mixture, and 0.056 ​μL of the primer mixture. The PCR cycling protocol comprised an initial hot start at 95 ​°C for 15 ​min, followed by 10 touchdown cycles (95 ​°C for 20 ​s; starting at 65 ​°C and decreasing by −1 ​°C per cycle for 25 ​s), and subsequently, 30 additional cycles of denaturation and annealing/extension steps (95 ​°C for 10 ​s; 57 ​°C for 60 ​s). PCR was performed using a Bio-Rad CFX Real-Time PCR System, and fluorescence signals were detected using the Bio-Rad CFX Manager 3.1 Software. To ensure accuracy, KASP markers underwent initial validation using the parental lines. Following validation, these markers were utilized for screening both RIL and natural populations.

Based on the Chinese Spring reference genome (IWGSC RefSeq v.1.1), the genes located within or adjacent to the physical intervals of the QTLs identified in this study were selected for screening. Using the flanking sequences of SNP markers during the confidence interval of QTL as probes, the candidate genes related to PH were explored by BLASTn in NCBI (http://www.ncbi.nlm.nih.gov/). Furthermore, the expression of candidate genes was evaluated using the Wheat Expression Browser (http://www.wheat-expression.com/) to identify genes with high expression levels in PH.

Nadir imaging is currently the most popular low-cost method for extracting crop height [[39], [40], [41], [42], [43]]. In addition to CCO imaging, nadir imaging was also applied in environment E5, with the same flight altitude and overlap settings as used in CCO imaging. To increase the variation in PH, an additional 24 plots outside the RIL population were included. Compared to nadir imaging, CCO imaging provided more complete details in point clouds, particularly in the front areas of the plots (Fig. S1 A and B). PH data were extracted using 11 height quantiles. Across all quantiles, CCO imaging consistently achieved higher r and lower RMSE values than nadir imaging (Fig. S1 C and D). Scatter plots comparing PH from optimal quantiles with FM-PH showed that CCO-derived PH aligned more closely with the 1:1 line (Fig. S1 E and F). Among the various quantiles of point clouds generated from CCO imaging, the 90 ​% quantile aligned most closely with FM-PH, while lower percentiles risked extracting stem height instead of canopy height. Therefore, in subsequent analyses, the 90 ​%, 92 ​%, 94 ​%, 96 ​%, 98 ​%, and 99.5 ​% quantiles were selected for multi-level PH extraction.

The detailed point clouds from CCO imaging of wheat plots at different heights are shown in Fig. 3 A–C. CCO imaging successfully reconstructed the canopy model at the organ scale, with spikes in the point clouds being clearly recognizable. This provides a basis for accurate multi-level PH extraction, as different height quantiles correspond to the different heights of plants within the plots. However, capturing side views can be challenging due to the close proximity of adjacent experimental plots, resulting in the absence of points at the lower part of the plot sides. When the distance between experimental plots is sufficient, the details of plants from various angles can be precisely reconstructed [16]. The Z-axis direction of the canopy was divided into four layers from top to bottom, and the distribution of point clouds projected to the bottom of each layer is shown in Fig. 3 D–F. The distribution pattern of point clouds was similar across plots of different heights. Points were mainly concentrated in the canopy portion (first and second layers), where spikes and leaves were concentrated. Fewer points were found in the lower and bottom portions (third and fourth layers). By statistically analyzing the points in the Z-axis direction (Fig. 3 G–I), the inhomogeneous distribution of point clouds can be understood in more detail. However, due to the accurate reconstruction of the canopy, this phenomenon had a negligible effect on the PH extraction.

Both the FM-PH and CCO imaging multi-level 3D-PH followed the normal distribution within each environment, as shown in Fig. 4 A and Fig. S2. The distribution pattern of BLUE values for each quantile height was consistent with that for FM-PH (Fig. 4 B). Moreover, 3D-PH extracted from different quantiles exhibited strong correlations with each other (Fig. 4 C). Among the six levels of 3D-PH, the 90 ​% height quantile closely aligned with the mean value of FM-PH. Additionally, the among-environment correlations for FM-PH (Fig. 4 D) and the 90 ​% height quantile (Fig. 4 E) were high and significant. Significant variances (P ​< ​0.001) were observed among genotypes across different environments, as well as within the genotype ​× ​environment (G ​× ​E) interactions, for both FM-PH and 3D-PH (Table S2). The broad-sense heritability within each environment ranged from 0.775 to 0.959 (Fig. 4 F), and from 0.975 to 0.982 across all seven environments. Notably, the heritability of most 3D-PH levels was higher than that of FM-PH.

The 3D-PH demonstrated strong correlations with FM-PH across all seven growth environments (Fig. 5 and Table 2). The 3D-PH measured at the 90 ​% and 92 ​% quantiles closely resembled FM-PH, with RMSE values less than 2 ​cm in most environments. For BLUE values, the correlation coefficient between FM-PH and 3D-PH reached an impressive 0.99, with an RMSE of only 0.65 ​cm in the 90 ​% height quantile. In the range of 94 ​%–99.5 ​% quantiles, the 3D-PH was slightly elevated compared to FM-PH, potentially reflecting the height of taller plants within the canopy. Overall, the success of CCO imaging in reconstructing detailed canopy models at the organ scale offers a valuable alternative for PH assessment in precision agriculture and breeding programs.

QTLs were identified for both FM-PH and multi-level 3D-PH across multiple growth conditions. Across the seven environments, a total of 106 QTLs associated with FM-PH and 3D-PH were identified on various chromosomes (Fig. 6): 1D (2), 2A (8), 2B (27), 2D (6), 3A (7), 3B (6), 4A (6), 4D (2), 5A (4), 5B (8), 5D (3), 6A (1), 6D (1), 7A (10), 7B (8), and 7D (7). These QTLs explained 1.1–18.8 ​% of the phenotypic variances. Among the 60 QTLs detected by FM-PH, 40 were also detected by 3D-PH. As expected, loci detected by 3D-PH extracted from different quantiles showed remarkable consistency (Fig. 6 A). Out of the 86 3D-PH loci, 38 were consistently identified by all six types of 3D-PH, while 33 loci were detected by two to five types of 3D-PH concurrently. Additionally, there were similarities in the distribution patterns of loci identified for FM-PH and 3D-PH across different environments and chromosomes (Fig. 6 B and C).

The stable loci detected by FM-PH and 3D-PH were summarized in Fig. 7, and their details can be found in Tables S3 and S4. For FM-PH, five stable loci were detected, distributed on 2B, 4A, 7A (2), and 7B. QPhzj.caas-2B.1, QPhzj.caas-7A.1, QPhzj.caas-7A.2, and QPhzj.caas-7B were detected in four environments, explaining 14.65 ​%, 3.69 ​%, 2.85 ​%, and 2.80 ​% of the phenotypic variances, respectively. For the QTLs on chromosome 7A, the positive alleles were contributed by ZM578, whereas those on 2B, 4A, and 7B were from JM22. Compared with FM-PH, multi-level 3D-PH detected more stable loci. Specifically, 3D-PH based on 90 ​%, 92 ​%, 94 ​%, 96 ​%, 98 ​%, and 99.5 ​% quantiles detected 6, 6, 8, 6, 6, and 6 stable loci, respectively. Excluding overlapping loci, a total of 11 stable loci were detected, distributed on 2B, 3A (2), 3B, 4A, 4D, 5A, 7A (2), 7B, and 7D. Different quantities may represent flora at different heights within a plot, which is difficult to measure manually and relate to genetic data. Among the 11 loci, QPhzj.caas-3A.2, QPhzj.caas-3B, QPhzj.caas-4A.2, QPhzj.caas-4D, QPhzj.caas-7A.2, and QPhzj.caas-7D were identified across four distinct environments, explaining 3.52 ​%, 3.02 ​%, 4.44 ​%, 2.30 ​%, 3.62 ​%, and 8.52 ​% of the phenotypic variations, respectively. Moreover, QPhzj.caas-7A.1 was detected in five environments (phenotypic variation explained ​= ​3.48 ​%). Notably, for the QTLs located on chromosomes 4D, 5A, and 7A, the positive alleles originated from ZM578, while those associated with 2B, 3A, 3B, 4A, 7B, and 7D were contributed by JM22. Among the stable loci detected by 3D-PH, Qphzj-caas-3B, QPhzj.caas-4A.2, and QPhzj.caas-7A.2 were detected simultaneously by all quartiles. It should be emphasized that three (QPhzj.caas-7A.1, QPhzj.caas-7A.2, and QPhzj.caas-7B) of the five stable loci detected by FM-PH could be detected by 3D-PH. The discovery of multiple stable loci through the multi-level 3D-PH approach suggests that this alternative metric may offer additional insights into the genetic architecture of PH, further enriching our understanding of this important trait.

After comparing the detected loci with the previously reported loci (See Discussion for details), QPhzj.caas-3A.2 (3D-PH) and QPhzj.caas-7A.1 (3D-PH and FM-PH) emerge as potential new loci. Flanking markers closely linked to these two loci, AX-94626731 (QPhzj.caas-3A.2) and AX-95217657 (QPhzj.caas-7A.1), were converted into KASP markers and named K_AX-94626731 and K_AX-95217657, respectively. The primer sequences are detailed in Table S5. The KASP markers K_AX-94626731 and K_AX-95217657,were successfully distinguished the two parental genotypes (ZM 578/JM 22). These two KASP markers were employed to genotype 120 natural population varieties from Xinxiang and Zhoukou under two irrigation conditions: full irrigation and limited irrigation treatments. The genetic effects were validated, and the PH of different allele genotypes were depicted in Fig. 8.

The genotyping results of the K_AX-94626731 marker indicated that there were 65 lines with the ZM578 genotype (TT) and 53 lines with the JM22 genotype (CC) (Fig. 8 A). T-test results reveal that the difference in PH between the two genotypes of QPhzj.caas-3A.2 across different environments has a P-value ranging from 0.001 to 0.03, reaching a significant level (P ​< ​0.05) (Fig. 8 B). Lines carrying the superior allele exhibit 4.6 ​%–9.4 ​% lower PH compared to those without the superior allele. Regarding the K_AX-95217657 marker, there are 53 lines with the ZM578 genotype (AA) and 65 lines with the JM22 genotype (GG) (Fig. 8 C). T-test results indicate that the difference in PH between the two genotypes of QPhzj.caas-7A.1 across different environments has a P-value ranging from 0.005 to 0.02, reaching a significant level (P ​< ​0.01) (Fig. 8 D). Lines carrying the superior allele exhibit 5.8 ​%–11.0 ​% lower PH compared to those without the superior allele (Table S6).

A total of two candidate genes were identified in this study. The physical interval of QPhzj.caas-3B on chromosome 3B spanned from 6.80 ​Mb to 35.34 ​Mb. Within this interval lies a previously named PH-related gene, Rht5. Previous studies have placed Rht5 within a 1 ​Mb region on chromosome 3BS. Based on the Chinese Spring reference genome (IWGSC RefSeq v1.1), TraesCS3B02G025600 is identified as the candidate gene for Rht5, encoding a cytochrome P450 involved in gibberellins (GAs) synthesis. The reduced PH conferred by Rht5 may result from a defect in GAs biosynthesis [44,45]. KASP marker developed based on Rht5 was used to genotype the parents, with ZM578 genotype being AA and JM22 genotype being GG. Subsequently, the effects of this gene were examined in the RIL population using the same KASP markers (Fig. 8 E). T-test results at five quantiles in seven environments indicated that lines carrying the ZM578 genotype (AA) exhibit significantly lower PH (P ​< ​0.05) (Fig. 8 F). On chromosome 5A, the physical interval of QPhzj.caas-5A spanned from 569.55 ​Mb to 572.83 ​Mb. Within this interval lies a cloned gene TaGL3-5A (TraesCS5A02G373900) associated with grain length and thousand grain weight. Moreover, this gene showed high expression in stem (Fig. S3). KASP markers developed based on this SNP were used to genotype the parents, with ZM578 genotype being AA and JM22 genotype being GG. Subsequent examination of its effects in the RIL population revealed, through T-test results at the 90 ​% and 98 ​% quantiles in seven environments (Fig. 8 G), that lines carrying the ZM578 genotype (AA) exhibit significantly higher PH (P ​< ​0.05) (Fig. 8 H). Therefore, TaGL3-5A probably serves as a candidate gene for QTLs on chromosome 5A.

Currently, digital extraction of PH using aerial photography stands out as one of the most promising phenotyping methods for practical application in breeding work. This method only requires point clouds to derive PH measurements without model training [14,16,30]. Digital camera-based surround imaging serves as the predecessor to CCO imaging and has long been utilized in crop 3D modeling [46]. However, the application of this method to large-scale field crop measurements is challenging. The emergence of multi-rotor UAV has extended the application of this technology from indoor to outdoor environments. The results of this study suggest that UAV-based CCO imaging is a robust approach for measuring PH, and it can be used in the future to capture the temporal variations in PH across various crops. Although CCO imaging has the disadvantages of not being able to penetrate to shaded plants and being susceptible to light variations compared to LiDAR, its simplicity and low cost allows it to be replicated at all levels of breeding institutions. The 3D-PH extraction error of the CCO imaging for all samples before averaging over three replications in each environment are summarized in Table S7. For Q90 and Q92, the 3D-PH extraction error in most environments was below 3 ​cm using FM-PH as the evaluation criterion. Additionally, we have summarized the accuracy of PH measurements across a range of crops utilizing different sensing platforms (Table S8). For wheat, the RMSE for PH assessment ranged from 4.6 to 10.0 ​cm, whereas for taller crops like maize, the RMSE spanned from 16.0 to 23.7 ​cm. For soybean, rice, potato, cotton, and oilseed rape, RMSE values fell within the range of 2.60–8.6 ​cm. The RRMSE values, reflecting the relative errors in PH extraction across various studies, varied from 4.98 ​% to 13.32 ​%. The reported errors in these studies notably surpass the errors observed in our study. While recognizing differences in plant material and growing conditions, this finding still underscores the great potential of CCO imaging methods for crop PH assessment. In the future, it will be necessary to systematically compare the accuracy of measuring PH using CCO imaging and various terrestrial laser scanning platforms within the same growth environment.

When exploring the proximity of different PH quantiles to the FM-PH, it was found that the 90 ​% and 92 ​% quantiles were more accurate than the 95–99 ​% quantiles typically preferred in previous studies [14,47]. This discrepancy may be due to differences in planting materials, growing conditions, or manual measurement methods, making it challenging to apply uniform standards across experiments. For instance, when measuring the height of wheat plants, we ignored the height of the awns, which may be roughly represented in the 3D model. This can result in the inclusion of awn point clouds in the higher quantiles, leading to deviations from FM-PH. Conversely, in other experiments, when manually measuring PH, the highest point of the crop canopy may be directly considered as the top of the plant, coinciding with the higher quantile of the canopy 3D model. To address this issue, we recommend conducting preliminary experiments, such as manually measuring PH in a randomly selected subset of plots in the field, to determine the optimal quantile for 3D-PH in a given environment. Subsequently, PH data can then be extracted for the majority of plots based on these determined quantiles. It is worth noting that despite the large error in 3D-PH at the high quantiles, they still found multiple stable loci (Fig. 7) and even the potential new loci. In manual measurements, taller plants may be overlooked, while medium-growth plants would be measured as the PH for the plot. The results this study suggest that these taller plants also harbor important genetic information, and their heights can be easily extracted through point cloud data.

Overall, this study presents a workable method for precise wheat PH measurement using multi-level 3D plant data obtained through UAV CCO imaging. Compared to LiDAR, it offers the advantage of lower cost, and it provides higher accuracy than nadir imaging. Multi-level 3D PH data extracted from point clouds reveal more detailed information about PH characteristics and help identify more QTLs associated with specific PH. This rich information is expected to provide valuable molecular markers for wheat breeding, enabling breeders to more precisely select and improve varieties with ideal PH, thereby increasing crop yield and adaptability.

Since CCO imaging is capable of organ-scale 3D reconstruction, it can also be used to assess crop leaf traits such as leaf length and leaf width, which is difficult to achieve with airborne LiDAR. However, the lower flight altitude can be time-consuming when applied in large-scale breeding experiments. For example, we used a DJI Phantom 4 aircraft equipped with 20 megapixels to execute the CCO route. With two drones working in tandem, it was able to complete the imaging in about two and a half hours. In practice, this efficiency is not satisfactory. While the addition of additional aircraft can reduce image acquisition time, it inevitably requires additional operators and carries the risk of multiple drone collisions. To address this challenge, in the future, the flight efficiency of CCO imaging can be improved through two aspects. Firstly, we can explore the implementation of autonomous flight systems utilizing advanced artificial intelligence algorithms. This system is expected to optimize the CCO routes of multiple aircraft, reducing overlap and maximizing coverage area, thereby ensuring safe and efficient operation. Importantly, this system does not require the involvement of additional operators, further enhancing operational convenience and efficiency. Secondly, raise the altitude of the flight to reduce the flight time. This method requires cameras with higher pixel counts, resulting in higher spatial resolution. In our previous study [16], we explored the relationship between flight altitude and spatial resolution for cameras with different pixels, and the results showed that when using a camera with 11648∗8736 pixels, the same spatial resolution as in this study could be obtained at a flight altitude of 17 ​m, which would greatly save flight time. Moreover, the process of 3D model reconstruction using the SfM-MVS algorithm is also time-consuming. Firstly, it requires improvements in computer resources. One feasible approach is to utilize distributed computing resources, decompose tasks, and run them simultaneously on multiple computers to enhance processing efficiency. Additionally, further research and optimization of 3D model reconstruction algorithms can be conducted to reduce computational complexity and improve algorithmic parallel performance.

QTLs or genes influencing PH in wheat have been identified across all 21 chromosomes [2]. However, no study has yet investigated QTLs using different quantiles of PH from 3D point clouds. While the number of stable loci (6–8) identified for each quantile did not notably exceed that of the FM-PH (5), complementation between different quantiles significantly increased both the quantity (11) and abundance of localized loci (Fig. 7). This phenomenon underscores the genetic diversity and environmental influences across various quantile ranges. This complementarity allows for a more complete understanding of changes in PH and helps to reduce errors or biases associated with single quantile measurements and manual measurements.

For FM-PH loci, two loci (QPhzj.caas-2B.1, QPhzj.caas-4A.1) that were not detected by 3D-PH were analyzed. QPhzj.caas-2B.1 (159.60–164.43 ​Mb) shares a similar position with the QTL reported by Hai et al. [48]. QPhzj.caas-4A.1 (46.12–59.32 ​Mb) is at similar positions to QPh.caas-4A.1 [6]. Nine out of the 11 loci identified by the 3D-PH were found to be consistent with those reported in previous studies. The QPhzj.caas-2B.2 spanned a physical interval from 767.13 to 770.68 ​Mb. This interval is consistent with a previously identified PH QTL from 767.13 to 775.17 ​Mb [6]. The physical intervals of QPhzj.caas-3A.1 and QPhzj.caas-3A.2 located on chromosome 3A were 466.11–477.79 ​Mb and 511.07–511.75 ​Mb, respectively. Near the physical interval of QPhzj.caas-3A.1, a previously identified PH-related QTL, with a physical position of 486.7 ​Mb, has been excavated by Rustgi et al. [49]. The interval where QPhzj.caas-3A.2 was located was relatively distant from all previously identified PH-related QTLs up to now [49,50]. Therefore, this locus may represent a novel QTL. QPhzj.caas-4A.2 was mapped to a physical interval of 570.26–570.47 ​Mb, at a similar position as V1.0:582124262 [51]. QPhzj.caas-4D was located within the physical interval of 35.35–35.99 ​Mb, shares a similar position with the QHt.cd-4D [52] and QPH. caas-4D [53]. On chromosome 7A, the physical intervals of QPhzj.caas-7A.1 and QPhzj.caas-7A.2 were 115.81–152.83 ​Mb and 671.47–675.39 ​Mb, respectively. The interval of QPhzj.caas-7A.1 was relatively distant from the currently identified PH-related QTLs on chromosome 7A. Therefore, it was speculated to be a potential novel QTL. A Previous study [6] have identified a QTL within the interval of QPhzj.caas-7A.2, associated simultaneously with thousand-kernel weight, kernel length, kernel width, average grain-filling rate and kernel number per spike. Further investigation revealed the presence of the cloned gene WAPO-A1 (TraesCS7A02G481600) within this interval, regulating spikelet number per spike and kernel number per spike. The previous study utilized the ZM578/JM22 RIL population to analyze the polymorphism between the parental gene sequences and developed KASP markers to identify favorable haplotypes. The results showed that lines carrying alleles from ZM578 exhibited higher thousand-kernel weight, kernel length, kernel width, and PH [6,54]. On chromosomes 7B and 7D, the physical intervals of QPhzj.caas-7B and QPhzj.caas-7D were 134.26 ​Mb–184.20 ​Mb and 248.80 ​Mb–340.68 ​Mb, respectively. A previously identified PH-related QTL, QHt.ipk-7B, coincides with the physical location of the QPhzj.caas-7B [55]. The position of QPhzj.caas-7D is similar to the locus reported by Zhang et al. [56].

Currently, numerous genes regulating PH have been identified. The wheat gene annotations provided by IWGSC RefSeq v.1.1 are crucial for screening candidate genes. Based on stable QTLs identified through 3D-PH, this study predicts two candidate genes. Previous studies have indicated that Rht5 exhibits pleiotropic effects on PH, spike length, stem length, and grain length [44]. Extensive research has revealed the significant role of gibberellins (GAs) in PH regulation [57]. Genes known to regulate wheat PH, based on their response to exogenous GAs, are classified into gibberellin-sensitive types such as Rht12 and Rht8, and gibberellin-insensitive types such as Rht1 and Rht3 [58]. Rht5 is a gibberellin-sensitive dominant gene induced by the Marfed mutation, which reduces PH by 25–55 ​% without affecting seedling vigor [59]. As for TaGL3-5A-G, previous studies have shown that a single nucleotide polymorphism (G/A) mutation in the exon of this gene leads to an amino acid change, affecting grain length and thousand grain weight. Varieties carrying the G allele (TaGL3-5A-G) exhibit longer grain length and higher thousand grain weight, while those with the A allele have shorter grain length and lower thousand grain weight [60]. Interestingly, this gene was identified within the loci associated with PH in our study. Given its high expression levels in the stem, this gene was hypothesized to have a dual role in regulating both PH and thousand grain weight. This dual functionality could be attributed to pleiotropic effects or the involvement in multiple growth and development pathways. The KASP markers developed for this gene was validated in the ZM578/JM22 RIL population triumphantly. Therefore, this gene was designated as a candidate gene of PH for further investigation. Both of these candidate genes were identified through 3D-PH. Compared to traditional manual measurement methods, this method is more effective in revealing phenotypic genetic analysis.

The plant height data for various environments and the detailed information of the genetic map can be downloaded from https://github.com/ILIKEWIND123/Plant-Phenomics.