Improving rice yield and quality through high-throughput phenomics, linear regression, and machine learning neural network models

doi:10.21203/rs.3.rs-4866351/v1

Download PDF

Research Article

Improving rice yield and quality through high-throughput phenomics, linear regression, and machine learning neural network models

https://doi.org/10.21203/rs.3.rs-4866351/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

To explore the potential of using high-throughput plant phenomics in rice breeding programs, one hundred elite rice varieties from southern rice-growing areas in China were subjected to high-throughput phenomic analysis. A total of 88 parameters were measured and obtained using RGB imaging, fluorescence imaging, and hyperspectral imaging at four key rice growth stages: tillering, jointing, grain filling, and 20 days after grain filling. These 88 parameters, which include RGB color and morphological features, chlorophyll fluorescence characteristics, and rice surface reflectance spectra, were analyzed to characterize high yield and high grain quality in rice using subset selection regression and deep learning neural network models. A total of 39 significant linear regression models were obtained for predicting rice yield and grain quality, with R-squared values ranging from 0.86 to 0.15, and an average R-squared of 0.41. The data from the 100 rice varieties were split into training and test sets to evaluate the prediction accuracies of the models using mean absolute error between predicted and actual values. The results indicated that the deep learning neural network model can be used to refine the linear regression model, improving the prediction accuracy. These findings suggest that high-throughput plant phenomics can be effectively utilized in rice breeding programs to select for high-yielding, high-quality rice varieties.

Rice

High-throughput phenomics

Yield

Grain quality

Linear regression

Machine learning

Rice is a staple crop that provides food for two-thirds of the world's population (Sen et al. 2020). Achieving high yield and high grain quality are longstanding goals for rice breeders. Both rice yield and grain quality are quantitative traits controlled by multiple quantitative trait loci (Zeng et al. 2017; Jin et al. 2023). The development of molecular marker technology and genetic linkage maps in rice has enabled the use of molecular markers to select for complex traits (Zeng et al. 2019a, b). Over the past 20 years, rapid advances in genomics research have led to the cloning of more than 200 genes controlling quantitative traits in rice, and the characterization of over 3,000 genes (Yao et al. 2018; Wei et al. 2021). The extensive use of molecular markers or genotyping-by-sequencing techniques allows simultaneous selection of desired genes underlying different traits in rice breeding programs (Poland et al. 2012). However, such molecular techniques are not yet widely adopted in crop breeding, as simultaneously selecting and pyramiding different genes does not necessarily yield the desired phenotype. Since decisions in plant breeding are ultimately based on the plant phenotype, a greater focus on phenotypic assessment may be beneficial.

Compared to traditional manual phenotyping methods, which are often subjective, inefficient, destructive, and error-prone (Yang et al. 2013), high-throughput plant phenomics offers a promising alternative for phenotypic selection. In the last decade, large-scale phenotyping data acquisition and processing for plant phenomics have become applicable in plant breeding, which can accelerate genetic improvements and promote the next green revolution in crop breeding (Yang et al. 2020).

Ground-based and aerial-based platforms are two types of high-throughput platforms for plant phenomics, employing various imaging sensors such as RGB, fluorescence, thermal, hyperspectral, and 3D imaging (Yang et al. 2020; Gill et al. 2022). RGB imaging can analyze morphological and color parameters (Awlia et al. 2016; Pavicic et al. 2017), while fluorescence imaging provides insights into the efficiency of photosynthesis and plant stress response by measuring chlorophyll fluorescence (Maxwell et al. 2000; Rolfe et al. 2010). Thermography can detect plant biotic and abiotic stress by monitoring leaf temperature (Pineda et al. 2021), and hyperspectral imaging can be used to predict plant traits (Heckmann et al. 2017; Silva-Perez et al. 2017; Fu et al. 2019; Vergara-Diaz et al. 2020; Grzybowski et al. 2021; Wang et al. 2021). Additionally, 3D reconstruction of plant structure can be achieved using algorithms that estimate the spatial information of plants from multiple view-angles (Klukas et al. 2014; Fuhrmann et al. 2015; Gibbs et al. 2017; Gao et al. 2021).

High-throughput phenomics research on rice has lagged behind other crops like maize and wheat. In this study, we used 100 elite rice varieties from China's southern rice-growing regions to collect 88 different traits across four growth stages, including plant shape, fluorescence intensity, and leaf hyperspectral reflectance, using RGB, fluorescence, and hyperspectral imaging. We also obtained yield and grain quality data for the 100 varieties after harvest. The goals of this study were to: (1) Assess the feasibility of using high-throughput phenomics data to predict rice yield and quality, and evaluate its potential application in rice breeding; (2) Compare the performance of subset selection regression and deep learning neural networks in predicting rice yield and quality; and (3) Explore how high-throughput phenomics data can be used to explain the underlying mechanisms of high yield and high grain quality in rice. This study will provide important insights for applying high-throughput phenomics in rice breeding programs.

Rice genotypes

One hundred rice genotypes were subjected to high-throughput phenomic analysis. Details about the 100 rice genotypes, including names, variety type, and the year they received the Variety Certificate in China, are listed in Table S1. These 100 rice genotypes were obtained by purchasing from domestic seed companies. The 100 rice genotypes were grown using the following method: On May 22, 2023, the rice genotypes were sown at the experimental farm of the China Rice Research Institute in Fuyang, Hangzhou, China. Ten days after sowing, the rice materials were transplanted into plastic buckets (18 cm high, 20 cm in diameter). Each rice genotype was planted in 3 plastic buckets (i.e., 3 replicates per genotype), with 2 seedlings planted in each bucket, totaling 6 individual plants per genotype. The potting soil used was paddy field clay loam, with about 4 kg of dry soil per bucket. These genotypes were studied using RGB imaging, visible and near-infrared (VNIR) hyperspectral imaging, short-wave infrared (SWIR) imaging, and fluorescence imaging at four rice growth stages: tillering, jointing, grain filling, and 20 days after grain filling.

Platform for high-throughput phenomic analysis

The high-throughput phenomic analysis was conducted using a PlantScreen™ System platform designed and constructed by Photon Systems Instruments (PSI, Czech R.) (https://psi.cz/). This PlantScreen™ System platform is located at the China National Rice Research Institute in Fuyang, Hangzhou, China (Fig. S1A). The PlantScreen™ System platform consisted of an RGB imaging station, a chlorophyll fluorescence imaging station, and VNIR and SWIR imaging stations (Fig. S1B, S1C). All the high-throughput rice images and data were obtained using this PlantScreen™ System and its attached computer software, including Plantscreen Data Analyzer (Version 3.3.16.3), Hyperspectral Analyzer (Version 1.0.0.14), Morpho Analysis (Version 1.0.12.0), and FluorCam10 (Version 1.0.0.1806).

RGB imaging

For 100 rice varieties, top-view and side-view RGB imaging were performed at 4 different growth stages. The RGB side view and top view cameras are mounted on robotic arms, which can take 12 megapixel resolution photos. Each camera is supplemented with a calibrated LED-based lighting source to ensure homogeneous illumination. Fish eye correction, white balance calibration, and color correction are applied to the original photos taken using the RGB cameras to reduce errors. The corrected images are then processed for masking and background subtraction to remove the background and keep only the rice plants. Color segmentation is then applied to the processed rice images. The RGB values of each pixel corresponding to the rice's surface area are extracted to serve as the dataset for the color segmentation. A k-means clustering algorithm is used to partition the pixels into 9 clusters based on their Euclidean distance in the RGB color space. These cluster centroids represent the base hues used for the color segmentation. The original pixel colors are approximated to the nearest cluster centroid, effectively color-segmenting the images. The image is segmented into 9 hues, which include RGB(34,38,22), RGB(45,54,13), RGB(45,55,36), RGB(57,71,46), RGB(59,71,20), RGB(72,84,58), RGB(73,86,36), RGB(90,98,58), and RGB(110,111,90) (Fig. 1). The detailed method for color segmentation in the PlantScreen™ System can be found in Awlia et al. (2016). After this step, a total of 18 color parameters are attained, including 9 colors for side view and top view images, respectively. The "side" and "top" suffixes are used to distinguish side view and top view, respectively, in this study. The RGB parameters attained at four rice growing stages for the 100 rice varieties were listed in Table S2 to Table S5, respectively.

Meanwhile, binary images are attained from the RGB rice images, and the binary images are subjected to morphological parameter calculation. Five parameters are attained for side view images, including AREA, PERIMETER, COMPACTNESS, WIDTH, and HEIGHT. For top view images, nine parameters are attained, including AREA, PERIMETER, ROUNDNESS, ROUNDNESS2, ISOTROPY, COMPACTNESS, ECCENTRICITY, RMS (Rotational Mass Symmetry), and SOL (Slenderness of Leaves) (Fig. 1). Area is calculated by counting pixels (PX) of the binary image. Perimeter is determined by calculating edge pixels of the rice binary image and transformed to millimeters (MM). Both pixels or millimeter can be used to indicate perimeter or area, which we use perimeter-PX or perimeter-MM to refer to pixels and millimeter, respectively. Because the perimeter-PX and perimeter-MM are the same in coefficient analysis or regression analysis, we neglect all the MM results and keep only the PX data in the present study. After obtaining the parameters of area and perimeter from the binary image, other morphometric rosette parameters can be determined, including: ROUNDNESS1, ROUNDNESS2, ISOTROPY, ECCENTRICITY, COMPACTNESS, Rotational Mass Symmetry (RMS), and Slenderness of Leaves (SOL) (PlantScreen™ analyzer, PSI) (Pavicic et al. 2017). These parameters were originally designed for studying Arabidopsis, and may not be suitable for studying rice, but since they are automatically obtained from the PlantScreen™ System, we also analyze these parameters (phenotypic traits) to see whether these rosette parameters influence rice or not. Fig.S2 is an example demonstrating the calculation of perimeter and eccentricity in rice. The introduction about these parameters can be found in Pavicic et al. (2017), and we introduce them briefly here:

$$\:\text{R}\text{O}\text{U}\text{N}\text{D}\text{N}\text{E}\text{S}\text{S}=\frac{4\ast\:\text{P}\text{I}\ast\:\text{A}\text{R}\text{E}\text{A}}{{\text{P}\text{E}\text{R}\text{I}\text{M}\text{E}\text{T}\text{E}\text{R}}^{2}}$$

$$\:\text{R}\text{O}\text{U}\text{N}\text{D}\text{N}\text{E}\text{S}\text{S}2=\frac{4\ast\:\text{P}\text{I}\ast\:\text{C}\text{O}\text{N}\text{V}\text{E}\text{X}\_\text{H}\text{U}\text{L}\text{L}\_\text{A}\text{R}\text{E}\text{A}}{{\text{C}\text{O}\text{N}\text{V}\text{E}\text{X}\_\text{H}\text{U}\text{L}\text{L}\_\text{P}\text{E}\text{R}\text{I}\text{M}\text{E}\text{T}\text{E}\text{R}}^{2}}$$

$$\:\text{I}\text{S}\text{O}\text{T}\text{R}\text{O}\text{P}\text{Y}=\frac{4\ast\:\text{P}\text{I}\ast\:\text{P}\text{O}\text{L}\text{Y}\text{G}\text{O}\text{N}\_\text{A}\text{R}\text{E}\text{A}}{{\text{P}\text{O}\text{L}\text{Y}\text{G}\text{O}\text{N}\_\text{P}\text{E}\text{R}\text{I}\text{M}\text{E}\text{T}\text{E}\text{R}}^{2}}$$

$$\:\text{R}\text{M}\text{S}=\frac{2\sqrt{{（0.5\ast\:\text{M}\text{A}\text{J}\text{O}\text{R}\_\text{A}\text{X}\text{I}\text{S}\_\text{L}\text{E}\text{N}\text{G}\text{T}\text{H}）}^{2}-{（0.5\ast\:\text{M}\text{I}\text{N}\text{O}\text{R}\_\text{A}\text{X}\text{I}\text{S}\_\text{L}\text{E}\text{N}\text{G}\text{T}\text{H}）}^{2}}}{\text{M}\text{A}\text{J}\text{O}\text{R}\_\text{A}\text{X}\text{I}\text{S}\_\text{L}\text{E}\text{N}\text{G}\text{T}\text{H}}$$

$$\:\text{C}\text{O}\text{M}\text{P}\text{A}\text{C}\text{T}\text{N}\text{E}\text{S}\text{S}=\frac{\text{A}\text{R}\text{E}\text{A}}{\text{C}\text{O}\text{N}\text{V}\text{E}\text{X}\_\text{H}\text{U}\text{L}\text{L}\_\text{A}\text{R}\text{E}\text{A}}$$

$$\:\text{S}\text{O}\text{L}=\frac{{\text{S}\text{K}\text{E}\text{L}\text{E}\text{T}\text{O}\text{N}\_\text{P}\text{E}\text{R}\text{I}\text{M}\text{E}\text{T}\text{E}\text{R}}^{2}}{\text{A}\text{R}\text{E}\text{A}}$$

$$\:\text{E}\text{C}\text{C}\text{E}\text{N}\text{T}\text{R}\text{I}\text{C}\text{I}\text{T}\text{Y}=\frac{\text{A}\text{R}\text{E}\text{A}\left(\text{C}\text{I}\text{R}\text{C}\text{L}\text{E}\:\text{O}\text{N}\text{L}\text{Y}\right)+\text{A}\text{R}\text{E}\text{A}\left(\text{C}\text{O}\text{N}\text{V}\text{E}\text{X}\:\text{H}\text{U}\text{L}\text{L}\:\text{O}\text{N}\text{L}\text{Y}\right)}{\text{A}\text{R}\text{E}\text{A}\left(\text{I}\text{N}\text{T}\text{E}\text{R}\text{S}\text{E}\text{C}\text{T}\text{I}\text{O}\text{N}\right)}$$

Fluorescence Imaging

The chlorophyll fluorescence imaging analysis can quantify the efficiency of plants in using light energy for photosynthesis and heat dissipation in rice. The fluorescence imaging station in the PlantScreen™ System uses a high-sensitivity CCD camera with a multi-color LED light panel. It utilizes pulse-modulated short-duration flashes for accurate measurement of minimal fluorescence, and it uses saturating light pulses for measurement of maximal fluorescence. The system has two types of actinic lights for light-adapted and quenching analyses.Thirteen parameters were measured and obtained for 100 rice varieties at four rice growing stages:

F0: Minimum fluorescence in the dark-adapted state.

Fm: Maximum fluorescence in the dark-adapted state.

Fv: Variable fluorescence in the dark-adapted state, calculated as Fm - F0.

QY_max: Maximum PSII quantum yield, calculated as Fv/Fm.

Fm_Lss: Steady-state maximum fluorescence in the light.

Ft_Lss: Steady-state fluorescence in the light.

F0_Lss: Steady-state minimum fluorescence in the light, calculated as F0 / ((Fv/Fm) + (F0/Fm_Lss)).

Fv_Lss: Variable fluorescence in the light-adapted state, calculated as Fm_Lss - F0_Lss.

Fq_Lss: Difference in fluorescence between Fm_Lss and Ft_Lss in the light-adapted state, calculated as Fm_Lss - Ft_Lss.

Fv/Fm_Lss: PSII quantum yield of the light-adapted sample in the steady-state, calculated as Fv_Lss/Fm_Lss.

QY_Lss: Steady-state PSII quantum yield, calculated as Fq_Lss/Fm_Lss.

qP_Lss: Coefficient of photochemical quenching in the steady-state, calculated as Fq_Lss/Fv_Lss.

Size_FC: The top view rice area under the fluorescence imaging system.

The "Lss" indicates that the parameters are measured in the light steady-state (Abdelhakim et al. 2024).

VNIR and SWIR hyperspectral imaging

The PlantScreen™ system uses two types of hyperspectral cameras: (1) Visible-Near-Infrared (VNIR) camera: Measures light in the 380–900 nm wavelength range at high spectral resolution. (2) Short-Wave Infrared (SWIR) camera: Measures light in the 900–1700 nm wavelength range at high spectral resolution. These cameras can capture hundreds of narrow spectral bands for each image pixel, providing detailed information about the optical properties of the samples being imaged. In this study, rice plants were imaged from the top-down at four different growth stages. To isolate the rice plants from the background, the VNIR images were processed using the formula: 1.2 * (2.5 * (R740 - R672) − 1.3 * (R740 - R556)), where R740, R672, and R556 represent the reflectance at 740 nm, 672 nm, and 556 nm, respectively. For the SWIR images, the background was removed using the formula (R960 - R1450) - (R960 - R1200). Pixels with values below a certain threshold were considered background, while those above the threshold were used to construct the rice plant imaging area. Several spectral vegetation indices were then calculated from the hyperspectral data using the PlantScreen™ Data Analyzer software, including PSRI (Plant Senescence Reflectance Index), NDVI (Normalized Difference Vegetation Index), PRI (Photochemical Reflectance Index), SIPI (Structure Insensitive Pigment Index), OSAVI (Optimized Soil-Adjusted Vegetation Index), and MCARI1 (Modified Chlorophyll Absorption in Reflectance Index) :

$$\:\text{P}\text{S}\text{R}\text{I}=\frac{\text{R}680-\:\text{R}500}{\text{R}750}$$

$$\:\text{N}\text{D}\text{V}\text{I}=\frac{\text{R}800-\:\text{R}670}{\text{R}800+\:\text{R}670}$$

$$\:\text{N}\text{D}\text{V}\text{I}2=\frac{\text{R}400-\:\text{R}670}{\text{R}400+\:\text{R}670}$$

$$\:\text{P}\text{R}\text{I}=\frac{\text{R}531-\:\text{R}570}{\text{R}531+\:\text{R}570}$$

$$\:\text{S}\text{I}\text{P}\text{I}=\frac{\text{R}790-\:\text{R}450}{\text{R}790+\:\text{R}650}$$

$$\:\text{O}\text{S}\text{A}\text{V}\text{I}=\frac{(1\:+\:0.16)\:\ast\:\:(\text{R}800-\:\text{R}670)}{\text{R}800-\:\text{R}670\:+\:0.16}$$

$$\:\text{M}\text{C}\text{A}\text{R}\text{I}1=1.2\:\ast\:\:[2.5\:\ast\:\:(\text{R}800-\:\text{R}670)\:-\:1.3\:\ast\:\:(\text{R}800-\:\text{R}550\left)\right]$$

For each of these indices, summary statistics were computed, including the average, standard deviation, median, minimum, and maximum values across the rice plant imaging area. The suffixes "-avg", "-stddev", "-median", "-min", and "-max" that follow the hyperspectral parameters indicate the average, standard deviation, median, minimum, and maximum values, respectively, based on a pixel-by-pixel inspection for each area defined by the rice mask. Additionally, a water index called WATER1 was calculated from the SWIR data as R1440 / R960.

Traits related to rice yield and grain quality measured at mature stage

At maturity, the aboveground portion of the rice plants (grown in plastic buckets) was harvested and separated into leaves, stem sheaths, and panicles. The plant materials were then dried at 105°C for 120 minutes and further dried at 80°C until they reached constant weight. The dry weights of the leaves, stem sheaths (shoot weight), and panicles were measured. The total dry matter weight was calculated as the sum of the leaf, stem sheath, and panicle weights. The number of effective panicles was also recorded. These five yield-related traits were measured for 100 rice varieties (Table S6), with three replicates (i.e., three plastic buckets) per variety.

After harvesting, the rice grains were dehulled using a rice huller (Otake-FC2R, Japan) to obtain the brown rice percentage, which was averaged across the three replicates. The brown rice was then milled using a precision rice mill (Puyun 2299, China) to determine the milled rice percentage. Approximately 30 grams of milled rice from each variety were analyzed using a rice grain appearance and quality tester (China Wanshen Testing Technology Co., Ltd) to measure the head rice percentage, chalky rice percentage, and degree of chalkiness. Each variety was tested in three replicates, with 30 milled rice grains per replicate. Head rice is defined as grains with a length of at least 3/4 the length of an intact milled rice grain. Chalkiness degree (%) refers to the percentage of the head rice area that is chalky. Chalky rice percentage (%) is the percentage of grains with chalkiness out of the total head rice. These five grain quality-related traits were measured for 91 rice varieties (no data available for 9 varieties, Table S7).

Regression Analysis

To identify the best subset of predictors from the 88 available parameters/traits related to RGB imaging, fluorescence imaging, VNIR and SWIR imaging, we performed subset selection regression using the "regsubsets" function in R. The goal was to use as few traits as possible to accurately predict rice yield and quality. We used the "regsubsets" function with the 'nbest = 8' argument, which returned the 8 "best" subsets of predictors. The number of predictors to include can be decided based on practical needs, ranging from 1 to 88.

Deep Learning Neural Network Model

We built a neural network model using the Keras library in R to predict rice yield and quality. First, we split the phenotypic data into 80% training and 20% testing sets using the "sample" function in R. The neural network model had two dense layers: the first had 10 units and used the ReLU activation function, and the second was the single-unit output layer. This was a simple feed-forward neural network. We trained the model for 500–2500 epochs, with a batch size of 32 and a validation split of 0.2 (see Data. S1 for details). We evaluated the model's performance on the test data using the "predict()" function and calculated the mean absolute error (MAE). Finally, we compared the neural network model's performance to a linear regression model. The data can be split into training and testing sets in different proportions depending on your needs, but here we used an 80/20 split.

Elite rice varieties grown in southern China

Of the 100 rice varieties evaluated, 98 have been awarded the Provincial or National Variety Certificate in China, while the remaining two varieties are still undergoing field trials and await certification. It is important to note that obtaining the Provincial or National Variety Certificate is a rigorous process, requiring a cultivar to pass a two-year regional trial and a subsequent one-year production trial at the provincial or national level. According to Chinese law, a cultivar cannot be planted in production without this Variety Certificate. For instance, rice varieties with poor blast disease resistance or low yield may fail to pass the variety testing and, therefore, cannot be awarded the Variety Certificate. In contrast, the 100 rice genotypes included in this study contain some exemplary "star" varieties, such as TianYouHuaZhan, a renowned hybrid recognized by the Ministry of Agriculture as a "super rice" and used as a control in national trials. Similarly, FengLiangYou4 and WuYou308 are also highly regarded and used as control varieties in national trials. The locations of the 75 varieties that have been awarded the Provincial Variety Certificate are shown in Fig. 2A, while the remaining 23 varieties have received the National Variety Certificate and are not shown in the map. Analysis of the trial timelines revealed that 56 (66.3%) of the 98 certified varieties were awarded their Variety Certificates in 2014 or later (Fig. 2B), suggesting that most of the varieties used in this study represent up-to-date cultivars that are widely grown in southern China's rice production areas.

Ten traits related to yield and grain quality were measured at the mature stage for the 100 rice varieties (Table S6, Table S7). The Shapiro-Wilk normality test showed that the frequency distributions of shoot weight, leaf weight, panicle weight, dry matter weight, and milled rice percentage were normal, while the distributions of number of panicles, brown rice percentage, head rice percentage, chalky rice percentage, and chalkiness degree were non-normal (Fig. 2C).

RGB image acquisition and analysis

Both top view and side view images of the rice plants were captured using an RGB imaging station. Background subtraction was first applied to the images to isolate the rice plants from the background. Color segmentation was then used to divide the rice plant pixels into nine distinct hues (Fig. 1). The color segmentation was performed on images of 100 rice varieties taken at four growth stages: tillering, jointing, grain filling, and 20 days after grain filling (Fig. 3A). The RGB coordinates of the nine green hues are summarized in Fig. 3B. By analyzing the pixels of these nine hues at different growth stages, we were able to observe the trends of color changes over the course of rice development (Fig. 3C) and the percentage of pixels in each hue (Fig. 3D). The frequency distributions of the nine hues for the 100 rice varieties are shown in Fig. 4, Fig.S3, Fig.S4, and Fig.S5 for the top view and side view images at the four growth stages.

The top view and side view images were also converted to binary format to extract morphological parameters. Five parameters were measured from the side view images, while nine parameters were obtained from the top view images (Fig. 1). The area and perimeter were directly measured from the original images (Fig. S2). These, along with the convex hull measurements, were then used to automatically calculate additional morphometric parameters, including roundness, roundness2, isotropy, eccentricity, compactness, RMS, and SOL, as defined by the PlantScreen™ analyzer (Pavicic et al. 2017). The frequency distributions of these morphological parameters for the 100 rice varieties at the different growth stages are shown in Fig. 4, Fig. S3, Fig.S4, and Fig.S5.

While these morphometric parameters were originally developed for rosette plants like Arabidopsis, we found them to be useful for studying rice as well. For example, we observed a strong negative correlation between panicle weight and “eccentricity” measured at the jointing (-0.58) and grain filling (-0.46) stages (Fig. 6). We split the data into a training set (70%) and a test set (30%) to evaluate linear regression models using eccentricity as a single predictor of panicle weight. The mean absolute error (MAE) between the predicted and measured panicle weight was 4.12 g when using eccentricity from the jointing stage, and 6.60 g when using eccentricity from the grain filling stage (see Data.S1 for the R script). Incorporating four additional variables ("AREA_PX_side", "HEIGHT_PX_side", "PSRI.stddev", "RGB.72.84.58.side") improved the prediction, resulting in an MAE of 3.52 g between the predicted and measured panicle weight (see below). Given that the average panicle weight for the 100 varieties is 31.2 g, an MAE of 3.52 g represents a substantial improvement in predictive accuracy that could be useful for rice breeding applications.

Parameters of Fluorescence Imaging

Twelve parameters are derived from chlorophyll fluorescence measurements, providing insights into the function and regulation of Photosystem II (PSII) in rice photosynthesis. For example, F₀ is the minimum fluorescence in the dark-adapted state, representing the fluorescence emitted when all PSII reaction centers are open. Fₘ is the maximum fluorescence in the dark-adapted state, representing the fluorescence emitted when all PSII reaction centers are closed. Fᵥ (Fᵥ = Fₘ - F₀) represents the variable part of the fluorescence associated with the opening and closing of PSII reaction centers. QY_max (QY_max = Fᵥ/Fₘ) is the maximum PSII quantum yield, representing the maximum efficiency of PSII in converting absorbed light energy into photochemical energy. Fᵥ/Fₘ_Lss refers to the PSII quantum yield of the light-adapted sample in the steady-state, calculated as Fᵥ_Lss/Fₘ_Lss (Fig. 5).

We calculated the correlation coefficients between the fluorescence parameters measured at four rice growth stages and yield-related traits (Fig. 6). The correlation coefficient between QY_max and panicle weight is 0.31 at the grain filling stage, while it was − 0.43 at 20 days after grain filling. This indicates that a higher QY_max value at the grain filling stage, and a lower QY_max value at 20 days after grain filling, are desirable to achieve a higher panicle weight. The correlation coefficients between fluorescence parameters and grain quality-related traits are listed in Fig. 7.

Parameters of SWIR and VNIR Imaging

The hyperspectral cameras in the PlantScreen™ System can measure hundreds of spectral bands between 350 and 1700 nm at a nm-level resolution for each image pixel. Hyperspectral imaging provides the visualization of differences in plant surface reflectance spectra by detecting the reflected radiation. The leaf water content can be determined by analyzing reflectance values at 1440 nm and 960 nm (Fig. 8).

It was found that the correlation coefficient between panicle weight and PSRI.stddev (standard deviation of PSRI) measured at the tillering stage was − 0.55, and the correlation coefficient increased to -0.61 when the PSRI.stddev value was measured at the jointing stage (Fig. 6). This result is very helpful for predicting panicle weight in rice breeding by using the PSRI.stddev value measured at the jointing stage. Rice breeders do not need to wait until the rice plants are mature to measure the panicle weight to make a breeding decision; they can simply predict panicle weight at the tillering or jointing stage by measuring the reflectance values at 680 nm, 500 nm, and 750 nm, because the PSRI value equals "(R680 - R500) / R750". It is also possible to use the NDVI.stddev value to predict panicle weight, as the correlation coefficients between NDVI.stddev and panicle weight are − 0.53 and − 0.61 when the NDVI values were measured at the tillering and jointing stages, respectively (Fig. 6). It is also possible to predict the head rice percentage using the MCARI1.max value measured at the jointing stage, as the correlation coefficient between them is -0.53 (Fig. 7).

Linear regression and deep learning neural network model for prediction of rice yield and grain quality at four rice growing stages

Subset selection regression was utilized to build linear regression models using eight independent parameters to predict rice yield and grain quality at four rice growing stages. A total of 40 regression models were built at four stages for predicting 10 traits relating to yield and grain quality (Fig. 9, Fig.S6, Fig.S7, Fig.S8, Table S8).

We chose eight independent parameters in this study because we wanted to use as few parameters as possible to predict rice yield and grain quality, and eight is about one-tenth of all the 88 parameters obtained from RGB imaging, fluorescence imaging, and hyperspectral imaging. It is possible to use one to 88 parameters to build the linear regression model according to the practical needs in rice breeding activities or research purposes.

In the above analysis, we had used one predictor, "eccentricity," to predict the panicle weight at the jointing stage, and the mean absolute error (MAE) between the predicted panicle weight and the measured panicle weight is 4.12 g. The subset selection regression yielded four predictors ("AREA_PX_side," "HEIGHT_PX_side," "PSRI.stddev," "RGB.72.84.58.side") to predict the panicle weight at the same stage, and the MAE is 3.52 g (See Table S8 for the linear regression model yielded by subset selection regression). The MAE is 4.54 if one predictor, "AREA_PX_side," was used to predict the panicle weight. The parameter "AREA_PX_side" has the largest correlation coefficient with panicle weight (Fig. 6). This indicates that the linear regression model developed by subset selection regression has higher prediction accuracy than choosing a predictor based on the largest correlation coefficient between the independent variable and the target variable.

We would like to predict rice yield and grain quality at as early stages as possible, such as the tillering or jointing stages, so that unwanted rice individuals can be eliminated to save labor, cost, and reduce the planting area in the rice field for breeding purposes. If predictions were made at late stages like the grain filling stage or 20 days after grain filling, rice plants would be mature and ready for harvesting, and the unwanted rice individuals kept until the grain filling stage or later would cause a large amount of labor, cost, and land usage.

Therefore, we used a deep learning neural network model for prediction of rice yield and grain quality at the tillering stage to compare the prediction accuracy between the linear regression and deep learning models (Fig. 10). The R script for the comparison between linear regression and machine learning models using MAE (mean absolute error) is provided as Data.S2.

The strategy we used is: First, we calculated the correlation coefficients between the 88 independent variables and the traits related to yield (or grain quality) to determine the eight parameters with the largest absolute value of the correlation coefficient. These eight parameters were used to build a linear regression model by subset selection regression. Then, the independent variables determined by subset selection regression were used in both the linear regression model and the neural network model to compare the MAE between them.

We used 500, 1000, 1500, 2000, and 2500 epochs to build the sequential neural network model created using the Keras library. The results show that there is an optimum number of epochs for training the neural network model for predicting panicle weight, panicle number, and chalky rice percentage at around 1500 epochs (Fig. 10), and the MAE value will increase gradually when the epochs are more than 1500, which indicates an overfitting of the model.

The results listed in Fig. 10 show that the predicting accuracy of the deep learning neural network model can reach a smaller MAE (Fig. 10E) or a MAE close to the linear regression model (Fig. 10D) by using the appropriate training epochs. This suggests that we can use the deep learning neural network model to refine the linear regression model to make the prediction more accurate.

High-throughput phenomics in rice

As a monocot model plant, high-throughput genomics, transcriptomics, proteomics, and metabolomics have boosted rice research in the past two decades, and yielded many novel insights for other monocot plants. However, the high-throughput phenomics research in rice has largely fallen behind its genomics, transcriptomics, or proteomics research. High-throughput phenotyping has become a bottleneck in accelerating rice biology research. The phenotype is the ultimate measure for evaluating the effects of a gene, protein, or metabolic process. Precise measurement of phenotypes will greatly support plant biology research. But compared to genomics, transcriptomics, proteomics, and metabolomics in rice, our knowledge about rice phenomics is the least. For example, we did not know whether the morphological trait of "eccentricity" influenced rice grain quality or not. And now we know that "eccentricity" indeed influences rice chalkiness degree, which can be predicted using parameters measured at the jointing stage: chalkiness_degree = -36.5 + 26.7*ECCENTRICITY_top + 106.9*SIPI.stddev + 0.005*HEIGHT_PX_side − 31.7*PSRI.min (R² = 0.33, p < 0.000001). Although we did not find reports in the previous literature on using large amounts of parameters obtained by RGB, fluorescence, and hyperspectral cameras at different rice growing stages to predict rice yield and grain quality for breeding purposes, there have been some reports on the limited use of phenotype evaluation using automatic machines or image analysis pipelines, such as for identification of broken rice kernels (Lin et al. 2012), identifying filled and unfilled rice spikelets (Duan et al. 2011), detecting plant-hopper infestation area on rice stems (Zhou et al. 2011), and detection of rice blast disease using near-infrared hyperspectral imaging (Yang et al. 2012) and image analysis pipelines (Hartmann et al. 2011). We may see many more reports published in the coming years as rice phenomics research is developing rapidly.

Can regression models with low R-squared values still be useful for rice breeding?

We developed a total of 39 significant linear regression model formulas (one formula is not significant) for the prediction of rice yield and quality. The R² of these 39 linear regression models varied from 0.86 to 0.15, with an average R² value of 0.41. Some of the formulas have an R² value less than 0.5 (Table S8). Can these linear regression models with low R² values (less than 0.5) be used in rice breeding? Do they have practical value? To answer this question, we split the data of 100 rice varieties into a training set (70%) and a testing set (30%), and used the mean absolute error (MAE) between predicted value and real value to evaluate those linear regression models with low R² values. We would like to predict the rice yield and grain quality. Since grain filling and 20 days after grain filling are late stages in rice, we did not test the MAE of the models at these two stages. We assessed the applicability of the linear regression models by comparing the mean absolute error (MAE) and average values of the target traits across the 100 rice varieties (Table S9). We believe that the linear regression models for the prediction of panicle weight, number of panicles, brown rice percentage, and milled rice percentage at both the tillering and jointing stages are applicable in breeding.

Let's examine the MAE, the average values of the target traits across the 100 rice varieties, and the R² value of the models carefully. At the tillering stage, the MAE, average values of panicle weight across the 100 rice varieties, and R² value for the prediction model were 4.96 g, 31.16 g, and 0.43, respectively. It indicated that the prediction accuracy for this model is about 31.16 ± 4.96 g (R² = 0.43). The prediction accuracy for the prediction of number of panicles, brown rice percentage, and milled rice percentage at the tillering stage were about 10.03 ± 1.41 (R² = 0.34), 74.66 ± 4.85 (R² = 0.39), and 64.73 ± 2.37 (R² = 0.30), respectively. The prediction accuracy for the prediction of panicle weight, number of panicles, brown rice percentage, and milled rice percentage at the jointing stage were about 31.16 ± 3.56 (R² = 0.57), 10.03 ± 0.97 (R² = 0.38), 74.66 ± 1.58 (R² = 0.30), and 64.73 ± 2.25 (R² = 0.26), respectively. We believe these models are applicable in rice breeding because most of their MAE values are less than 10% of the average value of the predicted target traits. This result suggested that a higher R² in the prediction model does not always yield a low MAE value. We need to carefully examine the prediction model to check whether it can be used in production or not.

Characterizing high yield and high grain quality in rice by phenomics

High yield is a basic standard a rice variety must meet to gain a Variety Certificate before it can be utilized in production in China. Low-yielding rice cannot be awarded a Variety Certificate in China. High-quality rice can sell at a higher price than low-quality rice and therefore has a higher value in the food market. Therefore, high yield and high grain quality are two of the most important rice breeding targets. Genetic studies have proven that rice yield and grain quality are quantitative traits controlled by many quantitative trait loci with small effects at each locus. The regression model built by high-throughput phenomics in the present study indicated that it must use four to six traits to predict yield or quality. For example, four traits were used to predict panicle weight at the tillering stage, including "HEIGHT_PX_side", "Plant_Height", "RGB.57.71.46.side", and "PRI.stddev". This suggests that rice height, leaf color (RGB.57.71.46), and reflectance value at 531 nm and 570 nm (PRI = (R531- R570) / (R531 + R570)) can influence panicle weight in rice. It is easy to understand that taller rice is usually stronger than shorter rice and therefore has higher panicle weight. The reason why the leaf color parameter "RGB.57.71.46" influences panicle weight may be explained as follows: Chlorophyll is the green pigment in plant leaves that is essential for photosynthesis, and the "RGB.57.71.46" parameter is related to chlorophyll content in rice. The amount of chlorophyll in rice leaves is closely linked to its ability to capture light energy and convert it into chemical energy (in the form of carbohydrates) through photosynthesis. Therefore, "RGB.57.71.46" is related to panicle weight. The reason why the reflectance value at 531nm and 570 nm influences panicle weight requires further study.

According to the regression models listed in Table S9, the brown rice percentage is positively correlated with MCARI1.max, but negatively correlated with OSAVI.max and Size.avg_VNIR at the tillering stage. In contrast, the brown rice percentage is positively correlated with NDVI2.median and NDVI.stddev, while it is negatively correlated with NDVI2.avg, NDVI.min, WATER1.avg, and Fo_Lss. These findings suggest that the parameters determining brown rice percentage may change greatly across different growth stages, indicating a dynamic process in the determination of grain quality. Using phenomics techniques to characterize rice with both high yield and high grain quality can yield models to guide breeding efforts and may also provide novel insights from a phenomics perspective.

Author contributions J.X. and G. F. designed the experiment, performed the experiments, analyzed the data, and wrote the manuscript. H. S., Y.W., W.F., B.F., W.W., T.C., Y.X., performed the experiments, analyzed the data. Y.Z. analyzed the data and wrote the manuscript. J.X. and G. F. supervised the project and designed the research.

Funding This work was supported bythe National Key Research and Development Program of China (2022YFD2300700), Zhejiang Provincial Natural Science Foundation key project of China (Z23C130001), National Natural Science Foundation of China (Grant No. 3230151416), and the “Sannongliufang” Science and Technology Cooperation Project of Zhejiang province (2023SNJF003).

Data availability All data supporting the findings of this study are available within the paper and its Supplementary Information.

Conflict of interest All authors declare that they have no conflict of interest.

Abdelhakim LOA, Pleskačová B, Rodriguez-Granados NY, Sasidharan R, Perez-Borroto LS, Sonnewald S, Gruden K, Vothknecht UC, Teige M, Panzarová K (2024) High Throughput Image-Based Phenotyping for Determining Morphological and Physiological Responses to Single and Combined Stresses in Potato. JoVE J. 10.3791/66255
Awlia M, Nigro A, Fajkus J, Schmoeckel SM, Negrão S, Santelia D et al (2016) High-Throughput Non-destructive Phenotyping of Traits that Contribute to Salinity Tolerance in Arabidopsis thaliana. Front Plant Sci 7:1414
Duan L, Yang W, Bi K, Chen S, Luo Q, Liu Q (2011) Fast discrimination and counting of filled/unfilled rice spikelets based on bi-modal imaging. Comput Electron Agric 75:196–203
Fu P, Meacham-Hensold K, Guan K, Bernacchi CJ (2019) Hyperspectral leaf reflectance as proxy for photosynthetic capacities: an ensemble approach based on multiple machine learning algorithms. Front Plant Sci 10:730
Fuhrmann S, Langguth F, Moehrle N, Waechter M, Goesele M (2015) MVE-An image-based reconstruction environment. Comput Graph 53:44–53
Gao T, Zhu F, Paul P, Sandhu J, Doku HA, Sun J, Pan Y, Staswick P, Walia H, Yu H (2021) Novel 3D Imaging Systems for High-Throughput Phenotyping of Plants. Remote Sens 13:2113
Gibbs JA, Pound M, French AP, Wells DM, Murchie E, Pridmore T (2017) Approaches to three-dimensional reconstruction of plant shoot topology and geometry. Funct Plant Biol 44:62–75
Gill T, Gill SK, Saini DK, Chopra Y, Koff JP, Sandhu KS (2022) A Comprehensive Review of High Throughput Phenotyping and Machine Learning for Plant Stress Phenotyping. Phenomics 2:156–183
Grzybowski M, Wijewardane NK, Atefi A, Ge Y, Schnable JC (2021) Hyperspectral reflectance-based phenotyping for quantitative genetics in crops: Progress and challenges. Plant Comm 2:100209
Hartmann A, Czauderna T, Hoffmann R, Stein N, Schreiber F (2011) HTPheno: an image analysis pipeline for high-throughput plant phenotyping. BMC Bioinformatics 12:148
Heckmann D, Schlüter U, Weber APM (2017) Machine Learning Techniques for Predicting Crop Photosynthetic Capacity from Leaf Reflectance Spectra. Mol Plant 10:878–890
Jin S-K, Xu L-N, Yang Q-Q, Zhang M-Q, Wang S-L, Wang R-A et al (2023) High-resolution quantitative trait locus mapping for rice grain quality traits using genotyping by sequencing. Front Plant Sci 13:1050882
Klukas C, Chen D, Pape JM (2014) Integrated analysis platform: An open-source information system for high-throughput plant phenotyping. Plant Physiol 165:506–518
Lin P, Chen Y (2012) Identification of Broken Rice Kernels Using Image Analysis Techniques Combined with Velocity Representation Method. Food Bioprocess Technol 5:796–802
Maxwell K, Johnson GN (2000) Chlorophyll fluorescence—a practical guide. J Exp Bot 51:659–668
Pavicic M, Mouhu K, Wang F, Bilicka M, Chovanček E, Himanen K (2017) Genomic and Phenomic Screens for Flower Related RING Type Ubiquitin E3 Ligases in Arabidopsis. Front Plant Sci 8:416
Pineda M, Barón M, Pérez-Bueno ML (2021) Thermal Imaging for Plant Stress Detection and Phenotyping. Remote Sens 13:68
Poland JA, Rife TW (2012) Genotyping-by-sequencing for plant breeding and genetics. Plant Genome 5:92–102
Rolfe SA, Scholes JD (2010) Chlorophyll fluorescence imaging of plant–pathogen interactions. Protoplasma 247:163–175
Sen S, Chakraborty R, Kalita P (2020) Rice - not just a staple food: A comprehensive review on its phytochemicals and therapeutic potential. Trends Food Sci Technol 97:265–285
Silva-Perez V, Molero G, Serbin SP, Condon AG, Reynolds MP, Furbank RT, Evans JR (2017) Hyperspectral reflectance as a tool to measure biochemical and physiological traits in wheat. J Exp Bot 69:483–496
Vergara-Diaz O, Vatter T, Kefauver SC, Obata T, Fernie AR, Araus JL (2020) Assessing durum wheat ear and leaf metabolomes in the field through hyperspectral data. Plant J 102:615–630
Wang S, Guan K, Wang Z, Ainsworth EA, Zheng T, Townsend PA, Li K, Moller C, Wu G, Jiang C (2021) Unique contributions of chlorophyll and nitrogen to predict crop photosynthetic capacity from leaf spectroscopy. J Exp Bot 72:341–354
Wei X, Qiu J, Yong K, Chan C, Hu X, Zeng D, Guo L, Fitzgerald M, Bonnett D, Wang J et al (2021) A quantitative genomics map of rice provides genetic insights and guides breeding. Nat Genet 53:243–253
Yang Y, Chai RY, He Y (2012) Early detection of rice blast (Pyricularia) at seedling stage in Nipponbare rice variety using near-infrared hyper-spectral image. Afr J Biotechnol 11:6809–6817
Yang W, Duan L, Chen G, Xiong L, Liu Q (2013) Plant phenomics and high-throughput phenotyping: accelerating rice functional genomics using multidisciplinary technologies. Curr Opin Plant Biol 16:180–187
Yang W, Feng H, Zhang X, Zhang J, Doonan JH, Batchelor WD et al (2020) Crop Phenomics and High-Throughput Phenotyping: Past Decades, Current Challenges, and Future Perspectives. Mol Plant 13:187–214
Yao W, Li G, Yu Y, Ouyang Y (2018) funRiceGenes dataset for comprehensive understanding and application of rice functional genes. Gigascience 7:1–9
Zeng Y, Shi J, Ji Z, Wen Z, Liang Y, Yang C (2017) Combination of twelve alleles at six quantitative trait loci determines grain weight in rice. PLoS ONE 12:e0181588
Zeng Y, Chen Y, Ji Z, Liang Y, Yang C (2019a) Twenty–four alleles at twelve quantitative trait loci act additively to control tiller angle in cultivated rice. Plant Growth Regul 88:195–203
Zeng Y, Chen Y, Ji Z, Liang Y, Zheng A, Wen Z, Yang C (2019b) Control of plant height by 24 alleles at 12 quantitative trait loci in rice. Crop Breed Appl Biotechnol 19:200–207
Zhou ZY, Zang Y, Luo XW, Wang P (2011) Color-based corner detection algorithm for rice plant-hopper infestation area on rice stem using the RGB color space. Paper presented at: ASABE Annual International Meeting; 2011; Louisville, Kentucky, USA. Paper number 1111374

Download PDF

Version 1

posted

You are reading this latest preprint version

Improving rice yield and quality through high-throughput phenomics, linear regression, and machine learning neural network models

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Rice genotypes

Platform for high-throughput phenomic analysis

RGB imaging

Fluorescence Imaging

VNIR and SWIR hyperspectral imaging

Traits related to rice yield and grain quality measured at mature stage

Regression Analysis

Deep Learning Neural Network Model

Results

Elite rice varieties grown in southern China

RGB image acquisition and analysis

Parameters of Fluorescence Imaging

Parameters of SWIR and VNIR Imaging

Discussion

High-throughput phenomics in rice

Can regression models with low R-squared values still be useful for rice breeding?

Characterizing high yield and high grain quality in rice by phenomics

Declarations

References

Supplementary Files

Status:

Version 1