Early screening of salt-tolerant and high-yielding rice varieties via cost-effective UAV data

Abstract

Breeding rice varieties that are both salt-tolerant and high-yielding is essential for utilizing saline-alkaline lands and ensuring food security. However, However, high-throughput and accurate phenotyping at early growth stages remains a major bottleneck in breeding programs. In this study, unmanned aerial vehicle (UAV) imaging was employed to screen salt-tolerant and high-yielding varieties among 60 rice varieties under saline-alkaline field conditions. Red-green-blue (RGB), multispectral, and thermal canopy images were acquired throughout the growing season by UAV, from which 41 phenotypic traits were extracted at each growth stage. These traits were categorized into early-stage (tillering and jointing), late-stage (booting, flowering, and maturity), and whole-growth-stage (from tillering to maturity) and subsequently used to screen salt-tolerant and high-yielding rice varieties. Results showed that: (1) An early high-throughput screening method for salt-tolerant rice varieties was developed based on the membership function and UAV phenotypes (MFuav), achieving high performance (Precision >0.8, OA > 0.7). MFuav demonstrated the highest accuracy at the early-stage, with Precision increasing by 0.29 and 0.43 compared to the late- and whole-stage models, respectively. (2) A machine learning based UAV phenotypes framework (MLuav) was developed to further improve salt-tolerance screening performance. Within this framework, the partial least squares regression (PLSR) was employed for early-stage salt-tolerance screening, which achieved a Precision of 0.97 and an OA of 0.78, outperforming the MFuav by 0.11 and 0.08, respectively. In addition, within the same MLuav framework, early-stage UAV phenotypes were further used for actual yield prediction using a Random Forest (RF) model. The model achieved a high Recall for high-yielding varieties (Recall = 1.00), ensuring that no potentially high-yielding germplasm was missed, although this was accompanied by a moderate Precision (0.51) and an overall accuracy of 0.70. (3) The MLuav consistently outperformed the MFuav in screening salt-tolerant and high-yielding varieties across all 60 rice varieties. Among the five referenced salt-tolerant and high-yielding rice varieties, the MLuav correctly screened four using early-stage phenotypes, whereas the MFuav only screened three. Overall, the proposed method enables early screening of salt-tolerant and high-yielding rice varieties, offering an efficient tool for the screening and utilization of elite stress-resilient germplasm.

Highlights

• •UAV phenotyping enabled efficient screening of salt-tolerant and high-yielding rice varieties.
• •UAV phenotypes combined with machine learning models realized early screening of salt-tolerant rice varieties.
• •Machine learning models outperformed membership function method for early screening of salt-tolerant and high-yielding varieties.

1. Introduction

Rice is one of the world's major staple crops, yet its production is increasingly threatened by various abiotic stresses. Soil salinity is among the most critical abiotic stressors limiting rice yield potential [1,2]. Currently, approximately 1.0 × 10⁹ ha (around 7%) of land worldwide is classified as saline-alkaline soils [3]. By 2050, the proportion of salinized arable land is projected to rise from 20% to 50% globally [4]. The screening of salt-tolerant rice varieties capable of efficiently utilizing saline-alkaline land is of vital importance for expanding arable land and enhancing global rice production.

The salt-tolerant rice varieties can be evaluated by analyzing their physiological, biochemical, or structural responses under salt stress. Traditional approaches for screening salt-tolerant rice varieties rely on manual measurements of various physiological and structural traits across developmental stages [5]. For instance, seed salt tolerance at the germination stage can be evaluated via germination rate, mean germination time, root length, seedling mortality, and other structural traits [[6], [7], [8]]. At the seedling stage, salt tolerance can be distinguished by measuring physiological and biochemical traits such as leaf chlorophyll content, potassium and sodium ion concentrations, fresh and dry biomass, and net photosynthetic rate [7,[9], [10], [11], [12]]. However, these methods depend on manual phenotyping, which is labor-intensive and time-consuming for screening a large number of salt-tolerant varieties. Therefore, there is an urgent need for high-throughput, low-cost, and efficient technologies to address the challenges of screening salt-tolerant rice varieties.

The rapid advancement of high-throughput phenotyping has provided new opportunities to overcome the aforementioned challenges. Depending on the research environment, high-throughput phenotyping can be conducted in controlled indoor conditions or open-field settings. For instance, low-cost and fixed-point red-green-blue (RGB) indoor imaging systems allow for the precise screening of salt-tolerant rice varieties by analyzing morphological characters such as shoot and root fresh weights and root length [13]. The response of rice varieties to different levels of salt stress was quantified based on total shoot area and senescent shoot area, calculated from indoor low-cost and high-throughput RGB images [14]. Computationally economic multi-angle RGB images collected via an indoor high-throughput phenotyping platform providing insights into the early responses of rice varieties to salinity, in particular into the effects of salinity on plant growth and transpiration [15]. Although indoor phenotyping offers advantages such as controlled conditions, high reproducibility, and shorter experimental cycles [16], it cannot fully substitute for field trials under natural field conditions. In field experiments, rice varieties are grown in saline-alkaline soils. The expensive hyperspectral sensors are used to collect canopy reflectance data at key growth stages, enabling the screening of hyperspectral indices sensitive to salt-induced nutrient changes [17]. However, the large number of genotypes, broad spatial scale, and long experiment duration make the screening of salt tolerance varieties based on field condition highly challenging, emphasizing the need for efficient and non-destructive field high-throughput approaches.

The effective and extensive ability of the unmanned aerial vehicle (UAV) platform equipped with multi-sensors has overcome the difficulties in studying crop stress under complex field conditions [18]. By quantifying canopy and projected area during the whole growth stage, UAV-based RGB and multispectral imagery enabled effective and non-destructive screening of salt-tolerant and high-yielding tomato accessions from 600 field-grown lines under salt stress [19]. An efficient high-throughput phenotyping method was developed with UAV images to quantitatively analyze drought response dynamics in 240 field-grown rice varieties [20]. The UAV-based imagery approach was similarly employed to screen drought-tolerant soybean genotypes among 116 field-grown accessions [21]. In addition, Thermal infrared imagery from the UAV captured during maize growth stages enabled accurate calculation of water stress indices for effective drought monitoring [22]. However, current UAV-based phenotyping has rarely addressed salt stress in the field, and lacks a standardized method for early screening of salt-tolerant genotypes, highlighting the need for phenotype-based early-stage field screening methods of salt-tolerant rice varieties.

Phenotype-based approaches for screening crop stress responses in the field can be broadly categorized into change detection [14,23], index construction [20], and machine learning (ML) based methods [24,25]. Change detection methods evaluate crop stress tolerance by analyzing differences in plant status between stress and control groups or before and after stress exposure [19,21]. Index construction methods first extract phenotypic traits under stress conditions, such as image-based or spectral features. The phenotypic traits are then put into mathematical operations or functional modeling to derive stress evaluation indices, which are compared with control group values to screen tolerant varieties [26]. The mathematical operations include the membership function (MF) method [27] and the ratio method [28]. The MF integrates phenotypic traits under stress conditions using the principal component analysis (PCA) to assign weights, while the ratio method calculates evaluation indices by taking ratios of stress-related traits and analyzing their variation. The ML methods predict phenotypic traits or evaluation indices, and are often integrated with change detection or index construction strategies to characterize varietal responses to stress [22]. All these methods require control groups, which providing comparison to distinguish the tolerance of varieties. However, due to the large and contiguous distribution of saline-alkaline soils, it is often difficult to establish salt-stressed and non-stressed control plots under the same climatic conditions, making traditional control-based index construction methods challenging to implement. Consequently, there remains a lack of high-throughput and quantitative methods suitable for screening salt-tolerant varieties under complex field conditions.

Therefore, this study aims to develop and validate an effective screening method using low-cost UAV-based phenotyping technology. The main objectives are: (1) to prove that the high-throughput UAV phenotypes can be used for variety screening by constructing a membership evaluation method based on UAV phenotypes; (2) to achieve early-stage screening of salt-tolerant rice varieties using UAV phenotypes combined with ML models, while also enabling early-stage screening of high-yielding varieties; (3) to validate the newly developed method for directly screening salt-tolerant and high-yielding rice varieties under field conditions.

2. Materials and methods

The workflow of this study mainly includes five steps (Fig. 1). (1) Data acquisition: Both manually measured data and UAV monitoring data were collected. The manually measured data included nutrient concentrations (N⁺, P⁺, K⁺, and Na⁺), leaf chlorophyll content (SPAD), plant height, biomass, and yield. UAV data were acquired during key growth stages of rice canopy, including RGB, multispectral (MSI), and thermal (TIR) imagery (Fig. 1a). (2) UAV data preprocessing: The raw imagery underwent ortho-mosaic images generation, geometric correction, radiometric correction, temperature calibration, plot segmentation, and canopy extraction (Fig. 1b). (3) High-throughput phenotype extraction: Including spectral, textural, and thermal phenotypic traits. Spectral and textural phenotypic traits were derived using spectral indices and gray-level cooccurrence matrix (GLCM) methods. Temperature traits were directly based on the average value of the TIR image values. Based on rice growth characteristics, all UAV-derived phenotypes were categorized into early-, late-, and whole-stage groups (Fig. 1c). The early-stage phenotypes corresponded to the tillering and jointing stages, which are critical for early vegetative growth and stress response. Seedling-stage data were excluded because incomplete canopy coverage and strong duckweed background effects may reduce the reliability of UAV-derived traits at this stage. The late-stage phenotypes included the booting, flowering, and maturity stages, which are closely associated with yield formation, while the whole-stage phenotypes integrated information across all stages to characterize cumulative growth performance. Multi-temporal observations within each stage were summarized into representative phenotypic values for subsequent screening and modeling analyses. (4) Ground-truth reference: Salt-tolerant rice varieties were screened using the MF method based on manually collected data. High-yielding rice varieties were determined by measured yield. The intersection of these two sets was the referenced group of salt-tolerant and high-yielding rice varieties (Fig. 1d). (5) Screening of salt-tolerant and high-yielding rice varieties using low-cost UAV data: Relevant features were selected from spectral, textural, and temperature traits. These features were then used in both the MF and the ML for screening candidate varieties. The screening accuracy across different growth stages was validated against ground-truth reference varieties (Fig. 1e)

Fig. 1.

Workflow of this study. (a) Data Collection. Manually measured data (N⁺, P⁺, K⁺, Na⁺, SPAD, plant height, tiller number, biomass, and yield) and UAV data (RGB, MSI, and TIR) were collected at the tillering, jointing, booting, flowering, and maturity stages, respectively. (b) UAV data preprocessing, including ortho-mosaic images generation, geometric correction, radiation geometric, temperature correction, plot segmentation, and canopy extraction. (c) High-throughput phenotype extraction. Three types of phenotypes were extracted, including spectral indices, textural indices, and temperature index. Meanwhile, these phenotypes were divided into three stages: early, late, and whole. (d) Ground truth. The referenced salt-tolerant and high-yielding rice varieties were obtained from manually measured data. (e) Variety screening. The screening methods of salt-tolerant and high-yielding rice varieties were constructed based on high-throughput phenotypes from UAV data at different growth stages.

2.1. Study area and data acquisition

The study was conducted at Dongxin Farm, Lianyungang City, Jiangsu Province, China (34° 33′ N, 119° 21′ E). The site is characterized by mildly saline-alkaline soil with a salt content of 1 g/kg and a pH of 8.41 [29]. A total of 60 widely cultivated rice varieties from the Huang-Huai region, representing different levels of salt tolerance, were selected for the experiment (Table A1). The experiment was conducted from June to October 2023. The experiment area was 6271.20 m², including 61 plots with an average plot size of 102.81 m². Particularly, one rice variety was planted in two different plots, and the other 59 varieties were planted in different plots. The rice seedlings were planted with a row spacing of 30 cm between rows and 12 cm within rows (Fig. 1a). Guard rows were established around the experimental area, and field management practices were consistent with those used in conventional large-scale rice production.

UAV data were acquired using a Matrice 300 RTK drone (DJI, Shenzhen, China) equipped with two types of sensors. One was a high-resolution RGB camera of 8192 × 5460 pixels (Zenmuse P1, DJI, Shenzhen, China) with a fixed-focus lens (focal length 50 mm, maximum aperture f/2.8, and field of view 46.8°). The other was an integrated sensor system comprising a multispectral of 1280 × 960 pixels and a thermal infrared camera of 640 × 512 pixels (AIRPHEN, Hi-phen, Avignon, France). The multispectral camera included six spectral bands covering the blue (450 nm), green (530 nm and 570 nm), red (675 nm), red-edge (730 nm), and near-infrared (850 nm). UAV data were collected five times (Fig. A1), corresponding to five critical growth stages of rice: tillering, jointing, booting, flowering, and maturity stages.

To balance data acquisition efficiency and spatial accuracy, the flight altitudes were set at 25m for the RGB camera and 20m for the multispectral camera, as the resolution of the latter is lower than that of the former [[30], [31], [32], [33], [34], [35]]. Both missions followed a zigzag-shaped flight path, with the cameras oriented vertically downward. Before each flight, the field was equipped with 12 fixed ground control points (GCPs) for geometric calibration, three spectral reflectance reference panels with reflectivity values of 5%, 30%, and 70% for radiometric calibration, and a heated calibration plate maintained at a constant temperature of 40 °C for thermal calibration.

Manually measured data were sampled in each plot during five key growth stages (seedling, tillering, jointing, booting, and maturity; Fig. A1). For each plot, three representative hills were randomly sampled to manually measure the traits after UAV data acquisition at all growth stages, except during the tillering stage, when ten hills were sampled. Specifically, nine traits were measured in the laboratory, including plant height, tiller number, leaf nutrient contents (N⁺, P⁺, K⁺, and Na⁺), SPAD, dry weight, and fresh weight. Moreover, a 2 m × 2 m quadrat at the center of each plot was used to measure yield at the maturity stage.

2.2. Data preprocessing and phenotype extraction

Data preprocessing includes ortho-mosaic images generation, geometric correction, radiometric correction, temperature calibration, plot segmentation, and canopy extraction. Ortho-mosaic images from UAV-acquired RGB and multispectral data were generated using Pix4Dmapper software (Pix4D SA, Lausanne, Switzerland). To improve both vertical and horizontal positional accuracy and to ensure precise spatial alignment across different time points and sensor types, 12 ground control points (GCPs) were incorporated into the ortho-mosaic image generation workflow for georeferencing. The spectral reflectance of the multispectral images was calibrated using a linear regression model based on the image-derived and actual reflectance values of three standard reference panels (5%, 30%, and 70%). Similarly, thermal images were calibrated using a linear model constructed from the known temperature and reflectance of the heated calibration plate (Appendix method 1). To enable plot-level phenotypic extraction, UAV-based RGB and multispectral data were segmented into individual plots using ArcGIS software (Esri, Redlands, USA). Considering the generalization of the model, this study used the enhanced vegetation index (EVI) and the fixed threshold method to extract the canopy of the sample plot. The parameterization was mainly determined by referring to existing studies [36] and the preliminary experimental analysis of this study (Figure A2).

Based on the plot canopy, a total of 41 phenotypes were extracted to screen salt-tolerant and high-yielding rice varieties. These phenotypes included canopy coverage, spectral, textural, and thermal features. Spectral traits were derived from MSI and comprised 36 vegetation indices (Table A2), which have been proven effective in yield prediction. Textural traits were extracted from RGB images using the GLCM method and included four features: mean, variance, entropy, and maximum probability [36]. Temperature traits were represented by the average temperature of each plot derived from TIR images. Additionally, the collected whole-stage data were categorized into early-stage (tillering and jointing) and late-stage (heading, booting, and maturity) (Fig. A1).

2.3. Salt-tolerant rice varieties screening via membership function method

To address the lack of control treatments under saline-alkaline field conditions, salt-tolerant rice in this study was evaluated based on relative phenotypic differentiation among varieties grown under uniform saline-alkaline stress, rather than by comparison with non-stressed controls. This strategy is commonly adopted in large-scale breeding trials, where establishing paired control plots is often impractical due to land and management constraints [[37], [38], [39], [40]]. Under such conditions, varietal salt tolerance is inferred from comparative performance across genotypes exposed to the same stress environment, which allows effective discrimination of tolerant and sensitive varieties.

To provide a reference for evaluating the UAV-based screening results, a classic MF method based on manually measured agronomic traits was adopted as a reference. The MF method integrates multiple stress-related traits into a comprehensive salt-tolerance index, enabling the ranking and discrimination of varietal tolerance under uniform stress conditions.

2.3.1. Salt-tolerant rice varieties screening via manually measured data

Firstly, the comprehensive salt tolerance value (D-value) derived from the MF was used to evaluate the salt tolerance of different rice varieties under salt stress. The calculation formula is as follows [41]:

$$D=\sum _{i=1}^n\left[u\left(X_i\right)\times W_i\right],i=1,2,\dots ,n$$ (1)

Where $X_i$, $\mathit{u}\left(X_i\right)$ and $W_i$ respectively represent the principal component value, the membership function value, and the weight of the i-th comprehensive index, with $\mathit{i}=1,2,\dots ,n$, and $\mathit{n}$ is the total number of varieties. Their calculation formulas are as follows (Equations (2), (3))) [12,27,42],

$$u\left(X_i\right)=\left(X_i-X_{\min }\right)/\left(X_{\max }-X_{\min }\right)$$ (2)

$$W_i=\frac{P_i}{\sum _{i=1}^n{P}_i},i=1,2,\dots ,n$$ (3)

where $X_{\max }$ and $X_{\min }$ represent the maximum and minimum values of an index of all the tested varieties, respectively; $P_i$ represents the contribution rate of the i-th comprehensive index of each tested material obtained by PCA.

Then, the D-value for each rice variety was calculated using nine manually measured data: plant height, tiller number, leaf nutrient contents (N⁺, P⁺, K⁺, and Na⁺), SPAD, dry weight, and fresh weight. Finally, based on the D-values, the hierarchical clustering method [43] was used to classify the rice varieties into three levels: salt-tolerant, moderately salt-tolerant, and salt-sensitive.

2.3.2. Salt-tolerant rice varieties screening via UAV phenotypes

To prove that high-throughput UAV phenotypes can be used for variety screening, this study constructed a membership evaluation method based on UAV phenotypes (hereafter as MFuav). Firstly, UAV phenotypes that showed a high correlation with the D-value were selected based on Pearson correlation coefficients throughout the whole growth stages (Fig. A3). Then, the selected UAV phenotypes, instead of manually measured data, were input into the MF model to compute the D-value. Finally, the D-value obtained through the MFuav was used to classify rice varieties into three categories (i.e., salt-tolerant, moderately salt-tolerant, and salt-sensitive) using the hierarchical clustering method. Additionally, the classification results of the MFuav at early-, late-, and whole-stage were compared to prove the potential of early screening from UAV phenotypes.

2.4. Early-stage screening of salt-tolerant and high-yielding rice varieties via machine learning

To achieve early-stage screening of salt-tolerant and high-yielding rice varieties, this study proposed an ML-based method with UAV phenotypes that includes three steps. The first step involved screening salt-tolerant rice varieties, while the second step focused on predicting rice yield levels. Both steps were implemented with ML models based on UAV phenotypes. In the third step, the intersection of the results from the first two steps was taken to achieve the final screening objective.

2.4.1. Early-stage screening of salt-tolerant rice varieties via machine learning

The proposed ML method with early-stage UAV phenotypes (MLuav) for classifying salt tolerance levels of rice varieties includes three steps. Firstly, the early-stage UAV phenotypes that showed high correlations with the D-values were selected based on Pearson correlation coefficients (Fig. A3). Secondly, these selected phenotypes were used to train six ML models, including multiple linear regression (MLR), Lasso regression (Lasso), Ridge regression (Ridge), partial least squares regression (PLSR), Random Forest (RF), and support vector machine (SVM), for predicting the D-values. In detail, the dataset of 61 rice samples was divided into a training set and a testing set at a 1:1 ratio. If the number of samples could not be evenly divided, the training set was rounded up, and the remaining samples were assigned to the testing set. To ensure strict data independence and avoid information leakage, all duplicate samples of the same variety are only assigned to the training set or the test set during the data splitting process. As a result, the numbers of salt-tolerant, moderately salt-tolerant, and salt-sensitive in the training set and testing set were 9, 17, 5, and 9, 16, 5, respectively. Then, all input phenotypic traits were standardized to eliminate differences between feature scales. On this basis, the ML models were trained and optimized using grid search for hyperparameter tuning, except for the MLR because it does not have a hyperparameter. Finally, the predicted D-values were clustered using the hierarchical clustering method to classify rice varieties into three categories: salt-tolerant, moderately salt-tolerant, and salt-sensitive.

2.4.2. Early-stage screening of high-yielding rice varieties via machine learning

The ML methods with early-stage UAV phenotypes were designed to screen high-yielding rice varieties, which include the following three steps. Firstly, early-stage UAV phenotypes that showed high correlation with the measured yield were selected based on Pearson correlation coefficients (Fig. A4). Then, the selected early-stage phenotypes were used to train six ML models (MLR, Lasso, Ridge, PLSR, RF, and SVM) for predicting yield. In detail, the dataset division, feature preprocessing, and hyperparameter tuning for the six ML models followed the same protocols described in Section 2.4.1. Finally, the predicted yields were clustered using the hierarchical clustering method to classify rice varieties into three levels: high-yielding, moderate-yielding, and low-yielding. Moreover, the three referenced yielding levels were classified using the hierarchical clustering method from the measured yield.

2.4.3. Early-stage screening of salt-tolerant and high-yielding rice varieties

Salt-tolerant and high-yielding rice varieties were obtained by taking the intersection of the results of the salt-tolerant variety screening and the high-yielding variety screening. To prove the superiority of the MLuav methods, the salt-tolerant levels of the MFuav and MLuav methods were respectively intersected with the yield levels predicted by ML models. In addition, the referenced salt-tolerant and high-yielding rice varieties were obtained by intersecting the MF-derived salt-tolerant varieties of the whole growth stage and the high-yielding varieties clustered from the manually measured data.

2.5. Evaluation criteria

To compare the screening performance among different methods, the confusion matrix between the prediction results of different methods and the reference results was constructed separately. Overall accuracy (OA), Precision, Recall, and F1-score (F1) were calculated from the confusion matrix. These four metrics range from 0 to 1, with higher values indicating better classification performance.

For machine learning-based methods, the dataset was first divided into an independent training set and testing set at a 1:1 ratio, as described in Section 2.4.1. The independent testing set was used exclusively for final performance evaluation and reporting of classification metrics.

Within the training set only, 10-fold cross-validation was applied for hyperparameter tuning and model optimization when predicting both D-values and yield. Given the limited sample size, cross-validation was used to reduce sensitivity to a single data split during model training rather than as the performance evaluation. The testing set was not involved in any stage of model training or cross-validation to ensure strict data independence.

3. Results

3.1. Salt-tolerant rice varieties screening via membership functions

Using the salt tolerance levels classified by the MF as a reference (Fig. A5), the classification accuracies of the MFuav (Fig. A6a-c) at different growth stages were evaluated. The results (Fig. 2) revealed that: (1) Overall, the early-stage yielded the highest classification performance for salt-tolerant rice varieties using the MFuav among the other compared stages. In contrast, the late- and whole-stage showed comparable but lower performance. (2) For salt-tolerant rice varieties, early-stage classification outperformed both the late- and whole-stage across all four metrics. In detail, the OA, Precision, Recall, and F1 of early-stage classification were 0.70, 0.86, 0.46, and 0.60, which were on average 0.04, 0.36, 0.07, and 0.17 higher than those of the other two compared stages. (3) For moderate salt-tolerant rice varieties, both the early- and late-stage yielded better classification accuracy than the whole-stage. For salt-sensitive rice varieties, the early- and whole-stage provided higher classification accuracy than the late-stage.

Fig. 2.

Classification accuracy of salt tolerance levels using the MFuav based on different UAV phenotypes at different stages. Early, late, and whole represent the results obtained via the MFuav method using early-, late-, and whole-stage UAV phenotypes. Red color font means the highest value.

3.2. Early-stage screening of salt-tolerant and high-yielding rice varieties via machine learning

3.2.1. Salt-tolerant rice varieties screening

To achieve early-stage screening of salt-tolerant rice varieties, the PLSR was selected for the MLuav approach because it exhibited the highest accuracy (R² = 0.71, RMSE = 0.05; Table A3) among the six ML models (MLR, Lasso, Ridge, PLSR, RF, and SVM). Furthermore, the early-stage classification accuracy of PLSR was the highest among the three compared stages (Fig. A7).

The classification accuracy (Fig. 3) between the MLuav (Fig. A6d) and MFuav (Fig. A6a) showed that: (1) Overall, the MLuav demonstrated superior performance over the MFuav in screening rice varieties across all salt tolerance levels, although both methods performed the same in classifying salt-sensitive varieties. (2) For salt-tolerant rice varieties, the MLuav outperformed the MFuav across all four metrics. In detail, OA, Precision, Recall, and F1 of the MLuav were 0.78, 0.97, 0.54, and 0.70, which were respectively 0.08, 0.11, 0.08, and 0.10 higher than those of the MFuav.

Fig. 3.

Classification accuracy of salt tolerance levels based on UAV phenotype using the MLuav and MFuav methods. Red font and underlined font represent the highest value.

3.2.2. High-yielding rice varieties screening

To achieve early-stage screening of high-yielding rice varieties, the RF was selected because it exhibited the highest accuracy (R² = 0.75 and RMSE = 14.77 kg/mu (Table A3); among the six ML models (MLR, Lasso, Ridge, PLSR, RF, and SVM).

Taking the measured yield classification (Fig. A8) as a reference, the predicted yield classification results by the RF (Fig. A9) were evaluated. The results (Fig. 4) showed that: (1) For the high-yielding rice varieties, Recall was the highest, followed by OA, F1, and Precision. (2) For the moderate-yielding rice varieties, OA was the highest, with the other three metrics showing no substantial differences. (3) For the low-yielding rice varieties, Precision was the highest, followed by OA, but Recall and F1 were relatively low. In summary, the screening performance for high-yielding rice varieties was characterized by the highest Recall among the three yield levels, indicating that the model effectively identified all high-yielding candidates. However, this was achieved at the expense of Precision, suggesting a trade-off between sensitivity and specificity in early-stage yield screening.

Fig. 4.

Classification accuracy of the rice yield levels by the Random Forest model.

3.3. Salt-tolerant and high-yielding rice varieties screening

The salt-tolerant and high-yielding rice varieties screening results showed that: (1) Using early-stage UAV phenotypes, the MFuav screened four salt-tolerant and high-yielding varieties, among which three were consistent with the reference. (2) In comparison, the MLuav screened five varieties based on early-stage phenotypes, of which four matched the referenced (Table 1). Therefore, MLuav outperformed MFuav in screening salt-tolerant and high-yielding rice varieties using the early-stage UAV phenotypes.

Table 1.

Screening results of salt-tolerant and high-yielding rice varieties by different methods, where ∗ represents the referenced salt-tolerant and high-yielding rice varieties calculated using the measured data, and the others are calculated using the UAV phenotypes. The green circle indicates consistency with the reference, while the yellow indicates inconsistency. The numbers in circles represent identities of different varieties.

Method	Stage	Salt-tolerant and high-yielding	Repetition
MF	Whole∗		5
MFuav	Early		3
MLuav	Early		4

4. Discussion

4.1. UAV phenotyping empowers efficient and early screening of elite varieties

Rapid and accurate data acquisition is a critical prerequisite for screening elite varieties in large-scale field experiments. Although manual measurements can provide fine-scale phenotypic data, it is time-consuming and labor-intensive. In contrast, UAV phenotyping offers rapid [44,45], non-destructive, and time-series collection capabilities [46], making it highly suitable for large-scale germplasm evaluation [47]. In this study, the UAV phenotyping platform significantly improved the efficiency of data collection across 61 field plots (6271.20 m²), reducing the measurement duration from two days of manual work to less than 1 h.

Meanwhile, this study demonstrated that UAV phenotypes enable early screening of elite varieties due to their time-series and multi-source data acquisition strength [20]. The early-stage yielded the highest screening performance for salt-tolerant rice varieties using MFuav among the other compared stages (Fig. 2). The possible reasons are that rice is highly sensitive to salt stress at early growth stages, particularly before tillering and during flowering [48]. This advantage can be attributed to two key factors: On the one hand, time-series UAV data show remarkable advantages in capturing the early-stage phenotype changes. On the other hand, multi-source spectral data showed high correlations with physiological status under salt stress, particularly variations in leaf pigment and nutrient content (ig. A10-14). The early-stage and multispectral UAV phenotypes were proved effective in distinguishing the salt tolerance of rice varieties using the PCA, especially with the clearest separation observed in salt-sensitive rice varieties (Fig. 5a). Among them, the structure insensitive pigment index (SIPI), the plant senescence reflectance index (PSRI), the ratio vegetation index (RVI), transformed the chlorophyll absorption in reflectance index (TCARI), and the chlorophyll vegetation index (CVI) were negatively correlated with the first principal component (PC1) (Fig. 5b), which are typically associated with plant senescence, chlorophyll content [49,50], and canopy structure [51]. Similarly, UAV phenotypes have also been used in screening crop germplasm such as drought tolerance [52], cold tolerance [53], and heat tolerance [54]. These studies all indicate that UAV is a rapid and non-destructive tool for in-situ and early screening of elite varieties.

Fig. 5.

Analysis of UAV phenotype contributions to the classification of salt-tolerant rice varieties. (a) Biplot of the early-stage UAV phenotypes used in the MLuav and MFuav. The colorful dots represent rice varieties with different salt tolerance. Green, yellow, and red represent salt-tolerant, moderately salt-tolerant, and salt-sensitive varieties, respectively. Light gray lines and black lines represent UAV phenotypes that are positively and negatively correlated with PC1, respectively. (b) Contributions of UAV phenotypes to PC1 and PC2.

4.2. ML outperforms MF in the accurate screening of salt-tolerant and high-yielding rice varieties

Accurate screening models are fundamental to the identification and breeding of stress-tolerant crop varieties under complex field environments [55]. This study proposed a data-driven approach (i.e., ML) that outperforms the model-driven method (i.e., MF) (Fig. 3; Table 1).

The MF method has been widely used in screening elite varieties because it synthesizes multiple phenotypic traits into a comprehensive evaluation index based on weighted functions, differing from index construction-based methods with only a few traits [28,56,57]. The MF can comprehensively represent varietal stress responses and has demonstrated effectiveness in salt tolerance assessments, when using physiological or biochemical traits, such as biomass, ion concentration, or nutrient content [18,58]. However, these traits are obtained from destructive sampling [8] and labor-intensive measurements [55,59,60], limiting their scalability in large-scale field screening.

To overcome these limitations, this study proposed a high-throughput and early screening strategy with UAV phenotypes (MFuav). On the one hand, UAV phenotypes allow for rapid, non-destructive measurement of structural and biophysical traits related to plant growth and yield [19]. On the other hand, UAV phenotyping provides rich data with broad spatiotemporal coverage, strong temporal resolution, and high spectral dimensionality (Fig. 5), supporting improved screening of salt tolerance. Our results showed that the MFuav based on the early-stage UAV phenotypes achieved the highest screening accuracy for salt-tolerant rice varieties (Fig. 2), aligning with the biological understanding that rice is most sensitive to salt stress during early tillering [48,61]. This suggests that integrating UAV phenotypes into the MF frameworks can retain reasonable accuracy while improving efficiency and throughput. However, the MF still faces limitations in its reliance on linear dimensionality reduction techniques (e.g., PCA) to determine variable weights (Equation (3)), which may not perform well under non-normal distributions or when feature correlations are weak [62]. Additionally, the MF requires expert-defined assumptions, reducing its adaptability to different datasets or environments.

This study further proposed the MLuav model, which offers higher flexibility, automation, and predictive power because it can robustly capture nonlinear and high-dimensional relationships between UAV phenotypes and varietal performance [18,63]. In this study, the ML models, particularly the RF and PLSR, effectively integrated spectral, structural, and textural UAV phenotypes to screen salt-tolerant and high-yielding rice varieties with significantly higher accuracy than the MFuav (Fig. 3, Table 1). The overall classification accuracy of the ML models improved by approximately 15% compared to the MFuav, which may be attributed to their inherent feature selection mechanisms (e.g., variable importance in RF) and robustness to multicollinearity [[64], [65], [66]]. In addition, the ML models can integrate spectral features, textural characteristics, and environmental variables, effectively capturing complex genotype-phenotype-environment interactions through nonlinear modeling [67], which can also benefit the screening accuracy.

4.3. Contributions and implications

This study developed a novel method for screening salt-tolerant and high-yielding rice varieties under saline-alkaline field conditions by integrating UAV phenotypes with ML models. The main contributions are as follows: (1) Proving that the high-throughput UAV phenotypes can be used for variety screening by constructing a membership evaluation method (MFuav). Meanwhile, the MFuav realized early-stage screening of salt-tolerant rice varieties, benefiting from the time-series UAV phenotypes. (2) Developing an accurate salt-tolerant screening method combining ML models with UAV phenotypes (MLuav), which outperformed the MFuav at the early-stage. (3) The MLuav was proved more accurate than the MFuav in screening salt-tolerant and high-yielding rice varieties. This approach provided a non-destructive and high-throughput method for screening elite varieties under field conditions at an early-stage.

However, several directions emerged from this study that merit further exploration in future research. (1) Advanced sensing technologies for deeper physiological insight: Although this study utilized cost-effective multispectral imaging, the integration of advanced technologies such as hyperspectral imaging [68] or chlorophyll fluorescence [11] may enhance the precision of physiological trait assessment and deepen our understanding of crop responses to salinity at a biochemical level. (2) Model generalization and scalability: While 10-fold cross-validation ensured robust model evaluation, further validation in larger, multi-location field trials is necessary to assess model applicability under diverse conditions [69]. Future studies could also explore transfer learning or ensemble learning frameworks to improve model robustness and generalizability across environments [18,70]. (3) Interpretable machine learning and genetic integration: Understanding which phenotypic features most influence ML predictions can provide physiological insights into salt tolerance. Tools such as SHAP (SHapley Additive exPlanations) [51] can help screen key UAV indicators. Combined with genome-wide association studies (GWAS), this may accelerate the discovery of salt tolerance-related QTLs and genes, ultimately advancing precision breeding [71]. (4) Temporal feature utilization and dynamic phenotyping: Although early-stage phenotypic analysis already demonstrated strong screening potential in this study, incorporating finer temporal information, including earlier growth stages, may further improve sensitivity and biological interpretability [46,72]. Future studies should deeply explore the UAV temporal phenotypes for screening high-yielding and salt tolerance rice varieties [73].

5. Conclusion

This study quantitatively evaluated the stress resistance of 60 rice verities under saline-alkaline field conditions and established an efficient UAV-based screening method. Firstly, a high-throughput screening method (MFuav) was proposed by replacing manual measurements in the MF with UAV phenotypes. The MFuav realized early-stage screening of salt-tolerant rice varieties (Precision >0.8, OA > 0.7). Secondly, a more accurate screening method was proposed by combining ML with UAV phenotypes (MLuav). The MLuav achieved a Precision of 0.97 and an OA of 0.78 at the early-stage, outperforming the MFuav by 0.11 and 0.08, respectively. Meanwhile, the MLuav enabled accurate high-yielding level identification (OA = 0.70, F1 = 0.67 and Recall = 1.00), which is advantageous for preliminary germplasm selection under field conditions. Finally, the MLuav successfully screened four out of five salt-tolerant and high-yielding reference rice varieties at the early-stage, outperforming the MFuav method, which screened only three.

Looking forward, future research should explore advanced remote sensing techniques (e.g., hyperspectral or fluorescence imaging) to gain deeper insights into plant physiological responses. Expanding field trials, utilizing temporal feature and incorporating transfer learning can enhance model generalization. Additionally, integrating interpretable ML methods (e.g., SHAP) with molecular genetics may help screen key phenotypic indicators and accelerate gene discovery for salt tolerance, thereby supporting breeding programs with greater precision and efficiency.

Author contributions

S. Jin and X. Zhang designed the idea of this study. S. Jin and X. Zhang performed data analysis. X. Zhang, S. Li, and T. Wang collected and preprocessed the data; X. Zhang and S. Jin wrote the paper and crafted all figures and tables. All authors helped to revise the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported in part by the National Key Research and Development Program of China (2022YFD2300700; 2022YFD1900700), the JBGS Project of Seed Industry Revitalization in Jiangsu Province (JBGS [2021] 007), the National Natural Science Foundation of China (32201656), 2025 AI for Graduate Education Program (2025AIYJSZX-YB031), and the Young Elite Scientists Sponsorship Program by CAST (YESS).

Data availability

The core code and data that support the findings of this study are available from the corresponding author (S. J.) upon reasonable request due to ongoing collaborative work.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Given their respective roles as Editor-in-Chief and Senior Editor, Yanfeng Ding and Shichao Jin had no involvement in the peer review of this article and had no access to information regarding its peer review. Full responsibility for the editorial process for this article was delegated to another journal editor.

Appendix.

Appendix method 1:

Multispectral Radiometric Calibration

For multispectral image calibration, three reference panels with known spectral reflectance values (5%, 30%, and 70%), representing low-, medium-, and high-reflectance targets, were used.

The mean digital number (DN) values of each reference panel were extracted from the corresponding regions in the raw multispectral images. A linear regression model was then established between the extracted DN values and the true reflectance values for each spectral band independently, and the resulting regression coefficients were used for band-wise radiometric calibration.

The calibration model for a single spectral band can be expressed as:

$$Y_i=a{X}_i+b,i=1,2,3$$ (1)

where $\mathit{a}$ and $\mathit{b}$ denote the radiometric calibration coefficients for the given bands; $X_1,X_2,X_3$ represent the mean DN values of the black (5%), gray (30%), and white (70%) reference panels in the uncalibrated image; and $Y_1,Y_2,Y_3$ correspond to their respective true reflectance values.

Thermal Image Calibration

Similarly, thermal images were calibrated using a constant-temperature heating plate set to 40 °C as a reference temperature source.

The mean DN value of the heating plate was extracted from the thermal image, and a linear proportional calibration model was applied to convert DN values to temperature values. The calibration relationship is defined as:

$$\frac{T_i}{DNi}=\frac{T_r}{DNr}$$ (2)

where $T_r$ and $\mathit{D}{N}_r$ represent the true temperature of the heating plate and its corresponding mean DN value in the raw thermal image, respectively; $\mathit{D}{N}_i$ is the DN value of pixel i in the thermal image; and $T_i$ is the calibrated temperature of that pixel.

Table A1.

The name and identity of rice varieties

Number	Name	Number	Name
N01	Huajing 5	N31	Suken 93002
N02	Huajing 9	N32	Fengjing 4
N03	Huajing 10	N33	Huajing 2022
N04	Huajing 12	N34	Ruijing 688
N05	Huajing 13	N35	Jindao 565
N06	Huajing 0098	N36	Zijinnuo 1
N07	Suxiu 852	N37	99-25
N08	Suxiu 852	N38	Huazhenuo 1
N09	Suxiu 1928	N39	Liannuo 1
N10	Suxiu 298	N40	Huainuo 52
N11	Lianjing 307	N41	Huainuo 51
N12	Lianjing 405	N42	Dahuanuo 168
N13	Lianjing 11	N43	Suxiunuo 889
N14	Lianjing 166	N44	Jinjing 667
N15	Xudao 15	N45	Runnong 17
N16	Jiayuanjing 1	N46	Runnong 802
N17	Fengjing 908	N47	Rundao 303
N18	Zhongyandao 88	N48	Jinjing 818
N19	Guangcan 1	N49	Shengdao 23
N20	Taijing 5201	Z01	Huajing 22022
N21	Jingdao 88	Z02	Huajing 12
N22	Sidao 23	Z03	Huajing 15
N23	Sidao 309	Z04	Huajing 16
N24	Sidao 21	Z05	Huajing 0034
N25	Shengdao 23	Z06	Huajing 0100
N26	Huaidao 5	Z07	Huajing 0111
N27	Nanjing 5718	Z08	Huajing 0119
N28	Nanjing 66	Z09	Huajing 2368
N29	Suken 218	Z11	Huajing 2387
N30	Zhongdao 1	Z12	Huajing 2325
N30	Zhongdao 1

Fig. A1.

Timeline of UAV-based and manual data collection.

Fig. A2.

Fixed-threshold determination procedure. (a) Random sampling of n canopy pixels and n background pixels from UAV images. (b) EVI image corresponding to the sampled pixels in (a). (c) Distribution of the n sampled EVI values. In this example, yellow and red points represent background and rice canopy pixels from an UAV imagery, respectively. The quartiles of the background pixels were 0.45 and 0.53, while those of the canopy pixels were 0.90 and 0.97, with the differences being statistically significant. To minimize background interference, the lower quartile of the canopy pixels (0.90) was selected as the fixed threshold for segmentation.

Table A2.

Definition of UAV spectral traits

Features	Formulation	References
Ratio vegetation index	RVI = NIR/R	[74]
Green chlorophyll index	GCI = (NIR/G) - 1	[75]
Red-edge chlorophyll index	RECI = (NIR/RE) - 1	[75]
Normalized difference vegetation index	NDVI = (NIR - R)/(NIR + R)	[76]
Green normalized difference vegetation index	GNDVI = (NIR - G)/(NIR + G)	[77]
Green-red vegetation index	GRVI = (G - R)/(G + R)	[78]
Normalized difference red-edge	NDRE = (NIR - RE)/(NIR + RE)	[79]
Normalized difference red-edge index	NDREI = (RE - G)/(RE + G)	[80]
Simplified canopy chlorophyll content index	SCCCI = NDRE/NDVI	[81]
The enhanced vegetation index	EVI = 2.5 ∗ (NIR - R)/(1 + NIR - 2.4 ∗ R)	[82]
Two-band enhanced vegetation index	EVI2 = 2.5 ∗ (NIR - R)/(1 + NIR + 2.4 ∗ R)	[75]
Optimized soil adjusted vegetation index	OSAVI = (NIR - R)/(NIR - R + 0.16)	[75]
Modified chlorophyll absorption in reflectance index	MCARI = [(RE - R) - 0.2 ∗ (RE - G)] ∗ (RE/R)	[83]
Transformed chlorophyll absorption in reflectance index	TCARI = 3 ∗ [(RE - R) - 0.2 ∗ (RE - G) ∗ (RE/R)]	[84]
Wide dynamic range vegetation index	WDRVI = (0.12 ∗ NIR - R)/(0.12 ∗ NIR + R)	[75]
Plant senescence reflectance index	PSRI = (R – G)/RE	[85]
Difference vegetation index	DVI = NIR - R	[86]
Green difference vegetation index	GDVI = NIR - G	[78]
Red-Edge chlorophyll index	CIRE = (NIR/RE) - 1	[87]
Red-Edge chlorophyll index 2	CIRE2 = (RE/G) - 1	[77]
Kawashima index	IKAW = (R - B)/(R + B)	[88]
Red green blue vegetation index	RGBVI = (G ∗ G – B ∗ R)/(G ∗ G + B ∗ R)	[89]
Modified green red vegetation index	MGRVI = (G ∗ G – R ∗ R)/(G ∗ G + R ∗ R)	[89]
Visible-Band Difference Vegetation Index	VDVI = (2 ∗ G – R + B)/(2 ∗ G + R + B)	[90]
chlorophyll Indices	CIgreen = (NIR/G) - 1	[91]
chlorophyll vegetation index	CVI = (NIR/G) ∗ (R ∗ G)	[92]
Modified Normalized Vegetation Index	MNVI = 1.5 ∗ (NIR2 − R)/(NIR2 + R + 0.5)	[93]
Modified Soil-Adjusted Vegetation Index	MSAVI = [(2 ∗ NIR + 1) - sqrt [(2 ∗ NIR + 1) ∗ 2 - 8 ∗ (NIR - R)]]/2	[94]
MERIS Terrestrial Chlorophyll Index	MTCI = (NIR − RE)/(RE − R)	[95]
Modified Triangular Vegetation Index	MTVI = 1.2 ∗ (1.2 ∗ (NIR − G) − 2.5 ∗ (R − G))	[96]
Normalized Leaf Index	NLI = (NIR2 − R)/(NIR2 + R)	[97]
Photochemical Reflectance Index	PRI = (R531 - R570)/(R531 + R570)	[98]
red-edge chlorophyll index	RECI = (NIR/RE) - 1	[87]
Structure Insensitive Pigment Index	SIPI = (NIR - B)/(NIR + R)	[99]
Simple Ratio Index	SRI = NIR/RE	[74]
Triangular Vegetation Index	TVI = 0.5 ∗ [120 ∗ (RE – G) - 200 ∗ (R – G)]	[100]

Fig. A3.

The Pearson correlation between UAV phenotypes and the comprehensive salt tolerance value (D-value) based on the membership function. 0725, 0812, 0831, 0914, and 1017 represent UAV flight dates corresponding to the tillering, jointing, booting, flowering, and maturity stages, respectively.

Fig. A4.

The Pearson correlation between UAV phenotypes and the measured yield. 0725, 0812, 0831, 0914, and 1017 represent UAV flight dates corresponding to the tillering, jointing, booting, flowering, and maturity stages, respectively.

Fig. A5.

The classification result of salt tolerance level for 61 rice samples based on the comprehensive salt tolerance value (D-value) from the membership function using measured data, including plant height, tiller number, leaf nutrient contents (N, P, K, and Na), leaf chlorophyll content, dry weight, and fresh weight. The classification was performed using the hierarchical clustering method. Green, yellow, and red represent salt-tolerant, moderately salt-tolerant, and salt-sensitive varieties, respectively.

Fig. A6.

The classification result of salt tolerance levels in the testing dataset. (a), (b) and (c) represent the hierarchical clustering results derived from the membership function method (MFuav) based on early-, late-, and whole-stage UAV phenotypes, respectively. (d) shows the classification result predicted using early-stage UAV phenotypes via the partial least squares regression method (MLuav). Red, yellow, and green represent salt-tolerant, moderately salt-tolerant, and salt-sensitive varieties, respectively.

Table A3.

The accuracy of machine learning (ML) models was assessed by 10-fold cross-validation. R² values were from the training dataset. The highest and second-highest values were highlighted in bold and underlined.

ML model	Salt-tolerant classification		Yield prediction
	RMSE	R2	RMSE	R2
MLSR	0.185	0.378	32.825	0.475
Lasso	0.577	0.507	27.947	0.509
Ridge	0.111	0.471	33.871	0.497
PLSR	0.047	0.712	17.834	0.655
RF	0.073	0.560	14.772	0.753
SVM	0.667	0.484	27.440	0.550

Fig. A7.

The prediction accuracy of the comprehensive salt tolerance value (D-value) from the membership function using the partial least squares regression model (PLSR). (a), (b) and (c) respectively represent results based on early-, late-, and whole-stage UAV phenotypes, using the PLSR model. Red font indicates the R² from 10-fold cross-validation on the training dataset, while black font indicates the R² and RMSE on the testing dataset.

Fig. A8.

The classification result of yield levels for 61 rice samples. The results were obtained based on measured yield using the hierarchical clustering method. Green, yellow, and red represent high-yielding, moderate-yielding, and low-yielding varieties, respectively.

Fig. A9.

The classification result of yield levels in the testing dataset using early-stage UAV phenotypes via the random forest model (RF). Green, yellow, and red indicate high-yielding, moderate-yielding, and low-yielding varieties, respectively.

Fig. A10.

The Pearson correlation between UAV phenotypes at the tillering stage and nine laboratory-measured traits at the five stages. The suffixes 1 to 5 of the labels on the vertical coordinate axis represent the tillering, jointing, booting, flowering, and maturity stages, respectively. The traits include plant height, tiller number, leaf nutrient contents (N, P, K, and Na), leaf chlorophyll content (SPAD), dry weight, and fresh weight.

Fig. A11.

The Pearson correlation between UAV phenotypes at the jointing stage and nine laboratory-measured traits at the five stages. The suffixes 1 to 5 of the labels on the vertical coordinate axis represent the tillering, jointing, booting, flowering, and maturity stages, respectively. The traits include plant height, tiller number, leaf nutrient contents (N, P, K, and Na), leaf chlorophyll content (SPAD), dry weight, and fresh weight.

Fig. A12.

The Pearson correlation between UAV phenotypes at the booting stage and nine laboratory-measured traits at the five stages. The suffixes 1 to 5 of the labels on the vertical coordinate axis represent the tillering, jointing, booting, flowering, and maturity stages, respectively. The traits include plant height, tiller number, leaf nutrient contents (N, P, K, and Na), leaf chlorophyll content (SPAD), dry weight, and fresh weight.

Fig. A13.

The Pearson correlation between UAV phenotypes at the flowering stage and nine laboratory-measured traits at the five stages. The suffixes 1 to 5 of the labels on the vertical coordinate axis represent the tillering, jointing, booting, flowering, and maturity stages, respectively. The traits include plant height, tiller number, leaf nutrient contents (N, P, K, and Na), leaf chlorophyll content (SPAD), dry weight, and fresh weight.

Fig. A14.

The Pearson correlation between UAV phenotypes at the maturity stage and nine laboratory-measured traits at the five stages. The suffixes 1 to 5 of the labels on the vertical coordinate axis represent the tillering, jointing, booting, flowering, and maturity stages, respectively. The traits include plant height, tiller number, leaf nutrient contents (N, P, K, and Na), leaf chlorophyll content (SPAD), dry weight, and fresh weight.