Machine Learning Models for Yield Estimation of Hybrid and Conventional Japonica Rice Cultivars Using UAV Imagery

Abstract

Advancements in unmanned aerial vehicle (UAV) multispectral systems offer robust technical support for the precise and efficient estimation of japonica rice yield in cold regions within the framework of precision agriculture. These innovations also present a viable alternative to conventional yield estimation methods. However, recent research suggests that reliance solely on vegetation indices (VIs) may result in inaccurate yield estimations due to variations in crop cultivars, growth stages, and environmental conditions. This study investigated six fertilization gradient experiments involving two conventional japonica rice varieties (KY131, SJ22) and two hybrid japonica rice varieties (CY31, TLY619) at Yanjiagang Farm in Heilongjiang Province during 2023. By integrating UAV multispectral data with machine learning techniques, this research aimed to derive critical phenotypic parameters of rice and estimate yield. This study was conducted in two phases: In the first phase, models for assessing phenotypic traits such as leaf area index (LAI), canopy cover (CC), plant height (PH), and above-ground biomass (AGB) were developed using remote sensing spectral indices and machine learning algorithms, including Random Forest (RF), XGBoost, Support Vector Regression (SVR), and Backpropagation Neural Network (BPNN). In the second phase, plot yields for hybrid rice and conventional rice were predicted using key phenotypic parameters at critical growth stages through linear (Multiple Linear Regression, MLR) and nonlinear regression models (RF). The findings revealed that (1) Phenotypic traits at critical growth stages exhibited a strong correlation with rice yield, with correlation coefficients for LAI and CC exceeding 0.85 and (2) the accuracy of phenotypic trait evaluation using multispectral data was high, demonstrating practical applicability in production settings. Remarkably, the R² for CC based on the RF algorithm exceeded 0.9, while R² values for PH and AGB using the RF algorithm and for LAI using the XGBoost algorithm all surpassed 0.8. (3) Yield estimation performance was optimal at the heading (HD) stage, with the RF model achieving superior accuracy (R² = 0.86, RMSE = 0.59 t/ha) compared to other growth stages. These results underscore the immense potential of combining UAV multispectral data with machine learning techniques to enhance the accuracy of yield estimation for cold-region japonica rice. This innovative approach significantly supports optimized decision-making for farmers in precision agriculture and holds substantial practical value for rice yield estimation and the sustainable advancement of rice production.

1. Introduction

Amid the continuously increasing demands for food production and consumption, coupled with the significant challenges posed by climate change to agricultural systems, ensuring sustainable agricultural production has become a critical research priority. This necessitates the efficient allocation and optimization of various factors in food production, guided by accurate crop yield estimation [1,2]. Traditional growth monitoring methods primarily rely on field measurements, which are labor-intensive, time-consuming, and prone to crop damage and can also be heavily influenced by weather conditions [3]. Consequently, these conventional approaches have been empirically shown to be inefficient and inadequate for effective crop growth monitoring in agricultural fields. In contrast, remote sensing technologies enable farmers and technicians to detect and monitor a wide range of crop traits [4,5]. Unmanned aerial vehicle (UAV)-based remote sensing, in particular, offers prominent advantages, including broad spatial coverage, high cost-effectiveness, and a well-established technological framework, leading to its extensive application in precision agriculture [6,7]. As a staple food crop of global importance, rice production stability and improvement are directly linked to food security worldwide. Therefore, accurate pre-harvest prediction of rice yield holds significant practical value for ensuring stable and high production while promoting the sustainable use of agricultural resources such as water, fertilizers, and pesticides [8,9].

Spectral indices in crop remote sensing can reflect crop growth status and stress levels, making them valuable for yield assessment during critical growth stages. Wang F. et al. [10] identified the optimal timing for yield evaluation using UAV-based multispectral data, while Jian Wang et al. [11] assessed the impact of lodging on the accuracy of rice yield estimation. Xi Su et al. [12] combined spectral mixture analysis with multispectral indices to enhance estimation accuracy by deriving parameters related to leaves, soil, and water. However, the limited resolution of multispectral cameras restricts the acquisition of organ-level phenotypic information in rice plants. To address this, Yang et al. [13] improved prediction accuracy by integrating multispectral sensor data with color and texture features from RGB images. Wang F. et al. [14] further enhanced yield forecasting by incorporating fluorescence spectroscopy, and Hou Longfei et al. [15] integrated phenological stages and agronomic management practices into rice yield prediction models. In terms of yield estimation modeling, ensemble learning has emerged as a prominent research focus. Sarkar et al. [16] demonstrated the advantages of ensemble learning approaches after evaluating multiple algorithms. Similarly, with advances in artificial intelligence, Yang Q. et al. [17] reviewed CNN- and ANN-based algorithms for yield estimation using imagery.

Although yield estimation based on spectral indices has shown promising results, Sun et al. [18] noted that the correlation between these indices and rice yield is often weak, making direct yield predictions highly sensitive to variations in experimental data. Furthermore, because vegetation indices are influenced by multiple factors related to crop growth, relying solely on these parameters makes it challenging to provide actionable agricultural guidance. In contrast, Zha H. et al. [19] emphasized that plant height (PH), leaf area index (LAI), and biomass exhibit strong correlations with rice yield throughout the growth stages.

This study conducted a field experiment involving two conventional and two hybrid rice varieties, tested under six fertilizer gradients with three replicates each. Both UAV-based remote sensing data and manual measurements were collected to support yield modeling and analysis. The objectives of this research are to (i) assess the accuracy of spectral index-based models in estimating rice biomass, yield, and PH; (ii) evaluate the feasibility of predicting yield through the preliminary assessment of key growth stage phenotypes; and (iii) identify the optimal growth stage and most effective modeling approach for rice yield estimation.

2. Materials and Methods

Figure 1 provides a comprehensive overview of the methodology and multi-stage procedures employed in this study, including field experiments, computational modeling, machine learning regression modeling, and statistical regression analyses. The illustrated workflow begins with field data collection and UAV data acquisition, followed by Spearman correlation analysis to determine the relationship between phenotypic parameters and yield. Then, machine learning regression models are constructed to invert crop phenotypic parameters, and linear and nonlinear regression analyses are conducted for yield estimation. Details regarding data acquisition equipment, model construction, software utilized, and analytical techniques applied at each stage are further described in the following.

2.1. Study Area and Experimental Setup

This field experiment was conducted in 2023 at Yanjia Gang Farm, Harbin, Heilongjiang Province, China (126°18′ E, 45°36′ N), which features a mid-temperate continental monsoon climate. The site has an average annual precipitation of 380 mm and an average annual temperature of 3.4 °C, and the annual average accumulated temperature is 2851 °C, belonging to the first accumulated temperature zone. The average annual frost-free period is approximately 160 days. An overview of the experimental site and plot layout is presented in Figure 2.

In this experiment, two conventional japonica rice varieties (Kongyu131 [KY131] and Songjing22 [SJ22]) and two hybrid japonica rice varieties (Chuangyou31 [CY31] and Tianlongyou619 [TLY619]) served as plant materials. Detailed information on these rice varieties is provided in

Table 1. Information on tested rice varieties used in the experiment.

Group	Variety	Abbreviation	Variety Characteristics	Days to Maturity
Conventional japonica rice	Kongyu131	KY131	Early-maturing	127 D
	Songjing22	SJ22	Late-maturing	144 D
Hybrid japonica rice	Chuangyou31	CY31	Early-maturing	129 D
	Tianlongyou619	TLY619	Late-maturing	139 D

Information was retrieved from the China Rice Data Center (https://www.ricedata.cn/variety/) Accessed on 1 May 2025.

. For each variety, 6 plots with different nitrogen concentrations were setup (among the 6 nitrogen fertilizer treatments, the nitrogen application rates of different treatments are 0 (N1), 45 (N2), 90 (N3), 135 (N4), 180 (N5), and 225 (N6) kg/ha, respectively), with 3 replicates for each concentration, resulting in a total of 72 experimental plots. Each plot measured 25 m² in area, and adjacent plots were separated by ridges approximately 0.5 m wide. Seeds underwent a standard soaking and germination process prior to sowing. Sowing was performed on 15 April 2023, and seedlings were transplanted on 15 May 2023, at 30 days of age. Plant spacing was set at 13.3 cm × 30.0 cm. Three seedlings per hill were transplanted for conventional japonica rice, and two seedlings per hill for hybrid japonica rice. To minimize yield reduction, refined management strategies—including appropriate pest, disease, and weed control—were implemented throughout the growing season in line with local agricultural practices. The data in this study were collected from June to October 2023. During the data collection period of this study (June–October 2023), the climate in Harbin showed significant seasonal variations, specifically as follows: In June, the average temperature was 15–25 °C, with moderate rainfall, and the temperature and humidity conditions were coordinated, which was suitable for the growth and development of the tested crops. From July to August, the average temperature rose to 20–28 °C, with concentrated precipitation and sufficient light and heat resources, and the risk of waterlogging needed to be prevented during this stage. In September and October, the average temperature dropped to 10–20 °C, the rainfall decreased, and the tested crops were harvested one after another, so that the yield data could be accurately obtained. The aforementioned climatic characteristics provided typical regional environmental conditions for field experiments and data collection.

2.2. Data Collection and Preprocessing

2.2.1. Data Collection

During field data collection, six rice plants were randomly selected from the inner rows of each plot at four growth stages: tillering (TL), panicle initiation (PI), heading (HD), and maturity (MT). The selected plants showed uniform growth and were representative of the plot condition. LAI, CC, PH, and AGB were measured using a crop growth monitor. All measurements were performed between 08:00 and 10:00 local time to minimize the influence of variations in light intensity on data collection.

LAI was measured non-destructively through PAR interception using AccuPAR LP-80 ceptometers (METER Group, Inc., Pullman, WA, USA). The AccuPAR LP-80 calculates LAI by integrating PAR measurements taken above and below the canopy, alongside additional parameters such as canopy structure and solar position. The instrument automatically computes the zenith angle based on input parameters, including longitude, latitude, and local time at the measurement location. To obtain LAI, photosynthetically active radiation (PAR) intensity was first measured at the base of the rice canopy (approximately 5 cm above the water or soil surface), and immediately afterward, the incident PAR at the top of the canopy was measured. The instrument automatically computes the LAI value. Measurements were performed under diffuse light conditions to reduce artifacts caused by solar angle variations. All sensors underwent cosine correction and a 5-point inter-calibration process, achieving an inter-sensor variability of less than 3%.

Chlorophyll content (CC) was assessed using a SPAD-502 chlorophyll meter (Konica Minolta Sensing, Inc., Tokyo, Japan). In each plot, six fully expanded functional leaves exhibiting consistent growth were selected from the top of the canopy, specifically the first to third leaves from the apex, which contribute more than 70% to yield. Each leaf was measured three times at different positions to avoid repeated measurements at the same position, and the average of these three readings was recorded as the CC value for that leaf. Prior to measurement, leaf surfaces were cleaned of dust, and care was taken to avoid veins and damaged areas. The mean and variance of CC readings are provided in the table.

Aboveground biomass (AGB) was determined through destructive sampling following standardized protocols (GB/T 32725-2016; General Administration of Quality Supervision, Inspection and Quarantine of the People’s Republic of China (AQSIQ), Standardization Administration of China (SAC): Beijing, China, 2016.). For each plot, three representative rice plants were selected and cut at ground level. PH was immediately measured from the base to the tip of the highest panicle using a graduated pole with an accuracy of ±0.1 cm. All aboveground tissues, including stems, leaves, and panicles, were sealed in pre-weighed paper bags and dried in a forced-air oven at 75 °C for 48 h until a constant mass was achieved (mass variation < 0.5% between consecutive weightings at 2 h intervals). Dry weight was determined using an analytical balance with a resolution of 0.01 g.

Multispectral UAV data were collected using a drone-mounted multispectral sensor, synchronized with the dates of field data acquisition. A UAV (DJI Matrice 300 RTK) served as the remote sensing platform, equipped with a multispectral camera (MS600 Pro; camera specifications are detailed in

Table A5. Key specifications of the MS600pro multispectral camera.

Parameter	Parameter Value	Parameter	Parameter Value
Configuration	6 Spectral Bands	Dimensions	129 mm × 157 mm × 148 mm
Spectral Bands	450 nm, 555 nm, 660 nm, 720 nm, 750 nm, 840 nm	Mounting Interface	X-Port
Effective Pixels	1.2 Mpx	Power Consumption	7 W/10 W
Shutter Type	Global Shutter	Image Format	16-bit Raw TIFF
Radiometric Resolution (Bit Depth)	12 bit	Storage Medium	Standard 64 GB, supports up to 128 GB micro SD car
Field of View (FOV)	49.6° × 38°	Processing Software	Yusense Map
Ground Sampling Distance	8.65 cm @ h120 m	Parameter Configuration	DJI Pilot2
Swath Width	110 m × 83 m @ h120 m	Shooting Trigger	Interval Timer Trigger
Configuration	Sapphire Optical Glass Window	Capture Frequency	1 Hz

). Image acquisition was conducted under clear weather conditions, with flight times scheduled between 10:30 and 14:00 to minimize the impact of variable light conditions on the imagery. The flight altitude was set at 25 m, with a forward overlap of 80% and a side overlap of 70%, and images were captured at regular time intervals.

2.2.2. Data Preprocessing

Following UAC multispectral imagery acquisition, Yusense Map (V2.2.0) software was employed for image stitching and radiometric calibration. Prior to each flight, a calibrated gray panel was photographed for radiometric calibration. Multispectral images underwent radiometric correction through processes including band registration, image stitching, and calibration using YUSENSE Map software.

The stitched multispectral images were then segmented according to the plot boundaries to calculate the vegetation indices per plot. The stitched images were further processed on ENVI 5.6 software. Initially, the soil background was removed from the stitched images. Subsequently, the Region of Interest (ROI) tool was used to delineate each measurement area, with careful attention paid to exclude sampling zones and edge effect areas to enhance result accuracy. When calculating vegetation indices, the average reflectance of all pixels within the marked area was used as the representative canopy reflectance for the plot.

Several vegetation indices were employed in this study, each with specific sensitivities and applications: the Normalized Difference Vegetation Index (NDVI) exhibits strong sensitivity to chlorophyll concentration and leaf density; the Green Normalized Difference Vegetation Index (GNDVI) optimizes chlorophyll detection through enhanced green spectrum bands; the Soil-Adjusted Vegetation Index (SAVI) reduces soil background interference via integrated correction factors; the Optimized Soil-Adjusted Vegetation Index (OSAVI) adjusts soil adjustment coefficients for rice canopy architecture; the Modified Simple Ratio (MSR) mitigates saturation effects in dense vegetation through nonlinear transformation; the Difference Vegetation Index (DVI) quantifies biomass via linear red–NIR spectral contrast; the Nitrogen Reflectance Index (NRI) offers sensitivity to foliar chlorosis progression (particularly effective for CC inversion); the Normalized Difference Red Edge Index (NDRE) detects subtle chlorophyll variations through red edge spectral dynamics (720 nm); and the Structure-Insensitive Pigment Index (SIPI) indicates photosynthetic stress via carotenoid/chlorophyll ratios. The calculation formulas and references for these vegetation indices are presented in

Table 2. Vegetation indices and their formulas.

Vegetation Index	Formula	Reference
NDVI	$(NIR-RED)/(NIR+RED)$	Rouse, J. W et al., 1974 [20]
GNDVI	$(NIR-GREEN)/(NIR+GREEN)$	Anatoly A et al., 1996 [21]
SAVI	$\left(\right(NIR-RED)\times (1+L\left)\right)/(NIR+RED+L), L=0.5$	A.R Huete et al., 1988 [22]
OSAVI	$(NIR-RED)/(NIR+RED+0.16)$	Geneviève Rondeaux et al., 1988 [23]
MSR	$\left(\right(NIR/RED)-1)/\surd \left(\right(NIR/RED)+1)$	Compton J et al., 1979 [24]
DVI	$NIR-RED$	Filella, 1995 [25]
NRI	$(GREEN-RED)/(GREEN+RED)$	Jordan et al., 1969 [26]
NDRE	$(NIR-RE)/(NIR+RE)$	Barnes et al., 2000 [27]
SIPI	$(NIR-BLUE)/(NIR-RED)$	Peñuelas et al., 1995 [28]

. Plot segmentation and vegetation index calculations were performed using ENVI 5.3.

2.3. Construction of Rice Phenotypic Parameter Inversion Models

2.3.1. Correlation Analysis

Spearman’s rank correlation coefficient was used as a non-parametric statistical measure to quantify the relationship between variables. This method provides particular advantages when datasets violate parametric assumptions, such as cases involving limited sample sizes or non-normally distributed variables, where Pearson correlation analysis can yield biased estimates [29].

The dataset as a whole displays non-normal distributions. Spearman’s correlation analysis is particularly well-suited for non-normally distributed data; hence, this study employed the Spearman method to analyze the correlation between predicted phenotypic parameters and measured yield. This approach is consistent with established best practices in agricultural data analysis, ensuring reliable statistical inferences even when data deviate from normal distributions.

2.3.2. Phenotypic Inversion Model Development

The extracted vegetation indices served as input data, while rice phenotypic parameters including LAI, CC, PH, and AGB served as target variables, forming the training dataset for machine learning regression modeling. The dataset was split into 70% training and 30% testing sets to ensure independent model training and evaluation.

A rigorous data cleaning protocol was implemented: (1) missing value analysis confirmed the completeness of data across all variables; (2) outliers were identified using the Median Absolute Deviation (MAD) method with a threshold of |modified Z-score| > 3.5; and (3) four outlier plots, deemed measurement artifacts, were permanently excluded. This conservative strategy retained 96% of the original observations while effectively removing technical noise.

During the training stage, several machine learning algorithms, including RF, XG Boost, SVR, and BPNN, were utilized to establish predictive relationships between input features and rice phenotypic data. The modeling procedure involved two main stages. First, feature selection and optimization were performed through correlation analysis or model-based feature importance evaluation methods. This step identified the most influential spectral features and vegetation indices, helping to reduce redundant information and improve model efficiency. Next, model training was performed by fitting parameters with the training set data. RF constructs multiple decision trees and employs ensemble learning techniques to enhance stability and mitigate overfitting [30]. XG Boost is a boosting-tree model that uses gradient-based optimization strategies to improve prediction accuracy while applying regularization to prevent overfitting [31]. SVR is suitable for modeling small-sample, high-dimensional datasets, effectively capturing complex relationships between spectral features and rice phenotypic data [32]. BPNN is a multilayer feedforward neural network that iteratively optimizes weights via the backpropagation algorithm, making it particularly effective at capturing nonlinear functions and providing strong fitting capabilities [33]. Each of these four models has distinct advantages; thus, all were selected for modeling in this study.

To mitigate overfitting and enhance the robustness of the evaluation outcomes, 3-fold cross-validation was adopted in this study. Field plots were used as grouping units to preserve spatial integrity; each treatment was set with 3 repetitions to ensure the accuracy of the data. To optimize model performance, parameter tuning and debugging were conducted for each model. Details of the adjustment and debugging of specific parameters can be found in Appendix A,

Table A1. Optimized hyperparameters for the RF model.

Hyperparameter Name	Value
n_estimators	70
max_depth	10
min_samples_split	2
min_samples_leaf	1
bootstrap	TRUE

,

Table A2. Optimized hyperparameters for the XGBoost model.

Hyperparameter Name	Value
n_estimators	70
learning_rate	0.05
max_depth	10
gamma	0
subsample	1
colsample_bytree	1
colsample_bylevel	1
lambda	1
alpha	0
objective	reg
tree_method	auto

,

Table A3. Optimized hyperparameters for the SVR model.

Hyperparameter Name	Value
kernel	rbf
C	60
gamma	auto
degree	3
coef0	0
tol	0.015

and

Table A4. Optimized hyperparameters for the BPNN model.

Hyperparameter Name	Value
learning_rate	0.07
hidden_units	100
epochs	1000
l2_lambda	0.5
activation	logistic
optimizer	lbfgs

.

Finally, after accuracy validation, the best-performing model was selected to predict rice phenotypic parameters for all plots, making it possible to generate the predicted phenotypic data.

2.4. Construction of Rice Yield Estimation Models

2.4.1. Model Construction

The predicted rice phenotypic data from inversion models were matched with corresponding measured yield data for each growth stage, generating comprehensive stage-specific datasets.

In this study, the dataset was split into 70% training and 30% testing to ensure objective model evaluation and generalization. Subsequently, MLR and RF were applied to model linear and nonlinear relationships between rice phenotypic data and yield at each stage (TL, PI, HD, and MT).

For comparative accuracy assessment, a consistent methodology was applied to develop a yield estimation model using VIs and actual yield data.

During modeling, each stage–model combination was trained and evaluated. The optimal model from the best-performing stage was selected as the final yield estimation model, identifying both the optimal growth stage and modeling approach. The model uses remotely sensed rice phenotypic data to estimate yield at the field scale across multiple growth stages (e.g., based on imagery acquired at key phenological stages). The established model was applied to both conventional and hybrid japonica rice, with yield estimation differences between the two compared to provide a scientific basis for precision agricultural management and rice yield forecasting.

2.4.2. Model’s Performance

At each rice growth stage, inversion models for rice phenotypic data were developed using RF, XG Boost, SVR, and BPNN, while yield estimation models were established employing MLR and RF methods. The performances of the four phenotypic inversion models and the two yield estimation models were evaluated and compared based on their coefficients of determination (R²) and root mean square errors (RMSEs).

The formula used to calculate RMSE is as follows [34]:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2/\mathrm{n}}$$

The formula used to calculate R² is as follows [34]:

$${\mathrm{R}}^2=1-{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}})}^2/{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}})}^2$$

where n denotes the total number of samples in the testing set and $\overline{\mathrm{y}}$ represents the average measured yield within the testing set.

In these metrics, higher R² values indicate greater model accuracy, whereas lower RMSE values indicate higher prediction precision. In this study, both evaluation criteria were considered comprehensively to select the model exhibiting the best overall performance.

3. Results

3.1. Comparison of Phenotypic Parameters

PH, as a key agronomic trait, can effectively reflect crop growth vigor and provide an important basis for decision-making in crop layout. The determination results of PH at MT (Figure 3a) showed that among conventional japonica rice varieties, the average PH of KY131 was 67.8 cm and that of SJ22 was 95.8 cm, with an overall average PH of 81.8 cm for this type of variety. Among hybrid japonica rice varieties, the average PH of CY31 was 82.1 cm, and that of TLY619 was 99.4 cm, with an overall average PH of 90.8 cm, which was 9 cm higher than that of conventional japonica rice in total.

The trend of CC change was basically consistent across all varieties, peaking at the HD. The determination data at the HD stage (Figure 3b) showed that among conventional japonica rice varieties, the average CC of KY131 was 51.7 SPAD, the average CC of SJ22 was 42.8 SPAD, and the overall average CC of this type of variety was 47.3 SPAD; among hybrid japonica rice varieties, the average CC of CY31 was 45.9 SPAD and the average CC of TLY619 was 45.4 SPAD, with an overall average CC of 45.6 SPAD, which was 1.7 SPAD lower than that of conventional japonica rice.

During the TL, PI, HD, and MT stages, the effect of variety on LAI was significant. Except for in the PI stage, the LAI of hybrid japonica rice was higher than that of conventional japonica rice, with the greatest difference (0.25 m²/m²) observed at the MT stage (see Figure 3c for details).

Similarly, there were significant differences in AGB among different variety types at several key growth stages, and the dry matter accumulation of each variety reached its maximum and remained stable at the MT stage. The determination data at the maturity stage (Figure 3d) showed that among conventional japonica rice varieties, the average AGB of KY131 was 7.93 t/ha, the average AGB of SJ22 was 13.42 t/ha, and the overall average AGB of this type of variety was 10.67 t/ha; among hybrid japonica rice varieties, the average AGB of CY31 was 11.76 t/ha, the average AGB of TLY619 was 12.62 t/ha, and their overall average AGB was 12.19 t/ha. On average, AGB of hybrid japonica rice was 1.52 t/ha higher than that of conventional japonica rice.

3.2. Correlation Analysis Between Rice Phenotypes and Yield

Rice phenotypic parameters such as LAI, AGB, CC, and PH can directly or indirectly influence rice yield. Therefore, correlation analysis was conducted to examine the relationships between these parameters and yield.

Spearman correlation analysis revealed that AGB, LAI and PH exhibited extremely significant and positive correlations with yield, with correlation coefficients greater than 0.60—specifically, AGB (ρ = 0.784, p < 0.01) and LAI (ρ = 0.622, p < 0.01). PH values also showed an extremely significant positive correlation with yield (ρ = 0.603, p < 0.01). CC had the lowest correlation with yield (ρ = −0.041, p > 0.05), where ρ denotes the correlation coefficient and p denotes significance. Correlation analysis results are shown in Figure 4a.

Additionally, the correlation between nine vegetation indices and yield was examined, with findings illustrated in Figure 4b. The analysis revealed that among the nine vegetation indices, the Structure-Insensitive Pigment Index (SIPI) exhibited the strongest correlation with yield, achieving a correlation coefficient of 0.58. The correlation coefficients for the remaining indices were all below 0.48, with the Nitrogen Reflectance Index (NRI) demonstrating the weakest correlation at 0.375.

3.3. Inversion Models of Vegetation Indices for Rice Phenotypic Data

Using data collected from four rice growth stages, including a total of 96 sets of measurements for LAI, CC, PH, and AGB, along with their corresponding UAV multispectral data for each stage and plot, inversion models were constructed to predict rice phenotypic parameters. Four modeling techniques—RF, XG Boost, SVR, and BPNN—were employed. For each phenotypic parameter, the model demonstrating the highest accuracy, defined by the highest R² and the lowest RMSE, was selected as the final prediction model.

Table 3. Comparison of the RMSE and R² obtained by RF, XG Boost, SVR, and BPNN regression models for predicting four rice phenotypic parameters: LAI, CC, PH, and AGB.

Model Type		LAI	CC	PH	AGB
RF	RMSE	0.38 (m2/m2)	1.70 (SPAD)	7.38 (cm)	1.74 (t/ha)
	R2	0.78	0.91	0.84	0.86
XG Boost	RMSE	0.31 (m2/m2)	2.76 (SPAD)	7.52 (cm)	1.94 (t/ha)
	R2	0.83	0.84	0.81	0.82
SVR	RMSE	0.40 (m2/m2)	4.30 (SPAD)	10.32 (cm)	3.46 (t/ha)
	R2	0.75	0.54	0.67	0.61
BPNN	RMSE	0.48 (m2/m2)	2.74 (SPAD)	8.67 (cm)	1.85 (t/ha)
	R2	0.78	0.74	0.76	0.73

Both R² and RMSE in this table refer to the accuracy of the test set.

summarizes the performance metrics of these models in predicting each rice variety’s phenotypic data. Among these, XG Boost achieved the highest accuracy for predicting LAI (RMSE = 0.31 m²/m², R² = 0.83). Meanwhile, RF exhibited the best overall accuracy for estimating CC, PH, and AGB, achieving RMSE values of 1.70 SPAD, 7.38 cm, and 1.74 t/ha and respective R² values of 0.91, 0.84, and 0.86. In contrast, SVR had the lowest prediction accuracy for all four phenotypic parameters. While BPNN accuracy was generally lower than the best-performing models, it showed stable prediction results with R² values ranging from 0.74 to 0.79 and consistently low RMSE values across all parameters. More detailed accuracy comparison of each model and parameter are depicted in Figure 5, Figure 6, Figure 7 and Figure 8.

3.4. Construction of Yield Estimation Models

3.4.1. Comparison of Yield Estimation Model Accuracy

This study developed yield estimation models using two distinct approaches: (1) integrating actual yield data with predicted phenotypic parameter values across four growth stages as input variables and (2) combining actual yield data with vegetation indices (VIs) across the same four growth stages as inputs. For both approaches, Multiple Linear Regression (MLR) and Random Forest (RF) algorithms were applied for modeling, and the model demonstrating the highest accuracy was selected as the final rice yield estimation model.

Table 4. Accuracy comparison of yield estimation models: predicted phenotypic parameters vs. direct vegetation indices (VIs) using RF and MLR algorithms across four growth stages: TL, PI, HD, and MT.

Model Type		TL		PI		HD		MT
		Phenotypic Parameters	VIs	Phenotypic Parameters	VIs	Phenotypic Parameters	VIs	Phenotypic Parameters	VIs
MLR	RMSE	1.17 (t/ha)	1.35 (t/ha)	0.88 (t/ha)	1.10 (t/ha)	0.83 (t/ha)	1.44 (t/ha)	0.84 (t/ha)	0.80 (t/ha)
	R2	0.62	0.31	0.70	0.39	0.77	0.41	0.72	0.37
RF	RMSE	0.81 (t/ha)	1.24 (t/ha)	0.69 (t/ha)	0.82 (t/ha)	0.59 (t/ha)	1.06 (t/ha)	0.73 (t/ha)	1.46 (t/ha)
	R2	0.75	0.44	0.79	0.38	0.86	0.50	0.78	0.38

Both R² and RMSE in this table refer to the accuracy of the test set.

presents the performance metrics of each model for yield estimation across the four rice growth stages. Among them, the RF model achieved the highest prediction accuracy during the HD stage, with RMSE = 0.59 t/ha and R² = 0.86, followed by the PI stage, with RMSE = 0.69 t/ha and R² = 0.79. The yield estimation model for the TL stage exhibited the lowest accuracy, with RMSE = 0.81 t/ha and R² = 0.75. Overall, the MLR model consistently demonstrated lower prediction accuracy compared to the RF model, with its highest accuracy also occurring at the HD stage (RMSE = 0.83 t/ha, R² = 0.77), followed by the MT stage (RMSE = 0.84 t/ha, R² = 0.72). The lowest accuracy for MLR was observed at the MT stage (RMSE = 1.17 t/ha, R² = 0.62). Detailed accuracy comparisons for each growth stage are shown in Figure 9. And the importance ranking of each feature in the optimal prediction model during the HD period is shown in Figure 10.

3.4.2. Comparison of Yield Between Conventional and Hybrid Japonica Rice

The average yield of hybrid japonica rice varieties was 5.50 t ha⁻¹, while the average yield of conventional japonica rice was 4.38 t ha⁻¹, representing a 25.6% yield advantage for hybrid japonica rice over conventional varieties. Among the tested varieties, the hybrid japonica rice variety CY31 achieved the highest average yield at 6.00 t ha⁻¹, whereas the conventional japonica rice variety KY131 exhibited the lowest average yield at only 3.59 t ha⁻¹. Yield results are presented in Figure 11.

Based on the estimates from the highest accuracy model, the predicted yield for KY131 was 3.86 t ha⁻¹ and for SJ22 it was 4.48 t ha⁻¹, resulting in an average predicted yield of 4.17 t ha⁻¹ for conventional japonica rice varieties. The predicted yield for CY31 was 5.31 t ha⁻¹ and for TLY619 it was 5.18 t ha⁻¹, yielding an average predicted yield of 5.25 t ha⁻¹ for hybrid japonica rice varieties.

4. Discussion

4.1. Comparison Between Direct and Phased Estimation Approaches

As shown in Table 4, the indirect estimation approach, first assessing key phenotypic parameters (PH, LAI, CC, and AGB) at critical rice growth stages using spectral indices, followed by yield estimation based on these phenotypes, demonstrated higher accuracy than direct yield estimation using spectral indices alone. Moreover, evaluating phenotypic traits at key growth stages during rice cultivation provides interpretable insights [16], offering greater practical value for agronomists in field management and decision-making.

These insights may prove useful for the integration of remotely sensed LAI and leaf nitrogen accumulation with the RiceGrow model, using a particle swarm optimization algorithm, for rice grain yield assessment or the evaluation of the feasibility of yield estimation based on leaf area dynamics measurements in farmer-managed rice paddy fields.

4.2. Optimal Growth Stage for Yield Estimation

The spectral reflectance characteristics of crop canopies at different phenological stages comprehensively reflect their internal biophysical and biochemical properties throughout growth [35]. During the vegetative stage (e.g., tillering), the rapid increase in LAI and CC enhances absorption in the red band while significantly increasing near-infrared (NIR) reflectance. As the crop enters the reproductive stage (e.g., heading), the canopy closes and LAI reaches its maximum, potentially causing saturation in broadband vegetation indices. By maturity, chlorophyll degradation and water loss lead to increased red-band reflectance and pronounced changes in short-wave infrared (SWIR) reflectance [35,36].

As shown in Table 4, yield estimation using the RF method at the HD stage achieved the highest accuracy, consistent with findings reported in [19]. This can be attributed to two main factors: (1) the heading stage represents the transition from vegetative to reproductive growth in rice, directly affecting yield-related traits such as seed-setting rate and 1000-grain weight, and (2) at this stage, the canopy is fully developed, enabling UAV-acquired imagery to effectively capture the grain-filling status of panicles. In contrast, by the maturity stage, leaf senescence and environmental interference in UAV imagery reduce prediction accuracy [10].

4.3. Comparative Analysis Machine Learning Models for Yield Estimation

Among the evaluated algorithms, linear regression, ensemble learning, and neural networks, the RF algorithm outperformed the others in phenotypic assessment across rice growth stages and in yield estimation, whereas XGBoost achieved the highest accuracy in LAI estimation. Although Multiple Linear Regression provides interpretability for individual parameters, its predictive accuracy is lower than that of ensemble learning methods due to its simplicity. Neural network algorithms, despite their computational complexity, require large datasets for effective training, leading to suboptimal performance in small-sample agronomic studies [37].

Residuals for each rice variety were calculated as the differences between predicted and actual yields, with the results presented in Figure 12. Considering the RMSE of the yield estimation model (0.59 t/ha), the residuals for conventional japonica rice varieties mostly fell within the RMSE range, with only two treatments of KY131 exceeding this threshold. In contrast, for hybrid japonica rice varieties, residuals for approximately half of the treatments exceeded the RMSE range, indicating relatively larger prediction errors.

The UAV-based remote sensing RF yield estimation model developed in this study demonstrated excellent accuracy in the Northeast region. However, its capacity for spatial and temporal generalization is limited due to region-specific environmental factors. The model was trained exclusively on data from cold-region rice-growing areas in Northeast China, without accounting for other regional rice varieties. As soil temperature critically influences material accumulation during the grain-filling stage [38], this thermal dependency may restrict the model’s applicability across diverse environmental conditions. Furthermore, the model does not consider the impacts of extreme climatic events, such as heatwaves or hail, nor does it incorporate key physiological parameters, such as the photosynthetic thermal inhibition coefficient [39]. Despite the robust performance achieved through cross-validation, our machine learning model inherits common limitations of data-driven approaches. These include dependence on large training datasets, potential performance degradation when extrapolated to regions or growing seasons with contrasting environmental conditions (e.g., different climate zones or extreme weather years), and the ongoing risk of overfitting due to site-specific noise, underscoring the need for continued validation with independent datasets.

Therefore, although the current model exhibits limitations in generalization, future research could incorporate Geographically Weighted Regression (GWR) to better capture spatially varying relationships, thereby enhancing the model’s adaptability and predictive stability across diverse regions [40]. Additionally, further studies should investigate the use of individual UAV spectral bands as direct inputs for yield modeling, alongside traditional vegetation indices, to improve feature interpretability and leverage underutilized spectral information [41].

By integrating spatially explicit modeling approaches such as GWR with richer spectral feature sets, future models are expected to achieve substantially improved spatiotemporal transferability and broader applicability under varying regional conditions.

5. Conclusions

By integrating UAV multispectral data with machine learning, this study systematically explored the feasibility of early-stage rice yield estimation. Key findings include (1) a strong correlation between rice phenotypes at critical growth stages and yield, with correlation coefficients for LAI and CC exceeding 0.85; (2) high evaluation accuracy of rice phenotypic traits using multispectral data, with the RF algorithm achieving an R² > 0.9 for CC and R² > 0.8 for PH and AGB and the XGBoost algorithm yielding R² > 0.8 for LAI; and (3) optimal yield estimation performance at the HD stage, where the RF model achieved R² = 0.86 and RMSE = 0.59 t/ha, outperforming other growth stages. In summary, this study underscores the critical influence of rice developmental stage, particularly the HD stage, and model selection on the accuracy of UAV-based yield estimation, with the RF model demonstrating superior performance. Integrating the HD stage with the RF model provides a reliable, high-throughput, and non-destructive phenotyping approach for precise early yield estimation in rice breeding and field production. These findings offer essential methodological support for improving rice breeding efficiency and advancing precision agriculture practices.