Estimation of maize LCC and FVC for full growth periods based on UAV remote sensing and ensemble learning

Experiment and experimental area

The research was conducted in Xingyang City, located in Zhengzhou, Henan Province, China (Fig. 1), situated in central Henan Province near the central and lower sections of the Yellow River. The region experiences a warm temperate continental monsoon climate, with an annual average temperature of approximately 14.8 °C and an average annual precipitation of about 608.8 mm. The experimental data were collected at the breeding trial fields of Henan Jinyuan Seed Industry Co., Ltd. A total of 160 maize planting plots (10 rows × 16 columns) were selected for the study.

Data collection was conducted over 12 time points from the seedling stage to the tasselling stage of the maize breeding lines, covering the full growth periods. The specific dates for data collection were: 30 June (P1), 5 July (P2), 8 July (P3), 16 July (P4), 22 July (P5), 27 July (P6), 10 August (P7), 18 August (P8), 1 September (P9), 7 September (P10), 14 September (P11), and 21 September (P12).

UAV data collection

As the remote sensing platform, the DJI Phantom 4 multispectral UAV was employed in this study. Weighing 1,487 g, this UAV is equipped with six camera sensors, each featuring a 1/2.9″ size. The monochrome sensors capture data in five bands: red (R), green (G), blue (B), red-edge (RE), and near-infrared (NIR). Considering the endurance capacity of small UAVs and their flight safety, the DJI Phantom 4M flight height was set to 20 m.

Data acquisition with the UAV was conducted under stable solar radiation conditions on clear, cloudless days between 10:00 and 14:00. The longitudinal overlap was adjusted to 80%, the lateral overlap was fixed at 75%. The UAV followed preprogrammed flight paths and settings for automatic image data acquisition.

Field LCC and LAI measurement

Field measurements included LAI and LCC. The LAI was measured using an LAI-2200C Plant Canopy Analyzer. LCC was measured using a portable SPAD-502 sensor. It is important to note that adjacent maize planting plots were sown with the same maize inbred lines under identical growing conditions. Therefore, measurements of physiological traits were taken in just one plot from each pair. In each growth stage, data were collected from 80 maize plots. The results of the field measurements are shown in Table 1.

The LAI was transformed into FVC according to the equation presented below Eq. (1) (Nilson, 1971; Song et al., 2017):

(1) $$FVC=1-e^{-G\times \mathrm{\Omega }\times {\displaystyle \frac{LAI}{\cos (\theta )}}}.$$

In the equation, G represents the spherical orientation of the leaf projection factor, ϴ denotes the solar zenith angle, and Ω corresponds to the aggregation index (G = 0.5, ϴ = 0, Ω = 1). The observations in this research were conducted within a suitable time window (12:00—12:30), during which the variation in solar zenith angle was limited. Approximating it as a constant ensures the accuracy and consistency of the results.

In the experiment, we measured LCC and LAI values in each maize planting plot. These ground-truth measurements were then compared with the remotely estimated LCC and FVC values to ensure a comprehensive machine learning modeling approach (Bochenek et al., 2017).

Methodological framework

The methodological framework for this study is illustrated in Fig. 2. It illustrates the estimation methods for LCC and FVC parameters, as detailed below: (1) Data analysis and cleaning of field measured LCC and LAI were performed to remove errors or outliers, followed by converting LAI into FVC. Preprocessing of UAV multispectral images was conducted, including image stitching and calibration. (2) VI calculation and TF extraction were performed on the preprocessed UAV multispectral images. Correlation analysis was carried out separately for LCC and FVC, focusing on the relationship between VI and TF. (3) Different features (VI, TF, VI+TF) were input into various regression models to evaluate the estimation performance. (4) The optimal estimation strategy was selected for the visualization and mapping of LCC and FVC.

Vegetation index features

VI primarily reflect the difference in reflection between vegetation in the visible and near infrared bands and the soil background, and they can be used to quantitatively describe the growth status of vegetation (Neinavaz et al., 2021; Yue et al., 2023). In this study, five commonly used VIs were selected for estimating LCC and FVC in maize, namely the green normalized difference vegetation index (GNDVI), leaf chlorophyll index (LCI), normalized difference red edge (NDRE), normalized difference vegetation index (NDVI), and optimized soil adjusted vegetation index (OSAVI). The detailed calculation equations are provided in Table 2.

Texture features

TFs are second order features that are not directly derived from the image but are first extracted into an intermediate matrix through computation, with statistical measures defined on it. TFs described the spatial color and intensity distribution of an image or its regions. TFs can represent the texture, structure, and spatial distribution of surface objects (León-Ecay et al., 2024). In this study, eight TFs were computed using the GLCM approach. Taking into account the maize cultivation conditions and the pixel resolution of the UAV images, a 3 × 3 window size was applied for the calculation of TFs. The specific calculation formulas are shown in Table 3.

Ensemble learning and base regression models

Correlation analysis

Correlation analysis of vegetation indices with LCC and FVC

We conducted a correlation analysis using Pearson correlation coefficient to examine the relationship between maize LCC, FVC, and VI obtained from UAV remote sensing, as shown in Fig. 3. The analysis reveals that the correlations between GNDVI, LCI, NDRE, NDVI, OSAVI, and LCC are 0.770, 0.660, 0.485, 0.803, and 0.847, correspondingly. Among them, OSAVI shows the strongest correlation with LCC (0.847), while NDRE exhibits the weakest correlation (0.485). In terms of FVC, the correlations with the five VIs are 0.658, 0.601, 0.500, 0.639, and 0.498, respectively. GNDVI has the strongest correlation with FVC (0.658), while OSAVI has the weakest (0.498). It is worth noting that in this study, the correlation between LCC and NDRE is not as high as that with OSAVI. This may be because the maize growth stage of interest primarily falls within the vegetative growth phase, during which the soil background could affect the relationship between LCC and NDRE. In contrast, OSAVI optimizes soil background, resulting in a higher correlation.

Correlation analysis of texture features with LCC and FVC

Based on the VI correlation analysis results, the VI images with higher correlations to LCC and FVC were selected. A correlation analysis was then performed between OSAVI image TF and LCC and between GNDVI image TF and FVC. The analysis results are shown in Fig. 4. It can be observed that LCC has a higher correlation with Mea (0.735), while FVC has a higher correlation with Mea (0.654), ASM (0.486), Con (−0.475), and Hom (0.464).

Estimation of LCC and FVC

Estimation of LCC and FVC based on VI + TF

The maize LCC and FVC were estimated by combining UAV remote sensing VI + TF as input variables for different models. As shown in Table 6, when VI and TF are utilized as input feature in the models, the Stacking model demonstrates strongest estimation performance for both LCC (Fig. 5A, R²: 0.945, RMSE: 3.701 SPAD units, MAE: 2.968 SPAD units) and FVC (Fig. 5b, R²: 0.645, RMSE: 0.045, MAE: 0.036). Based on the estimation performance of FVC and LCC under different models (Table 6), this research selects the Stacking regression model with VI + TF used as input feature to estimate LCC and FVC.

Table 6:

Estimation performance based on the VI + TF.

Name	Model	R2	RMSE	MAE	MAPE	Name	Model	R2	RMSE	MAE	MAPE
LCC	GPR	0.923	4.387	3.478	10.737	FVC	GPR	0.577	0.049	0.039	5.057
	LR	0.912	4.430	3.511	10.831		LR	0.617	0.048	0.036	4.751
	RR	0.894	5.135	4.275	13.524		RR	0.579	0.049	0.039	5.113
	RFR	0.925	4.317	3.286	9.265		RFR	0.635	0.047	0.037	4.693
	SVR	0.920	4.389	3.576	10.978		SVR	0.560	0.050	0.040	5.116
	KNNR	0.923	4.386	3.549	10.449		KNNR	0.617	0.047	0.038	4.738
	Stacking	0.945	3.701	2.968	9.095		Stacking	0.645	0.045	0.036	4.585
	Bagging	0.905	4.858	3.566	10.158		Bagging	0.603	0.047	0.037	4.839
	Blending	0.861	5.879	4.517	12.480		Blending	0.379	0.059	0.047	6.130

DOI: 10.7717/peerj-cs.3356/table-6

Comparative analyses of the estimated and measured values of LCC (Fig. 6A) and FVC (Fig. 6B) were conducted for S1, S2, S3, and the S1-S3 periods. As shown in Fig. 6A, the estimated values of LCC closely align with the measured values overall. Similarly, Fig. 6B demonstrates that during the S1 to S3 periods, the estimated FVC values were generally consistent with the measured values.

Mapping of LCC and FVC combining VI and TF

We mapped the LCC and FVC from P1 to P12 stages using the VI + TF as input features and Stacking ensemble learning, as shown in Fig. 7. During the P1 to P4 stages, there is a considerable variation in the LCC of the maize in different planting plots, which is due to phenotypic differences in maize breeding lines. During the P6 to P7 stages, the LCC is generally higher across all plots. At P1, the FVC of all maize planting plots is relatively low, and the differences are minimal. From P2 to P3, some maize breeding lines grow faster, leading to a considerable variation in FVC across planting plots. During the P10 to P12 stages, maize breeding lines are in the reproductive growth stage, and both LCC and FVC gradually begin to decrease. Overall, both LCC and FVC show an upward trend followed by a downward trend, consistent with the field measurement results (Table 1).

The relationship between LCC and FVC throughout the growth stages

LCC and FVC are two critical parameters in crop growth research, and they exhibit distinct variation patterns during different growth stages while being interrelated. LCC and FVC are complementary indicators of crop growth. LCC reflects the nutritional status and photosynthetic capacity of the crop, while FVC is closely related to the biomass and canopy coverage. The variation trends of LCC and FVC during different growth stages can provide a more comprehensive understanding of crop growth and physiological processes.

In the early growth stage of the crop, both LCC and FVC are at relatively low levels. During this stage, crops have just begun to germinate or grow, with fewer leaves and relatively low chlorophyll content, leading to a low LCC value. At the same time, due to the smaller leaf area, the FVC value is also low. As the crop progresses into the mid growth stage, leaf expansion accelerates, photosynthesis increases, and the LCC value rises rapidly. Concurrently, the FVC value also increases, as the growth of leaves and stems increases the canopy coverage, forming a denser vegetation layer. In the maturity stage of the crop, the relationship between LCC and FVC changes. Although the number of leaves stabilizes at this stage, LCC may decrease, particularly when the crop enters the physiological senescence phase, where chlorophyll content gradually declines. Meanwhile, FVC may remain at a high level until the crop reaches the final harvesting stage, at which point some leaves may wilt, and the FVC may decrease.

The impact of VI and TF on crop parameter estimation

Based on spectral and TF combined with ensemble learning models, a study was conducted on the estimation of maize LCC and FVC. As shown in Table 6, when VI and TF were combined as input features for the model, the Stacking model achieved the best estimation performance for LCC (R²: 0.945, RMSE: 3.701 SPAD units, MAE: 2.968 SPAD units) and FVC (R²: 0.645, RMSE: 0.045, MAE: 0.036). A comparative analysis of the two input scenarios (VI and TF): when VI is used as input, the estimation of LCC is superior to when TF is used as input. This is because VI, as an indicator that represents the vegetation status on the surface, is more sensitive to changes in vegetation color, making it better suited for estimating LCC. Conversely, when TF is used as input, the estimation of FVC is better than when VI is used. This is because FVC reflects the structure and spatial distribution of crops, and TF is more sensitive to fine structural changes on the surface of vegetation, which makes it more effective for estimating FVC.

UAV remote sensing technology provides high spatiotemporal resolution multi-spectral imagery, which is essential for rapid and detailed monitoring of crop stress, growth conditions, and other agricultural parameters. Compared to traditional remote sensing methods, small agricultural UAV can easily adjust the timing and frequency of data collection, making them adaptable to research needs across different temporal and spatial scales. The VI obtained from the sensors onboard these UAV can be used to diagnose crop growth status on the surface. However, factors like soil background can affect the VI, making it difficult for spectral VI alone to fully capture the dynamic changes in crop FVC and LCC. Additionally, during the mid and late growth stages of crops, where canopy coverage is high, VI may saturate, losing sensitivity to LCC and FVC changes. This study attempts to combine spectral VI with crop canopy TF to improve the accuracy of FVC and LCC estimation. Image TF, which describe the surface properties of objects in the image or region, are highly sensitive to crop canopy structures. Combining VI and TF helps further enhance the ability of remote sensing technologies to capture changes in LCC and FVC, leading to more accurate estimates.

In this study, the low quality of LAI data may have affected the estimation accuracy of FVC. LAI is a key indicator reflecting the leaf area of crops and is commonly used to estimate biomass and coverage. LAI data can be influenced by factors such as sensor errors, variations in environmental conditions, and soil background, which lead to inaccuracies in estimation and, consequently, affect the FVC prediction. When the quality of LAI data is insufficient, it is difficult to accurately reflect the crop growth status, which in turn impacts the FVC values derived from LAI. Future research will focus on improving the quality of LAI data. In addition, FVC changes are also influenced by factors such as soil type, climate conditions, and crop management, which exhibit considerable variability in remote sensing data, increasing the complexity of FVC prediction. Although ensemble learning models have effectively improved prediction accuracy, challenges remain in addressing these diverse factors.

The sensitivity of specific spectral regions—particularly the RE, R, and NIR bands—played a critical role in capturing maize phenological dynamics. As demonstrated in our analysis, OSAVI and NDRE exhibited strong correlations with LCC. This aligns with known biophysical principles: the R band corresponds to chlorophyll absorption minima, the RE band is highly sensitive to subtle changes in chlorophyll concentration and leaf internal structure, while the NIR band responds to canopy structural properties like leaf area and density.

Notably, emerging research (Bartold & Kluczek, 2023) suggests chlorophyll fluorescence (ChF), especially solar-induced fluorescence (SIF) in the red and far-red spectrum, may provide an even more direct and sensitive indicator of photosynthetic activity and phenological transitions than LCC alone. Future work with UAVs equipped with hyperspectral or specialized fluorescence sensors could bridge this gap, enabling combined LCC-ChF monitoring for enhanced phenological tracking.

While our correlation analysis provided initial insights into feature-target relationships, the study by Bartold & Kluczek (2024) which utilized gain parameters, and research employing Mean Decrease Accuracy (MDA) (Bartold et al., 2024), rightly emphasize the broader scientific value of formal feature importance assessment within machine learning frameworks. Future studies should consider incorporating these metrics to strengthen the experimental design.

The advantages of ensemble learning estimation models

Machine learning, while powerful, relies heavily on large training datasets and complex hyperparameter tuning. Its training process typically requires substantial computational resources and time (Yang & Shami, 2020). The generalization performance of single machine learning models is closely related to the volume of data and the regularization techniques employed. Despite their strong learning capabilities, single-machine learning model can suffer from overfitting or underfitting when training data is limited or has substantial distribution shifts (Lou, Lu & Li, 2024). Compared to single machine learning models, Stacking models have lower computational complexity. The individual base learners can be trained in parallel and then combined for prediction, which makes the model more computationally efficient.

Ensemble learning generally offers good generalization capabilities. Since ensemble learning models combine the strengths of multiple individual models, they often maintain high performance across different datasets and application scenarios, reducing the risk of overfitting. Ensemble learning can also integrate various types of models, including linear and nonlinear models, making it more adaptable and flexible when handling complex problems. With its ability to improve accuracy, robustness, generalization ability, and the flexibility of model integration, ensemble learning has become an indispensable method in modern machine learning. Stacking models are particularly advantageous when resources are limited or quick model iteration is needed, making them a viable option for real-time monitoring or decision support systems in agricultural applications. By selecting the Stacking ensemble learning model, which has strong generalization capabilities, and combining VI and TF as input features, the advantages of VI and TF in different aspects of crop parameter estimation can be fully utilized, resulting in more accurate estimation outcomes.

The importance of ensemble learning models in this study is reflected in several aspects. First, by combining multiple base learners, ensemble learning mitigates the overfitting issues associated with single models, thereby enhancing prediction stability. Second, ensemble learning demonstrates greater robustness to noise and outliers, effectively improving the reliability of the results. Lastly, ensemble learning can flexibly adapt to complex agricultural data, enhancing the computational efficiency and generalization ability of the model, making it particularly suitable for real-time monitoring and rapid model iteration applications.

Although the ensemble model achieved the highest FVC estimation accuracy (R² = 0.645), this value remains modest, suggesting room for improvement. Future studies should explore incorporating structural features (e.g., canopy height distribution from LiDAR) or auxiliary data (e.g., meteorological variables) to potentially enhance prediction performance.

Limitations of UAV remote sensing

UAV data present limitations, such as sensitivity to soil background reflectance during early growth stages, relatively limited spectral resolution (often lacking continuous coverage in key regions like the red-edge) compared to dedicated airborne or satellite hyperspectral sensors, and practical constraints on area coverage compared to satellite data like Copernicus Sentinel-2. As demonstrated by Gurdak & Bartold (2021) vegetation indices derived from Copernicus satellite data for maize and sugar beet effectively tracked changes associated with plant aging, such as the SIPI index. Despite these limitations, UAVs excel at high-frequency, detailed field-scale monitoring, which is crucial for precision breeding and management applications.

Finally, we explicitly acknowledge a key limitation inherent to UAV-based optical monitoring: significant sensitivity to soil background noise, particularly during early growth stages when crop canopy coverage is sparse. Soil reflectance variations can substantially confound spectral signals from vegetation, impacting VI accuracy for both LCC and FVC estimations. While the soil-adjusted OSAVI employed in this study partially mitigates this issue, more robust solutions are needed (Dąbrowska-Zielińska et al., 2012).

SAR systems are uniquely advantageous due to their all-weather operational capability and sensitivity to soil properties like moisture content and surface roughness. Integrating SAR-derived soil moisture and roughness parameters with UAV-acquired spectral and textural features offers a multi-sensor approach to decouple the influence of soil background from crop canopy signals. This synergistic data fusion holds significant potential for not only minimizing early-stage estimation errors caused by soil noise but also enabling the retrieval of complementary biophysical parameters relevant to maize growth assessment. Future investigations will explore this integrated UAV-SAR framework.

A critical limitation of UAV optical remote sensing is its susceptibility to soil background noise, particularly during early growth stages when vegetation cover is sparse. SAR systems uniquely penetrate cloud cover and are insensitive to solar illumination conditions. Crucially, SAR backscatter coefficients directly respond to soil biophysical properties—including surface roughness and moisture content—which are key parameters influencing crop growth but inaccessible to optical sensors Soil spectral variations can significantly interfere with vegetation signals, reducing the accuracy of VI-based estimations for both LCC and FVC.