Next Article in Journal
Field Study of UAV Variable-Rate Spraying Method for Orchards Based on Canopy Volume
Previous Article in Journal
Enhanced Spring Wheat Soil Plant Analysis Development (SPAD) Estimation in Hetao Irrigation District: Integrating Leaf Area Index (LAI) Under Variable Irrigation Conditions
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Agriculture
Volume 15
Issue 13
10.3390/agriculture15131373
agriculture-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Winter Wheat Nitrogen Content Prediction and Transferability of Models Based on UAV Image Features

by
Jing Zhang
1,
Gong Cheng
2,*,
Shaohui Huang
1,
Junfang Yang
1,
Yunma Yang
1,
Suli Xing
1,
Jingxia Wang
1,
Huimin Yang
1,
Haoliang Nie
1,
Wenfang Yang
1,
Kang Yu
3 and
Liangliang Jia
1,*
1
Hebei Key Laboratory of Soil Fertility Improvement and Agricultural Green Development, Institute of Agricultural Resources and Environment, Hebei Academy of Agriculture and Forestry Sciences, No. 598 West Heping Road, Shijiazhuang 050051, China
2
Key Laboratory of Agricultural Water Resources, Hebei Laboratory of Water-Saving Agriculture, Center for Agricultural Resources Research, Institute of Genetic and Developmental Biology, The Chinese Academy of Sciences, Shijiazhuang 050021, China
3
Precision Agriculture Laboratory, School of Life Sciences, Technical University of Munich, Dürnast 9, 85354 Freising, Germany
*
Authors to whom correspondence should be addressed.
Agriculture 2025, 15(13), 1373; https://doi.org/10.3390/agriculture15131373
Submission received: 6 May 2025 / Revised: 24 June 2025 / Accepted: 25 June 2025 / Published: 26 June 2025
(This article belongs to the Section Digital Agriculture)
Browse Figures
Versions Notes

Abstract

Accurate and timely monitoring of plant nitrogen content (PNC) is essential for precision agriculture (PA) and food security. While multispectral unmanned aerial vehicle (UAV) imagery has shown promise in PNC estimation, the optimal feature combination methods of spectral and texture features remain underexplored, and model transferability across different agricultural practices is poorly understood. This study aims to present an innovative approach by integrating 40 texture features and 22 spectral features from UAV multispectral images with machine learning (ML) methods (RF, SVR, and XGBoost) for winter wheat nitrogen content prediction. In addition, through analysis of an 8-year long-term field experiment with rigorous data, the results indicated that (1) the RF and XGboost models incorporating both spectral and texture features achieved good prediction accuracy, with R2 values of 0.98 and 0.99, respectively, RMSE values of 0.10 and 0.07, and MAE values of 0.07and 0.05; (2) models trained on Farmers’ Practice (FP) data showed superior transferability to Ecological Intensification (EI) conditions (R2 = 0.98, RMSE = 0.08, and MAE = 0.05 for XGBoost), while EI-trained models performed less well when applied to FP conditions (R2 = 0.89, RMSE = 0.45, and MAE = 0.35 for XGBoost). These findings established an effective framework for UAV-based PNC monitoring, demonstrating that fused spectral–textural features with FP-trained XGboost can achieve both high accuracy and practical transferability, offering valuable decision-support tools for precision nitrogen management in different farming systems.

1. Introduction

Wheat, the world’s most widely cultivated cereal crop, plays an essential role in global food security. As a leading wheat producer, China’s production significantly influences worldwide supply [1]. Accurate monitoring of plant nitrogen content (PNC) is crucial for wheat production. Precise PNC assessment enables optimized nitrogen application, ensuring food security while promoting sustainable agriculture [2].
However, the traditional method for PNC assessment relies on destructive sampling, incurring high costs and labor intensity [3]. While traditional methods face these limitations, advances in remote sensing now enable the non-destructive, real-time monitoring of crop nitrogen status through rapid spectral data acquisition [4]. Many scholars have conducted extensive research on the rapid monitoring and accurate prediction of crop nitrogen content. Kaur et al. studied 2145 winter wheat (Triticum aestivum L.) genotypes using hyperspectral imaging and predicted phenotypes with partial least squares regression (PLSR) and random forest (RF) regression [5]. Ma et al. used UAV hyperspectral imaging to predict wheat nitrogen content, enabling precise nitrogen management [6]. Jia et al. used hyperspectral imaging to derive vegetation indices for assessing nitrogen status and yield in spring maize, providing stage-specific fertilization guidance [7]. Also, multispectral and hyperspectral images have also been successfully used to predict grain yield [8], vegetation indices, and biomass [9,10]. Hyperspectral-based nitrogen prediction has boosted the use of low-cost UAV multispectral data for rapid crop monitoring.
Although multispectral vegetation indices are widely used, their accuracy is often compromised by soil interference and canopy saturation [11]. To address this, recent studies have explored texture features as complementary indicators which capture spatial canopy patterns and enhance nitrogen signal sensitivity [11,12]. Zheng et al. found that incorporating UAV multispectral texture features improved rice nitrogen prediction by reducing background interference and canopy saturation effects [11]. Zhang et al. showed that RGB image texture features enhanced winter wheat nitrogen prediction across all growth stages [12]. Despite the cost benefits, RGB images’ limited spectral range (red, green, blue) reduces their sensitivity to nitrogen-related features, unlike near-infrared and red-edge bands used in key vegetation indices like NDVI [8,13]. Vegetation indices based on RGB images, such as visible light atmospheric resistance index (VARI) and excess green index (EXG), are sensitive to saturation under high-density canopy or high-biomass conditions, thus reducing their predictive accuracy in these complex environments [14,15]. In addition, texture features have also been used to predict the PNC of potatoes [16], cotton [17], corn [18], and other crops [19,20]. Nevertheless, developing nitrogen prediction models that maintain both high accuracy and wide applicability is complex, and the most effective way to combine spectral and texture features for this goal has yet to be determined.
The integration of spectral and texture features requires advanced analytical tools [14]. Here, machine learning (ML) algorithms demonstrate unique advantages by handling high-dimensional data and nonlinear relationships [19,21]. Hyperspectral-based N prediction shows that combining vegetation indices with ML algorithms (e.g., RF, PLSR) improves accuracy over linear regression alone [5]. An RF model outperformed support vector regression (SVR) and PLSR in predicting rice canopy nitrogen content from UAV images [21]. Ma et al. proposed an effective method with deep learning and the N-based PROSAIL model to predict the plant nitrogen of wheat [6]. The results from the above studies showed that ML methods often performed better than traditional models in terms of accuracy, while ensemble learning methods often performed better than general ML methods.
Yet, transferability is an important evaluation parameter of these models, referring to their applicability in different environments or crop types. Thompson and Puntel developed a UAV-based multispectral sensing system for maize nitrogen management, achieving robust predictive accuracy while demonstrating consistent performance across diverse soil types and climatic zones [22]. Their framework integrated vegetation indices with machine learning algorithms to optimize nitrogen recommendation accuracy under varying agronomic conditions. And subsequent research in 2023 further established that UAV-derived remote sensing data could effectively support the development of crop nitrogen status prediction models, with demonstrated cross-crop and cross-regional transferability in agricultural applications [23]. These findings highlight the growing potential of drone-based remote sensing for cross-regional and cross-crop nitrogen management, yet they also reveal the need for more robust feature integration to enhance model performance and adaptability.
Based on these gaps, we hypothesize that a fusion of spectral–textural features with ML can simultaneously improve PNC estimation accuracy and cross-practice transferability. This study took winter wheat as its research object, set up two fertilization recommendations on EI (Ecological Intensification) and FP (Famers’ Practice), obtained critical growth-period UAV multispectral images (including blue (B), green (G), red (R), red-edge (Rededge), and near-infrared (NIR)), and extracted the spectral and texture features of the UAV images, with the goals of (1) identifying the optimal spectral and texture features from UAV multispectral imagery for winter wheat PNC prediction through correlation analysis; (2) constructing and comparing ML models (RF, SVR, XGBoost) to evaluate the performance of fused spectral–textural features in improving PNC estimation accuracy; (3) rigorously evaluating and analyzing model transferability between contrasting agricultural practices (EI and FP) to ensure practical applicability.

2. Materials and Methods

2.1. Study Site

The study area was located at the Comprehensive Experimental Station of Hebei Academy of Agriculture and Forestry, Dahe Town, Luquan District, Shijiazhuang, Hebei Province (38°07′ N, 114°29′ E, WGS84). This area belongs to the semi-humid continental monsoon climate. The average annual precipitation is 300–600 mm, with approximately 70–80% occurring from June to September; the mean annual air temperature is 14.3 °C; the frost-free period is 198 days; and the volume water content of wheat fields the during growth period is 18–22%. The soil type of the experimental site was fluvo-aquic soil. The main nutrient contents of the 0–20 cm soil layer before the experiment at the study area were as follows: pH, 7.1; organic matter content, 16.4 g/kg; total N, 1.14 g/kg; available N, 27.9 mg/kg; available P, 13.6 mg/kg; and available K, 96.6 mg/kg. The precipitation in the area is mostly concentrated from June to September, and this accounts for 70–80% of the annual precipitation; the average annual precipitation is 300–600 mm. The average annual air temperature is 14.3 °C, with 198 frost-free days in this region [24].

2.2. Field Experiment Design

One of the main crops in the research area was winter wheat, which is sown in mid-October and harvested in early June of the following year. The long-term positioning experiment we used in this study was set up in June 2009. It was established as a split-plot design, with two different management modes as the main treatments, which were EI and FP. EI is based on the recommendation of nutrient management by the Nutrient Expert system, optimizing fertilizer applications, management measurements, planting with a new variety (Chunmiao 618), and other agronomic measures to form a comprehensive management model. The Nutrient Expert system can help develop fertilizer recommendations for specific plots and adopt 4R nutrient management strategies to meet the nutrient needs of crops at different stages. EI treats 1/3 of nitrogen fertilizer and all phosphorus and potassium fertilizers as base fertilizer, and the remaining 2/3 of the nitrogen fertilizer as topdressing at the jointing period; FP is managed similarly to local farmers and the variety grown is Xingmai 6. The ratio of nitrogen fertilizer base to topdressing in FP treatment is 1:1. The side treatment was designed as including three different nitrogen fertilizer application strategies (with 3-year cycles), named 0 N (no nitrogen fertilizer applied), 2/3N (no nitrogen fertilizer applied in the first of the 3 years, followed by 2 years of nitrogen fertilizer application), and 3/3N (nitrogen fertilizer applied). Except for the different nitrogen fertilizer applications, the application rates of phosphorus and potassium fertilizers, as well as other agronomic measures, for each side treatment were the same. The fertilization types, amounts, and methods were the same as in Huang et al. [24]. Each treatment was repeated four times, for a total of 24 plots, all of which had the same area of 9 × 5 m2 (Figure 1).

2.3. Data Acquisition and Processing

2.3.1. Multispectral Images from UAV

An M300 quad-copter UAV remote sensing platform was equipped with a K6 multispectral imager to obtain multispectral canopy images of winter wheat during the returning green stage, jointing stage, and grouting period in 2023. Specifically, data was collected on 15 April and subsequently collected at 7-day intervals. The UAV multispectral image acquisition parameters are shown in Table 1. The preprocessing of images mainly included image format conversion, image filtering, image stitching, orthorectified correction, and radiometric calibration [25].

2.3.2. Ground Data

Ground data acquisition was synchronized with UAV multispectral image acquisition. Specifically, during the returning green stage, jointing stage, and grouting period of winter wheat, a uniform growing area was selected for each plot, with 2 rows (0.15 m) × 1 m fixed. Ten single-stem samples were taken and placed in sealed bags. The samples were separated into leaves, stems, and spikes in the laboratory and placed in respective paper bags. The samples were first heated at 105 °C to inactivate enzymatic activity and then dried at 80 °C to constant weight. After organ crushing, nitrogen content was determined by the Kjeldahl nitrogen determination method [24].

2.4. Research Method

2.4.1. Feature Extraction

After preprocessing through methods such as radiometric correction, the UAV remote sensing images were subjected to spectral and texture feature extraction. In this study, the radiometric calibration employed empirical line correction using spectral on reference panels to establish linear relationships between image DN values and surface reflectance, with validation confirming high accuracy (R2 > 0.95) across all spectral bands. This standardized method ensured physically meaningful reflectance data while eliminating sensor-specific radiometric distortions. Spectral characteristic data included reflectance data from five bands—blue (B), green (G), red (R), red-edge (Rededge), and near-infrared (NIR)—as well as 22 different vegetation indices, including the Green-Wave-Normalized Vegetation Index (NGBDI) and the Green-Wave-Optimized Soil-Regulated Vegetation Index (GOSAVI), which were obtained through band combination calculations. The texture features included 8 features corresponding to each of the 5 bands, namely contrast (con), second moment (sm), variance (var), mean (mean), correlation (cor), dissimilarity (dis), homogeneity (hom), and entropy (ent). In this study, spectral data were extracted using ENVI 5.6, and we separated each plot using the Region of Interest (ROI) tool in ENVI 5.6 to correlate with plant nitrogen content. Texture features were derived from ENVI, while vegetation indices were calculated using R 4.2.2.
(1)
Vegetation indices
The vegetation indices had a simple structure and certain mechanistic properties which could reduce the impact of soil and other factors on the vegetation spectra. Therefore, they have been widely used in the qualitative and quantitative evaluation of vegetation cover and its growth trends [15,26]. Crops exhibit obvious epigenetic features such as reduced coverage, reduced leaf area, and the yellowing of leaves when they are deficient in nitrogen. These features can provide a basis for predicting PNC using vegetation indices. In total, 22 commonly used vegetation indices were selected in this study, and their related formulas are shown in Table 2.
(2)
Texture features
Texture features are inherent attributes of images, containing important information about the organization and arrangement of object surface structures and their relationships with the surrounding environment. They have the advantages of rotation invariance and strong resistance to noise [15,39,40]. The main methods for extracting texture features are as follows: statistical methods (grayscale co-occurrence matrix, texture spectrum, geometry), model methods (random field model, classification model), signal processing methods, and structural analysis methods [41]. The gray-level co-occurrence matrix method is currently recognized by the academic community as an image recognition technology with strong robustness and adaptability which can efficiently achieve image classification and retrieval and maximize the improvement of classification processing accuracy. It is widely used in texture feature extraction from remote sensing images. The information reflected by the texture features of each band is different, and in this study, we used the gray-level co-occurrence matrix method to extract the texture features of five bands in the multispectral images, resulting in a total of 40 types of data.

2.4.2. Data Analysis

To identify sensitive features for winter wheat plant nitrogen content (PNC) prediction, Pearson correlation analysis was performed on 22 spectral features and 40 texture features with measured PNC. The PNC prediction model was developed using only those features that demonstrated statistically significant correlations at α = 0.01. The ML models (RF, SVR, and XGBoost) were implemented using standardized parameters (Table 3) to ensure consistent comparative analysis of spectral–texture feature fusion. While hyperparameter tuning could yield marginal improvements, this study prioritized computational efficiency and reproducibility given the study’s focus on feature fusion evaluation. The training time of each model was also recorded.
(1)
RF
RF employs bootstrap sampling, which involves randomly selecting data subsets with replacement, to construct multiple decision trees. As a bagging-based ensemble learning method, RF combines these trees to form a supervised machine learning model. By introducing randomness during decision tree training, the method enhances model robustness against overfitting and noise interference. The RF algorithm also supports parallel training, improving computational efficiency while providing feature importance metrics for model interpretation [42,43]. For PNC modeling, each bootstrap sample generates a decision tree with unique splitting rules. The ensemble of these trees forms the final regression model. Tree depth, determined by node splitting, influences model complexity. To control overfitting, key RF parameters were optimized to maintain balanced model performance.
(2)
SVR
SVR projects data into high-dimensional space to find optimal prediction boundaries. The theoretical basis of SVR is convex quadratic programming, which determines that its final result is the global optimum. It utilizes nonlinear mapping to project data into a high-dimensional space, facilitating linear regression, and then maps the results back to the original space. At the same time, by introducing kernel functions, the inner product operation in high-dimensional space can be well solved [44]. The commonly used kernel functions include linear, polynomial, radial basis, sigmoid, Fourier, etc. Among them, linear kernel functions have the advantages of high efficiency and wide application range. In this study, the spectral and texture features had linear separability and linear kernel functions could meet the requirements and improve efficiency.
(3)
XGBoost
XGBoost utilizes gradient boosting, an ensemble technique that sequentially builds decision trees to correct prediction errors from previous models through weighted aggregation. It is an ensemble learning algorithm based on the Gradient Boosting Decision Tree (GBDT) algorithm, which is an improvement on the gradient boosting algorithm. XGboost provides parallel tree boosting, which can quickly and accurately solve many problems in data science. Compared to similar boosting tree methods, XGBoost demonstrates superior computational efficiency, higher prediction accuracy, and stronger generalization performance. The basic principle of XGboost is using decision trees as the basic learners. Each decision tree is a binary tree that grows through feature splitting, with the aim of fitting the predicted residuals of the sample. A new tree is added after each iteration of XGboost, and the residuals of the input and prediction results of the new tree are directly related to the actual values of the previous iteration. Then, new iterations are started until the specified number of iterations or threshold is reached. The final prediction model is obtained by adding up the prediction results of all the decision trees [45].
It should be noted that the initial model evaluation was trained using a mixture of the FP and EI datasets, corresponding to result Section 3.2. The transferability evaluation of the models was conducted using treatment-specific models, corresponding to result Section 3.3.

2.4.3. Model Evaluation

The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) were used to evaluate the predictive performance of the machine learning models to predict nitrogen content in winter wheat plants. The specific formulas are as follows:
R M S E = 1 n ∑ i = 1 n ( y i − y i ^ ) 2
M A E = 1 n ∑ i = 1 n y i − y i ^
R 2 = 1 − ∑ i = 1 n ( y i − y i ^ ) 2 ∑ i = 1 n ( y i − y i ¯ ) 2
where n is the number of samples, y i is the observation value, y i ^ is the predicted value, and y i ¯ is the average of the observation value.

2.4.4. Model Transferability Evaluation

The model transferability is the model’s capacity to maintain prediction accuracy when applied to alternate management systems (FP ↔ EI). And the model transferability in this study was systematically evaluated using a cross-validation approach. Specifically, we conducted bidirectional validation by (1) training models on EI data and testing on FP datasets, and (2) vice versa. This rigorous validation framework enabled the quantitative assessment of model generalizability across different management systems. Transferability performance was comprehensively evaluated using three established metrics: R2 to measure prediction accuracy, RMSE to assess prediction deviations, and MAE to evaluate absolute prediction errors. Through this comparative analysis, we were able to determine which training dataset (EI or FP) yielded models with superior transfer capability, providing critical insights for practical agricultural applications.

3. Results

3.1. Sensitivity Analysis of Plant Nitrogen Content

This study initially derived 62 image features, comprising 22 vegetation indices and 40 texture features, from five spectral bands of canopy reflectance. Through correlation analysis, we identified and selected the features most strongly associated with winter wheat nitrogen content for subsequent modeling (Figure 2 and Figure 3). The highest correlation coefficient between spectral characteristics and PNC occurs for RERDVI and VREI, at 0.85. The predictive capability likely stems from the inclusion of red-edge and near-infrared (NIR) spectral bands, which exhibit strong sensitivity to both chlorophyll concentration and canopy nitrogen levels [13]. The highest correlation coefficient between texture features and PNC was found for mean-nir at 0.84. This reflects the ability to capture structural variations in canopy density that correlate with nitrogen uptake efficiency. The correlation coefficients between PNC and image features were tested at a highly significant level of 0.01, including 18 vegetation indices, excluding VARI, EVI, MTVI, and CARI, and 35 texture features, excluding hom_green, con_red, dis_red, hom_red, and var_red. By using this approach, this study used 53 spectral and texture features to preserve the sensitivity characteristics of PNC.

3.2. Model’s Performance

Three machine learning models (RF, SVR, and XGBoost) were developed to predict PNC, utilizing 53 significantly correlated spectral and texture features. The relationship between the measured and predicted values of the test data, as well as the evaluation indicators of the model, are shown in Figure 4. There are differences in the predictive effects of the models constructed by different methods on nitrogen content of winter wheat. It was shown that the measured and predicted values of RF and XGboost model were mostly concentrated near the 1:1 straight line. The R2, RMSE, and MAE of the prediction models constructed by different ML methods are different. The prediction model constructed with the XGboost method has the highest R2 and lowest RMSE and MAE, with values of 0.99, 0.07, and 0.05, respectively. The prediction model constructed by the RF method has a moderate R2, RMSE and MAE, with values of 0.98, 0.10, and 0.07, respectively. For SVR, its R2 is the lowest and its RMSE and MAE are the highest, with values of 0.96, 0.07, and 0.05, respectively. The results showed that the winter wheat PNC prediction model constructed with XGboost methods performed the best. XGBoost’s superior performance over RF and SVR can be explained by its built-in regularization mechanisms and boosted tree architecture, which effectively handles the high-dimensional feature space while preventing overfitting. The relatively lower performance of SVR may stem from its linear kernel’s limited capacity to capture nonlinear relationships between spectral–textural features and PNC.

3.3. Transferability of Nitrogen Content Prediction Models for Winter Wheat

Based on RF, SVR, and XGboost methods, PNC prediction models were established using the measured data of EI and FP as training sets, respectively, and the PNC of the wheat treated with FP and EI was predicted. The performances of the three models in the prediction of nitrogen content in wheat treated with EI and FP were similar to those shown in Section 3.2 in this study. The relationships between the measured and predicted values are shown in Figure 5. The R2 values of the EI-trained models constructed with RF, SVR, and XGboost were 0.86, 0.75, and 0.89, respectively, when predicting PNC transferability in winter wheat treated with FP (Figure 5e). On the contrary, the R2 values of the FP-trained models constructed with RF, SVR, and XGboost were 0.84, 0.81, and 0.98, respectively, when predicting PNC transferability in winter wheat treated with EI. The RMSE and MAE values of the transferability prediction models are higher with RF than in the SVR and XGboost models (Figure 5f). Based on R2, RMSE, and MAE, the migration prediction ability of the XGboost model was better than others. We found asymmetric transferability patterns, where the FP-trained models performed better on EI data (XGBoost R2 = 0.99) than vice versa (R2 = 0.83). This likely reflects the broader nitrogen response spectrum captured under conventional farming conditions. The FP plots exhibited greater variability in nitrogen status, enabling the models to learn more robust feature–response relationships adaptable to different management systems.

4. Discussion

4.1. Driving Factors for Differences in Model Performance

ML is an intelligent data analysis method which achieves accurate prediction by fully mining the information in the model construction dataset for model construction. It has gained increasing attention in predictive research across various fields. This study focused on exploring the effects of three ML methods, RF, SVR, and XGboost, on predicting nitrogen content in winter wheat. And there are differences in the prediction performance and learning efficiency of the PNC models of winter wheat constructed with the three ML methods.
Firstly, the computational efficiency of these three models is different. This study conducted statistical analysis on the training time of each model under the same conditions of data segmentation, shuffling methods, and cross-validation. There were significant differences in the training time of the models, with the SVR model taking the shortest time of 1.8 s, the XGboost model taking the longest time of 2.4 s, and the RF model taking 2.1 s. This result is similar to those of Guo et al. [1], Lin and Liu [45], Du et al. [46], Panahi et al. [44], Fernández-Habas et al. [42], and Jeung et al. [43], but different from that of Yin [47].
Another driving factor behind the different performances of the models is their different algorithm characteristics. The principles, data requirements, and model generalization ability of the three methods are different, resulting in differences in their prediction results. SVR, as a supervised learning algorithm, is mainly used for regression analysis. SVR may encounter computational bottlenecks when processing high-dimensional data as it needs to solve a quadratic programming problem, which can become very time-consuming when dealing with large amounts of data. This indicates that the different data sizes caused the difference in SVR time consumption between this study and the research results of Yin [47]. On the other hand, the SVR method is useful for solving multidimensional problems, with strong generalization ability and a relatively low dependence on data. However, it is difficult to select the appropriate kernel function, and the accuracy of the model is easily affected. The kernel function selected in this study was a linear kernel function, which improved the running speed but may have reduced the accuracy of the model.
Both the RF and XGboost methods extract and recombine large subsets from the original dataset based on the bootstrap strategy. The parameters required for the two methods are relatively simple and the computation speed is faster. On the other hand, the great performance of RF and XGboost depends on the fact that the two methods are ensemble learning approaches which combine several learners based on the ideas of bagging and boosting, respectively. They combine several learners to create a new one, achieving better learning results, and the two methods fully embody the “collective intelligence” of ML. When dealing with complex and nonlinear data, the XGboost model learns columns and blocks in parallel, sorts the features, saves them in a block structure, and stores them in a sparse matrix format, greatly reducing the computational load. In addition, XGboost can implement distributed or multithreaded computation because it can simultaneously perform shard searches for different features when looking for the best split point, realizing the parallelization of features. In addition, XGboost also supports missing value processing for each non-leaf node in the tree. If a feature value of a sample is missing, XGboost can automatically learn its default splitting direction and classify it into the default branch, further enhancing the robustness and generalization ability of the model. These features make XGboost more efficient at learning large amounts of data [48].
RF models construct different training sets by random row sampling and column sampling, obtaining the final prediction result by the weighted average method or majority voting method. RF is able to learn in parallel, and it has a good filtering effect on noise and abnormal data, which is widely used in various datasets. Although RF has high accuracy and stability, its learning efficiency may not be as good as XGboost when processing large-scale data as it relies on the voting or averaging of multiple decision trees, resulting in relatively high computational complexity [47]. Therefore, the PNC models constructed with RF and XGboost methods had good predictive performances and migration abilities [48].

4.2. Agricultural Explanation of Model Transferability

ML methods have been widely applied into many fields such as science, engineering, agriculture, and medicine. Supervised ML especially has greatly improved the accuracy of predictions [49]. But the transferability of these models is a crucial indicator of whether these models are reliable. The differences in the statistical values for the training and test models are shown in Figure 6. The asymmetry in model transferability between FP and EI may be caused by several reasons, including data variability, differing canopy structures, and environmental conditions. Firstly, there are two varieties of winter wheat with different genotypes, which results in different datasets for training and testing, leading to different prediction effects and model transferability in this study. The point-to-point graph between the measured and predicted values of the test data showed a high degree of local fitting for the three models (Figure 5). This may be caused by the values of PNC in this study, which were mainly concentrated in the range of 1.1 to 2.3 g kg−1, meaning that the models performed well within this range. The difference between the training dataset and the testing dataset, as well as the good performance of local fitting, are closely related to the canopy image information. In particular, when compared with different ML methods, the transferability of the RF and XGboost methods were more prominent. Jiang [50] evaluated various ML methods, and their results showed that the RF and Adaboost methods performed well in estimating the total nitrogen content of Miyun Reservoir. Studies on predicting PNC, dry weight, leaf area index, soil texture, and soil nitrogen content by ML methods also showed that RF and ensemble learning methods performed well [49,51,52].
Secondly, texture features represent the detailed information and organizational structure of an image. They not only contain the structural arrangement information of the object surface, but also reflect the spatial distribution pattern of grayscale information of pixels in the image. Texture features of spectral images can represent the canopy structure, especially under dense canopy coverage. Texture features have special significance for improving the sensitivity of remote sensing data to biophysical properties [11]. Therefore, models trained with treatments exhibiting different canopy characteristics also demonstrate varying degrees of transferability.
In addition, management measures also have a significant impact on the stability of features. Protective fertilization is very important, and effective nitrogen monitoring can ensure stable crop production while implementing protective fertilization. The method of drone monitoring is accurate and efficient, and the migration ability of the FP-trained model is significantly higher than that of the EI-trained model. This indicates that the model constructed with traditional fertilization as the training set is more suitable for actual use.

4.3. Limits and Future Research

While this study demonstrates the effectiveness of UAV-based multispectral imaging for plant nitrogen content prediction, several practical implementation challenges merit consideration. First, the computational demands of image processing and feature extraction, particularly for texture analysis, may limit real-time field applications without adequate computing infrastructure. Second, data quality can be compromised by variable field conditions, including changing illumination angles (particularly during morning/afternoon flights), sensor calibration drift (±2–3% reflectance variation) [53], and atmospheric interference that disproportionately affects NIR bands. Third, while the models performed well under controlled experimental conditions, their broader applicability would require validation across more diverse agroecological zones and farming systems. These practical constraints highlight important trade-offs between analytical precision and field applicability that should guide future research toward more computationally efficient algorithms and cost-effective implementation strategies.
In subsequent work, it is necessary to focus more on integrating transfer learning techniques such as domain adaptation and expanding the datasets to validate the performance of the model in heterogeneous environments. In addition, this prediction system can be combined with existing agricultural machinery to achieve variable nitrogen application; this technology is closely related to smart agriculture and can provide real-time nitrogen supplementation based on the growth period of wheat. It can also achieve the precise management of nitrogen fertilizer application in large fields.

5. Conclusions

This study identified 53 highly significant spectral and texture features (p < 0.01) for PNC based on five-band multispectral data. The XGboost model performed the best among the three tested ML models, and NGBDI, RDVI, RENDVI, mean3, and VREI were among the top five in the variable importance ranking of the model. Notably, the FP-trained models showed promising transferability to EI conditions (R2 = 0.81–0.99 across models), suggesting their potential utility in operational settings. While these results provide valuable insights for winter wheat nitrogen monitoring, further validation under diverse field conditions would be necessary before widespread implementation. The methodology developed in this study offers a foundation for future research, which could productively explore deep learning methods, additional spectral bands to further enhance PNC prediction accuracy, and the field validation of variable-rate fertilization protocols based on these predictions.

Author Contributions

Writing—original draft, J.Z.; writing—review and editing, S.H., J.Y., Y.Y., S.X., J.W., H.Y., H.N., W.Y. and K.Y.; visualization, G.C.; supervision, L.J. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by Hebei Natural Science Foundation (C2024301124), the Basic Research Funds of Hebei Academy of Agriculture and Forestry Sciences (2024130202), and the HAAFS Science and Technology Innovation Special Project (2022KJCXZX-ZHS-5, 2022KJCXZX-ZHS-6). The APC was funded by the HAAFS Science and Technology Innovation Special Project (2022KJCXZX-ZHS-5, 2022KJCXZX-ZHS-6).

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The data presented in this study are available on request from the corresponding author.

Acknowledgments

We thank all the foundation, and we thank Tongtong Zhao for the help on the graph. Finally, we appreciate the constructive suggestions from the reviewers.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Guo, Y.; Jing, Y.H.; Gang, W.L.; Yi, H.J.; Jia, H.; Wei, F.; Qing, Z.G. UAV multispectral image-based nitrogen content prediction and the transferability analysis of the models in winter wheat plant. Sci. Agric. Sin. 2023, 56, 850–865. [Google Scholar] [CrossRef]
  2. Yue, J.; Yang, H.; Yang, G.; Fu, Y.; Wang, H.; Zhou, C. Estimating vertically growing crop above-ground biomass based on UAV remote sensing. Comput. Electron. Agric. 2023, 205, 107627. [Google Scholar] [CrossRef]
  3. Verma, B.; Prasad, R.; Srivastava, P.K.; Singh, P.; Badola, A.; Sharma, J. Evaluation of simulated AVIRIS-NG imagery using a spectral reconstruction method for the retrieval of leaf chlorophyll content. Remote Sens. 2022, 14, 3560. [Google Scholar] [CrossRef]
  4. Garofalo, S.P.; Modugno, A.F.; De Carolis, G.; Sanitate, N.; Negash Tesemma, M.; Scarascia-Mugnozza, G.; Tekle Tegegne, Y.; Campi, P. Explainable Artificial Intelligence to Predict the Water Status of Cotton (Gossypium hirsutum L., 1763) from Sentinel-2 Images in the Mediterranean Area. Plants 2024, 13, 3325. [Google Scholar] [CrossRef]
  5. Kaur, S.; Kakani, V.G.; Carver, B.; Jarquin, D.; Singh, A. Hyperspectral imaging combined with machine learning for high-throughput phenotyping in winter wheat. Plant Phenome J. 2024, 7, e20111. [Google Scholar] [CrossRef]
  6. Ma, X.; Chen, P.; Jin, X. Predicting wheat leaf nitrogen content by combining deep multitask learning and a mechanistic model using UAV hyperspectral images. Remote Sens. 2022, 14, 6334. [Google Scholar] [CrossRef]
  7. Jia, Z.; Zhao, S.; Zhang, Q.; Zhang, X.; Zhang, Y.; Gao, Q. Multi-stage fertilizer recommendation for spring maize at the field scale based on narrowband vegetation indices. Comput. Electron. Agric. 2023, 213, 108236. [Google Scholar] [CrossRef]
  8. Fu, Y.; Yang, G.; Pu, R.; Li, Z.; Li, H.; Xu, X.; Song, X.; Yang, X.; Zhao, C. An overview of crop nitrogen status assessment using hyperspectral remote sensing: Current status and perspectives. Eur. J. Agron. 2021, 124, 126241. [Google Scholar] [CrossRef]
  9. Qiao, L.; Tang, W.; Gao, D.; Zhao, R.; An, L.; Li, M.; Sun, H.; Song, D. UAV-based chlorophyll content estimation by evaluating vegetation index responses under different crop coverages. Comput. Electron. Agric. 2022, 196, 106775. [Google Scholar] [CrossRef]
  10. Zhang, Y.; Xia, C.; Zhang, X.; Cheng, X.; Feng, G.; Wang, Y.; Gao, Q. Estimating the maize biomass by crop height and narrowband vegetation indices derived from UAV-based hyperspectral images. Ecol. Indic. 2021, 129, 107985. [Google Scholar] [CrossRef]
  11. Zheng, H.; Ma, J.; Zhou, M.; Li, D.; Yao, X.; Cao, W.; Zhu, Y.; Cheng, T. Enhancing the nitrogen signals of rice canopies across critical growth stages through the integration of textural and spectral information from unmanned aerial vehicle (UAV) multispectral imagery. Remote Sens. 2020, 12, 957. [Google Scholar] [CrossRef]
  12. Zhang, L.; Song, X.; Niu, Y.; Zhang, H.; Wang, A.; Zhu, Y.; Zhu, X.; Chen, L.; Zhu, Q. Estimating winter wheat plant nitrogen content by combining spectral and texture features based on a Low-Cost UAV RGB system throughout the growing season. Agriculture 2024, 14, 456. [Google Scholar] [CrossRef]
  13. Gitelson, A.A.; Viña, A.; Ciganda, V.; Rundquist, D.C.; Arkebauer, T.J. Remote estimation of canopy chlorophyll content in crops. Geophys. Res. Lett. 2005, 32. [Google Scholar] [CrossRef]
  14. Dube, T.; Mutanga, O. Investigating the robustness of the new Landsat-8 Operational Land Imager derived texture metrics in estimating plantation forest aboveground biomass in resource constrained areas. Isprs J. Photogramm. Remote Sens. 2015, 10, 12–32. [Google Scholar] [CrossRef]
  15. Zhou, Y.; Lao, C.; Yang, Y.; Zhang, Z.; Yang, N. Diagnosis of winter-wheat water stress based on UAV-borne multispectral image texture and vegetation indices. Agric. Water Manag. 2021, 256, 107076. [Google Scholar] [CrossRef]
  16. Fan, Y.; Feng, H.; Yue, J.; Jin, X.; Liu, Y.; Chen, R.; Bian, M.; Ma, Y.; Song, X.; Yang, G. Using an optimized texture index to monitor the nitrogen content of potato plants over multiple growth stages. Comput. Electron. Agric. 2023, 212, 108147. [Google Scholar] [CrossRef]
  17. Chen, P.F.; Liang, F. Cotton nitrogen nutrition diagnosis based on spectrum and texture feature of images from low altitude unmanned aerial vehicle. Sci. Agric. Sin. 2019, 52, 2220–2229. [Google Scholar]
  18. Wu, L.; Gong, Y.; Bai, X.; Wang, W.; Wang, Z. Nondestructive determination of leaf nitrogen content in corn by hyperspectral imaging using spectral and texture fusion. Appl. Sci. 2023, 13, 1910. [Google Scholar] [CrossRef]
  19. Su, Q.; Lv, J.; Fan, J.; Zeng, W.; Pan, R.; Liao, Y.; Song, Y.; Zhao, C.; Qin, Z.; Defourny, P. Remote sensing-based classification of winter irrigation fields using the random forest algorithm and GF-1 Data: A case study of Jinzhong Basin, North China. Remote Sens. 2023, 15, 4599. [Google Scholar] [CrossRef]
  20. Zheng, M.; Zhu, P.; Zheng, J.; Xue, L.; Zhu, Q.; Cai, X.; Cheng, S.; Zhang, Z.; Kong, F.; Zhang, J. Effects of soil texture and nitrogen fertilization on soil bacterial community structure and nitrogen uptake in flue-cured tobacco. Sci. Rep. 2021, 11, 22643. [Google Scholar] [CrossRef]
  21. Yue, J.; Yang, G.; Li, C.; Liu, Y.; Wang, J.; Guo, W.; Ma, X.; Niu, Q.; Qiao, H.; Feng, H. Analyzing winter-wheat biochemical traits using hyperspectral remote sensing and deep learning. Comput. Electron. Agric. 2024, 222, 109026. [Google Scholar] [CrossRef]
  22. Thompson, L.J.; Puntel, L.A. Transforming unmanned aerial vehicle (UAV) and multispectral sensor into a practical decision support system for precision nitrogen management in corn. Remote Sens. 2020, 12, 1597. [Google Scholar] [CrossRef]
  23. Silva, L.; Conceição, L.A.; Lidon, F.C.; Maçãs, B. Remote monitoring of crop nitrogen nutrition to adjust crop models: A review. Agriculture 2023, 13, 835. [Google Scholar] [CrossRef]
  24. Huang, S.; Yang, J.; Yang, Y.; Jiang, R.; He, P.; Jia, L. Long-term nutrient expert management improves nitrogen use efficiency and enhances soil organic carbon sequestration in wheat-maize rotation system. Acta Agric. Boreali-Sin. 2021, 36, 154–161. [Google Scholar] [CrossRef]
  25. Messina, G.; Praticò, S.; Badagliacca, G.; Fazio, S.D.; Modica, G. Monitoring onion crop “Cipolla Rossa di Tropea Calabria IGP” growth and yield response to varying nitrogen fertilizer application rates using UAV imagery. Drones 2021, 5, 61. [Google Scholar] [CrossRef]
  26. Yu, F.; Xing, S.; Guo, Z.; Bai, J.; Xu, T. Remote sensing inversion of the nitrogen content in rice leaves using character transfer vegetation index. Trans. Chin. Soc. Agric. Eng. 2022, 38, 175–182. [Google Scholar] [CrossRef]
  27. Erunova, M.G.; Pisman, T.I.; Shevyrnogov, A.P. The technology for detecting weeds in agricultural crops based on vegetation index VARI (PlanetScope). J. Sib. Fed. Univ. Eng. Technol. 2021, 14, 347–353. [Google Scholar] [CrossRef]
  28. Rondeaux, G.; Steven, M.; Baret, F. Optimization of soil-adjusted vegetation indices. Remote Sens. Environ. 1996, 55, 95–107. [Google Scholar] [CrossRef]
  29. Jordan, C.F. Derivation of leaf-area index from quality of light on the forest floor. Ecology 1969, 50, 663–666. [Google Scholar] [CrossRef]
  30. Huete, A.; Justice, C.; Van Leeuwen, W. MODIS Vegetation Index (MOD13) Algorithm Theoretical Basis Document; NASA Goddard Space Flight Center: Greenbelt, MD, USA, 1999. [Google Scholar]
  31. Lamm, R.D.; Slaughter, D.C.; Giles, D.K. Precision weed control system for cotton. Trans. ASABE 2002, 45, 231. [Google Scholar] [CrossRef]
  32. Gitelson, A.A.; Kaufman, Y.J.; Merzlyak, M.N. Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sens. Environ. 1996, 58, 289–298. [Google Scholar] [CrossRef]
  33. Gilabert, M.; González-Piqueras, J.; Garcıa-Haro, F.; Meliá, J. A generalized soil-adjusted vegetation index. Remote Sens. Environ. 2002, 82, 303–310. [Google Scholar] [CrossRef]
  34. Rouse, J.W.; Haas, R.W.; Schell, J.A.; Deering, D.W.; Harlan, J.C. Monitoring the Vernal Advancement and Retrogradation (Green Wave Effect) of Natural Vegetation; NASA/GSFC Type III, Final Report; NASA: Greenbelt, MD, USA, 1974. [Google Scholar]
  35. Haboudane, D.; Miller, J.R.; Pattey, E.; Zarco-Tejada, P.J.; Strachan, I.B. Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: Modeling and validation in the context of precision agriculture. Remote Sens. Environ. 2003, 90, 337–352. [Google Scholar] [CrossRef]
  36. Reyniers, M.; Walvoort, D.J.; De Baardemaaker, J. A linear model to predict with a multi-spectral radiometer the amount of nitrogen in winter wheat. Int. J. Remote Sens. 2006, 27, 4159–4179. [Google Scholar] [CrossRef]
  37. Cao, Q.; Miao, Y.; Shen, J.; Yu, W.; Yuan, F.; Cheng, S.; Huang, S.; Wang, H.; Yang, W.; Liu, F. Improving in-season estimation of rice yield potential and responsiveness to topdressing nitrogen application with Crop Circle active crop canopy sensor. Precis. Agric. 2016, 17, 136–154. [Google Scholar] [CrossRef]
  38. Deering, D.W.; Rouse, J.W., Jr.; Haas, R.H.; Schell, J.A. MSS data. In Proceedings of the 10th International Symposium on Remote Sensing of Environment, Ann Arbor, MI, USA, 6–10 October 1975. [Google Scholar]
  39. Usha, S.G.A.; Vasuki, S. Significance of texture features in the segmentation of remotely sensed images. Optik 2022, 249, 168241. [Google Scholar] [CrossRef]
  40. Wan, L.; Li, Y.; Cen, H.; Zhu, J.; Yin, W.; Wu, W.; Zhu, H.; Sun, D.; Zhou, W.; He, Y. Combining UAV-Based vegetation indices and image classification to estimate flower number in oilseed rape. Remote Sens. 2018, 10, 1484. [Google Scholar] [CrossRef]
  41. Yang, Y.; Chen, J. Comprehensive analysis of water carrying capacity based on wireless sensor network and image texture of feature extraction. Alex. Eng. J. 2022, 61, 2877–2886. [Google Scholar] [CrossRef]
  42. Fernández-Habas, J.; Cañada, M.C.; Moreno, A.M.G.; Leal-Murillo, J.R.; González-Dugo, M.P.; Oar, B.A.; Gómez-Giráldez, P.J.; Fernández-Rebollo, P. Estimating pasture quality of Mediterranean grasslands using hyperspectral narrow bands from field spectroscopy by Random Forest and PLS regressions. Comput. Electron. Agric. 2022, 192, 106614. [Google Scholar] [CrossRef]
  43. Jeung, M.; Baek, S.; Beom, J.; Cho, K.H.; Her, Y.; Yoon, K. Evaluation of random forest and regression tree methods for estimation of mass first flush ratio in urban catchments. J. Hydrol. 2019, 575, 1099–1110. [Google Scholar] [CrossRef]
  44. Panahi, M.; Sadhasivam, N.; Pourghasemi, H.R.; Rezaie, F.; Lee, S. Spatial prediction of groundwater potential mapping based on convolutional neural network (CNN) and support vector regression (SVR). J. Hydrol. 2020, 588, 125033. [Google Scholar] [CrossRef]
  45. Lin, F.; Hai, C.L.; Mei, L.Q. Multi-source aerodynamic data fusion modeling with XGBoost. Acta Aerodyn. Sin. 2023, 42, 27–34. [Google Scholar] [CrossRef]
  46. Du, B.; Lund, P.D.; Wang, J.; Kolhe, M.; Hu, E. Comparative study of modelling the thermal efficiency of a novel straight through evacuated tube collector with MLR, SVR, BP and RBF methods. Sustain. Energy Technol. Assess. 2021, 44, 101029. [Google Scholar] [CrossRef]
  47. Yin, Q.G. Research on Batch Evaluation of Real Estate Tax Base Based on Ensemble Learning: Agriculture. Master’s Thesis, Qingdao University of Technology, Qingdao, China, 2023. [Google Scholar] [CrossRef]
  48. Li, Y.M.; Tan, Z.Y.; Yang, C.; He, F.; Meng, D.; Luo, J.H.; Duan, H.T. Extraction of algal blooms in Dianchi Lake based on multi-source satellites using machine learning algorithms. Adv. Earth Sci. 2022, 37, 1141–1156. [Google Scholar] [CrossRef]
  49. Lin, L.; Liu, X. Mixture-based weight learning improves the random forest method for hyperspectral estimation of soil total nitrogen. Comput. Electron. Agric. 2022, 192, 106634. [Google Scholar] [CrossRef]
  50. Jiang, Q.; Xu, L.; Sun, S.; Wang, M.; Xiao, H. Retrieval model for total nitrogen concentration based on UAV hyper spectral remote sensing data and machine learning algorithms–A case study in the Miyun Reservoir, China. Ecol. Indic. 2021, 124, 107356. [Google Scholar] [CrossRef]
  51. Shen, Z.; Zhang, R.; Long, H.; Xu, A. Research on spatial distribution of soil texture in Southern Ningxia based on machine learning. Sci. Agric. Sin. 2022, 55, 2961–2972. [Google Scholar] [CrossRef]
  52. Shi, P.; Wang, Y.; Xu, J.; Zhao, Y.; Yang, B.; Yuan, Z.; Sun, Q. Rice nitrogen nutrition estimation with RGB images and machine learning methods. Comput. Electron. Agric. 2021, 180, 105860. [Google Scholar] [CrossRef]
  53. Gallegos, Á.; Gavito, M.E.; Ferreira-Medina, H. Foliar nitrogen estimation with artificial intelligence and technological tools: State of the art and future challenges. Int. J. Adv. Comput. Sci. Appl. 2024, 15, 375–386. [Google Scholar] [CrossRef]
Agriculture 15 01373 g001
Figure 1. Schematic diagram of field experimental design (split-plot arrangement with Ecological Intensification (EI) and Farmers’ Practice (FP) as main treatments, and three nitrogen application strategies (0 N, 2/3 N, 3/3 N) as sub-treatments).
Figure 1. Schematic diagram of field experimental design (split-plot arrangement with Ecological Intensification (EI) and Farmers’ Practice (FP) as main treatments, and three nitrogen application strategies (0 N, 2/3 N, 3/3 N) as sub-treatments).
Agriculture 15 01373 g001
Agriculture 15 01373 g002
Figure 2. Pearson correlation coefficients between vegetation indices and measured PNC. Note: Color scale represents correlation strength (red: positive; blue: negative). Significant features displayed with “**” at p < 0.01 and “*” at p < 0.05.
Figure 2. Pearson correlation coefficients between vegetation indices and measured PNC. Note: Color scale represents correlation strength (red: positive; blue: negative). Significant features displayed with “**” at p < 0.01 and “*” at p < 0.05.
Agriculture 15 01373 g002
Agriculture 15 01373 g003
Figure 3. Pearson correlation coefficients between texture features and measured plant nitrogen content. Note: Color scale represents correlation strength (red: positive; blue: negative). Significant features displayed with “**” at p < 0.01 and “*” at p < 0.05.
Figure 3. Pearson correlation coefficients between texture features and measured plant nitrogen content. Note: Color scale represents correlation strength (red: positive; blue: negative). Significant features displayed with “**” at p < 0.01 and “*” at p < 0.05.
Agriculture 15 01373 g003
Agriculture 15 01373 g004
Figure 4. Performance comparison of winter wheat plant nitrogen content prediction models using different machine learning models. Note: Black solid line = 1:1 line.
Figure 4. Performance comparison of winter wheat plant nitrogen content prediction models using different machine learning models. Note: Black solid line = 1:1 line.
Agriculture 15 01373 g004
Agriculture 15 01373 g005
Figure 5. Transferability assessment of plant nitrogen content prediction models between management systems (a,c,e) when applying EI-trained models to FP plots and (b,d,f) FP-trained models to EI plots. (Density refers to kernel density estimation of prediction residuals).
Figure 5. Transferability assessment of plant nitrogen content prediction models between management systems (a,c,e) when applying EI-trained models to FP plots and (b,d,f) FP-trained models to EI plots. (Density refers to kernel density estimation of prediction residuals).
Agriculture 15 01373 g005
Agriculture 15 01373 g006
Figure 6. Performance of training and validation models.
Figure 6. Performance of training and validation models.
Agriculture 15 01373 g006
Table 1. UAV multispectral image acquisition parameters.
Table 1. UAV multispectral image acquisition parameters.
Parameter CategoryDetails
Spectral bandsCenter wavelength/nm
Blue450
Green550
Red685
Red-edge725
Near-infrared780
Height of flight50 m
Lens orientationVertically downward
Field of view30°
Forward overlap80%
Side overlap75%
Table 2. Definition and calculation formulas of vegetation indices used in this study.
Table 2. Definition and calculation formulas of vegetation indices used in this study.
Vegetation Indices (Abbreviation)FormulasReference
Atmospherically resistant vegetation index (VARI)(Green − Red)/(Green + Red − Blue)[27]
Chlorophyll absorption ratio index (CARI)(RE − Red) − 0.2 × (RE + Red)[28]
Difference vegetation index (DVI)NIR − Red[29]
Enhanced vegetation index (EVI)2.5 × (NIR − Red)/(NIR + 6 × Red − 7.5 × Blue + 1)[30]
Excessive green index (EXG)2 × Green − Red − Blue[31]
Green-band-normalized vegetation index (GNDVI)(NIR − Green)/(NIR + Green)[32]
Green-band-optimized soil-adjusted vegetation index (GOSAVI)1.16 × (NIR − Green)/(NIR + Green + 0.16)[33]
Modified simple ratio index (MSR)((NIR/Red) − 1)/(((NIR/Red) + 1)0.5)[34]
Modified triangle vegetation index (MTVI)(1.5 × (1.2 × (NIR − Green) − 2.5 × (Red − Green)))/(((2 × NIR + 1)2 − 6 × NIR − 5 × (Red)0.5 − 0.5)0.5)[35]
Normalized blue-green band difference vegetation index (GBNDVI)(NIR − (Green + Blue))/(NIR + Green + Blue)[32]
Normalized blue-green difference index (NGBDI)(Green − Blue)/(Green + Blue)[13]
Normalized difference vegetation index (NDVI)(NIR − Red)/(NIR + Red)[34]
Optimized soil adjusted vegetation index (OSAVI)(NIR − Red)/(NIR + Red + 0.16)[28]
Optimized vegetation index (VIplot)1.45 × (NIR2 + 1) × (Red + 0.45)[36]
Ratio vegetation index (RVI)NIR/Red[27]
Red-edge-band-optimized soil-adjusted vegetation index (REOSAVI)1.16 × (NIR − Red)/(NIR/Red + 0.16)[37]
Red-edge-band-renormalized difference vegetation index (RERDVI)(NIR − RE)/((NIR + RE)0.5)[32]
Red-edge-normalized difference vegetation index (RENDVI)NIR − RE[32]
Renormalized difference vegetation index (RDVI)(NIR − Red) × (NIR + Red)0.5[35]
Soil-adjusted vegetation index (SAVI)1.5 × (NIR − Red)/(NIR + Red + 0.5)[30]
Triangle vegetation index (TVI)0.5 × (120 × (NIR − Green) − 200 × (Red − Green))[38]
Vegetation red-edge index (VREI)NIR/RE[32]
Table 3. Configuration parameters of machine learning models (RF, SVR, and XGBoost) used for plant nitrogen content estimation, including parameter settings and optimization methods.
Table 3. Configuration parameters of machine learning models (RF, SVR, and XGBoost) used for plant nitrogen content estimation, including parameter settings and optimization methods.
ModelParametersParameter Value
RFData cut0.75
Data shufflingYes
Cross-validation3-fold cross-validation
Identity node split evaluation criterionMSE
Minimum number of samples for internal node splitting2
Minimum number of samples of leaf nodes1
Maximum depth of tree10
Maximum number of leaf nodes50
Numbers of decision trees500
SVRData cut0.75
Data shufflingYes
Cross-validation3-fold cross-validation
Penalty factor1
Kernel functionlinear
Kernel function coefficientscale
Maximum number of terms in kernel function3
Error convergence condition0.001
Maximum number of iterations1000
XGboostData cut0.75
Data shufflingYes
Cross-validation3-fold cross-validation
Learning rate0.3
gamma0.001
Maximum depth of tree10
subsample0.7
nrounds1000
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhang, J.; Cheng, G.; Huang, S.; Yang, J.; Yang, Y.; Xing, S.; Wang, J.; Yang, H.; Nie, H.; Yang, W.; et al. Winter Wheat Nitrogen Content Prediction and Transferability of Models Based on UAV Image Features. Agriculture 2025, 15, 1373. https://doi.org/10.3390/agriculture15131373

AMA Style

Zhang J, Cheng G, Huang S, Yang J, Yang Y, Xing S, Wang J, Yang H, Nie H, Yang W, et al. Winter Wheat Nitrogen Content Prediction and Transferability of Models Based on UAV Image Features. Agriculture. 2025; 15(13):1373. https://doi.org/10.3390/agriculture15131373

Chicago/Turabian Style

Zhang, Jing, Gong Cheng, Shaohui Huang, Junfang Yang, Yunma Yang, Suli Xing, Jingxia Wang, Huimin Yang, Haoliang Nie, Wenfang Yang, and et al. 2025. "Winter Wheat Nitrogen Content Prediction and Transferability of Models Based on UAV Image Features" Agriculture 15, no. 13: 1373. https://doi.org/10.3390/agriculture15131373

APA Style

Zhang, J., Cheng, G., Huang, S., Yang, J., Yang, Y., Xing, S., Wang, J., Yang, H., Nie, H., Yang, W., Yu, K., & Jia, L. (2025). Winter Wheat Nitrogen Content Prediction and Transferability of Models Based on UAV Image Features. Agriculture, 15(13), 1373. https://doi.org/10.3390/agriculture15131373

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop