Enhanced Estimation of Rice Leaf Nitrogen Content via the Integration of Hybrid Preferred Features and Deep Learning Methodologies

Abstract

Efficiently obtaining leaf nitrogen content (LNC) in rice to monitor the nutritional health status is crucial in achieving precision fertilization on demand. Unmanned aerial vehicle (UAV)-based hyperspectral technology is an important tool for determining LNC. However, the intricate coupling between spectral information and nitrogen remains elusive. To address this, this study proposed an estimation method for LNC that integrates hybrid preferred features with deep learning modeling algorithms based on UAV hyperspectral imagery. The proposed approach leverages XGBoost, Pearson correlation coefficient (PCC), and a synergistic combination of both to identify the characteristic variables for LNC estimation. We then construct estimation models of LNC using statistical regression methods (partial least-squares regression (PLSR)) and machine learning algorithms (random forest (RF); deep neural networks (DNN)). The optimal model is utilized to map the spatial distribution of LNC at the field scale. The study was conducted at the National Agricultural Science and Technology Park, Guangzhou, located in Baiyun District of Guangdong, China. The results reveal that the combined PCC-XGBoost algorithm significantly enhances the accuracy of rice nitrogen inversion compared to the standalone screening approach. Notably, the model built with the DNN algorithm exhibits the highest predictive performance and demonstrates great potential in mapping the spatial distribution of LNC. This indicates the potential role of the proposed model in precision fertilization and the enhancement of nitrogen utilization efficiency in rice cultivation. The outcomes of this study offer a valuable reference for enhancing agricultural practices and sustainable crop management.

1. Introduction

Nitrogen serves as a vital nutrient in fostering the growth and maturation of rice plants. The suitable application of nitrogen fertilizer stands as a pivotal strategy in attaining both abundant yields and superior quality in rice crops [1]. According to statistics, approximately 200 million tons of nitrogen are applied globally to farmland, yet its utilization efficiency remains below 50%. The resultant nitrogen loss results in problems such as atmospheric and water contamination, biodiversity depletion, and the exacerbation of climate fluctuations [2,3]. Accurate estimations of nitrogen levels in rice leaves can be used to overcome these challenges, facilitating precise fertilization tailored to the plant’s nutritional requirements. This can consequently reduce the volume of chemical fertilizers used, enhance their efficacy, and mitigate nitrogen loss [4]. However, conventional chemical testing methods are costly, time-consuming, and inherently lag behind real-time agricultural demands [5]. Unmanned aerial vehicle (UAV)-based hyperspectral technology can overcome these bottlenecks due to its exceptional spectral resolution, spatial precision, and detection sensitivity. Leveraging these attributes, UAV-based monitoring has emerged as a potent tool for gauging the nitrogen status of rice paddies. It could invert nitrogen information based on UAV spectral data and calculate the rice nitrogen deficit in combination with critical nitrogen concentration curve, effectively supporting dynamic nitrogen monitoring and optimizing the management of nitrogen fertilizers [6,7]. This significantly contributes to the management of nitrogen pollution in farmland, promoting a more sustainable and environmentally friendly agricultural practice.

Currently, estimations of rice nitrogen content based on UAV hyperspectral technology primarily concentrate on two key steps: the screening of characteristic variables and the optimization of estimation models. The former, which is crucial for accurately estimating leaf nitrogen content (LNC), can be categorized into two groups based on the methodological approach. The first approach screens the characteristic variables using the correlation between the variables and LNC. Methods such as the maximum information coefficient, Pearson correlation coefficient (PCC), and uninformative variables elimination are commonly employed for this approach. However, due to the inherent weakness of spectral information and the challenges in its acquisition, relying solely on statistical analysis methods can often result in an incomplete determination of optimal feature bands. The second approach centers on feature-importance-based variable screening. Algorithms such as random forest, gradient boosting decision tree, and extreme gradient boosting are examples of this category. These methods iteratively select a subset of features from the initial feature set, train the learner, evaluate the subset based on the learner’s performance, and ultimately identify features with higher importance [8,9]. However, scholars have identified inconsistencies in selected wavebands when using various methods to estimate crop nitrogen content. The multi-method ensemble for wavebands selection may improve the definition of particular spectral regions in relation to specific absorption features, thereby increasing the reliability of the results, surpassing the capabilities of a single method [10,11]. In addition, given the intricate relationship between spectra and rice nitrogen, employing just one screening method may result in the omission of crucial information, thereby compromising inversion accuracy.

At present, the direct inversion method and vegetation index construction method are the two primary approaches used for rice nitrogen estimation based on UAV hyperspectral technology. The direct inversion method determines nitrogen information by establishing a relationship model between nitrogen content and spectral reflectance. For example, Yang et al. [12] utilized characteristic variables identified through principal component analysis (PCA) and employed the support vector machine (SVM) algorithm to construct a relationship model between nitrogen content and these variables. Their findings demonstrate the efficacy of the PCA-SVM method in assessing LNC. However, combining bands has been reported to provide more comprehensive information and enhance the accuracy of nitrogen estimations compared to the use of a single band. Consequently, scholars have established nitrogen estimation models through the construction of vegetation indices. This method is currently the most widely adopted approach. For example, Yu et al. [13] employed the successive projections algorithm to extract characteristic bands and proposed the nitrogen characteristic transfer index (NCTI) composed of these bands. The authors then constructed an LNC estimation model using the linear regression method, achieving an R² value of 0.774. Similarly, Wang et al. [8] integrated vegetation, color and texture indices, hyperspectral parameters, and machine learning algorithms to construct an estimation model of nitrogen concentration in rice stems and leaves, with R² values of 0.7 and 0.8, respectively. These studies offer valuable insights for nitrogen estimation in rice, yet the complex relationship between spectra and nitrogen requires the introduction of novel algorithms to enhance model accuracy.

The contribution of this work was to develop a novel method for LNC estimation in rice that combines the screening of characteristic variables with model optimization. In this method, spectral variables were screened using a hybrid method involving PCC and the extreme gradient boosting (XGBoost) algorithm. A deep neural network algorithm was subsequently employed to establish a model capturing the relationship between characteristic variables and LNC. This model was then compared against commonly employed methods such as partial least squares regression (PLSR), and random forest (RF). The optimal model was subsequently utilized to map the spatial distribution of LNC.

2. Materials and Methods

2.1. Field Experimental Design

The experiment was performed at the National Agricultural Science and Technology Park, Guangzhou, located in Baiyun District of Guangdong, China (23°23′38″ N, 113°25′37″ E). This area has a subtropical monsoon climate, with an average annual temperature of approximately 24 °C. We selected the conventional rice variety Meixiangzhan No. 2 as the experimental material, with an average plant spacing of 25 cm. The experiment comprised six nitrogen fertilizer gradients, namely, N1 (0 kg/hm²), N2 (37.5 kg/hm²), N3 (75.0 kg/hm²), N4 (112.5 kg/hm²), N5 (150.0 kg/hm²), and N6 (187.5 kg/hm²). Each gradient was replicated five times, resulting in a total of 30 experimental blocks. The phosphorus and potassium fertilizer application followed the local standard rates. Figure 1 presents the specific experimental layout. The area of each block was approximately 19.0 m² for No. 1–6 and 22.5 m² for No. 7–30.

2.2. Data and Pre-Processing

2.2.1. Collection and Pre-Processing of UAV Hyperspectral Imagery

Hyperspectral images were collected during the tillering (3 September 2022), jointing-to-booting (27 September 2022), and heading-to-flowering (21 October 2022) stages of rice using a UAV-based hyperspectral imaging system (Gaiasky mini3-VN, Dualix Spectral Image Technology Co. Ltd., Wuxi, China). This system offers a spectral range spanning from 400 to 1000 nm, with a spectral resolution of 5 nm and a total of 224 bands (Figure 2). Data collection took place on days of clear weather, with the flight altitude maintained at 50 m. To ensure data accuracy, calibration was conducted with a standard white plate and standard gray cloth prior to flights. The acquired images were pre-processed using SpecView 2.9.3.10, HySpectralStitcher 1.0.1, and ENVI 5.6. Pre-processing steps included geometric correction, radiometric correction, image stitching, clipping, etc. We computed the average of the spectra in the region of interest (ROI) for each block. Gauss filtering was then employed to denoise the spectral data, yielding canopy spectral reflectance data for each block (

Table 1. Basic information of UAV hyperspectral images.

Instrumentation	Flight Altitude	Block Size	Spectral Range	Forward Overlap	Side Overlap
Gaiasky mini 3-VN	50 m	No.1–6: 19.0 m2 No. 7–30: 22.5 m2	400–1000 nm	80%	65%

).

2.2.2. Acquisition of Rice Agronomic Parameters

Rice plant samples were gathered from five points in an “X” shape (Figure 3) at each block and the acquisition time coincided with the collection of the UAV hyperspectral images. The collected samples were then washed and the fresh leaves of each rice plant were weighed, labeled, and placed in bags for oven-drying at 105 °C for 30 min. The temperature was then adjusted to 80 °C for drying until a constant weight was achieved. After drying, the samples were crushed using a pulverizer through a 0.25 mm sieve. The Kjeldahl method was subsequently employed to determine the LNC.

Table 2. Statistical parameters of leaf nitrogen contents from the rice plant samples (unit: %).

Growth Stages	Minimum	Maximum	Mean	SD	CV (%)
tillering	2.02	3.66	2.63	0.44	16.90
jointing-to-booting	1.69	2.55	2.03	0.23	11.56
heading-to-flowering	1.42	2.06	1.69	0.18	10.38

Note: SD, standard deviation; CV, coefficient of variation.

reports the statistical results. The LNC was observed to progressively decrease with rice growth. The mean LNC during the tillering, jointing-to-booting, and heading-to-flowering stages was 2.63%, 2.03%, and 1.69%, respectively. Moreover, LNC exhibited the highest standard deviation (0.44) and coefficient of variation (16.90%) at the tillering stage, indicating greater data variability.

2.3. Vegetation Index Construction

Vegetation indices (VIs) serve as valuable indicators of crop health and growth, reflecting the nutritional status of crops [14]. Drawing from the relevant studies [15,16,17,18,19], 20 VIs were selected as variables for estimating the LNC (

Table 3. Definitions and formulae of the selected vegetation indices.

VI	Formula	Reference	VI	Formula	Reference
NPCI	(R670 − R460)/(R670 + R460)	[15]	SIPI	(R800 − R445)/(R800 − R680)	[16]
SR	R750/R550	[15]	PSRI	(R680 − R500)/R750	[17]
MSR	(R800/R760 − 1)/(R800/R670 + 1)0.5	[15]	GI	R554/R677	[17]
PBI	R810/R560	[15]	PSND	(R800 − R470)/(R800 + R470)	[17]
LCI	(R850 − R710)/(R850 + R680)		PSSR	R800/R500	[17]
GNDVI	(R750 − R550)/(R750 + R550)	[16]	RARS	R760/R500	[17]
SRPI	R430/R680	[16]	OSAVI	1.16 × (R800 − R670)/(R800 + R670 + 0.16)	[17]
PRI	(R570 − R531)/(R570 + R531)	[16]	RENDVI	(R750 − R705)/(R750 + R705)	[18]
MTCI	(R750 − R710)/(R710 − R680)	[16]	DCNI	(R720 − R700)/(R700 − R670)/(R720 − R670 + 0.03)	[19]
NDRE	(R790 − R720)/(R790 + R720)	[16]	NDVI	(R800 − R670)/(R800 + R670)	[19]

Note: Rx denotes the spectral reflectance at wavelength x nm; NPCI, normalized pigment chlorophyll ratio index; SR, simple ratio vegetation index; MSR, modified simple ratio; PBI, plant biochemical index; LCI, leaf chlorophyll index; GNDVI, green normalized difference vegetation index; SRPI, simple ratio pigment index; PRI, photochemical reflectance index; MTCI, MERIS terrestrial chlorophyll index; NDRE, normalized difference red edge index; SIPI, structure intensive pigment index; PSRI, plant senescence reflectance index; GI, greenness index; PSND, pigment-specific normalized difference; PSSR, pigment-specific simple ratio; RARS, ratio analysis of reflectance spectra; OSAVI, optimized soil adjusted vegetation index; NDVI, normalized difference vegetation index; RENDVI, red edge NDVI; DCNI, double-peak canopy nitrogen index.

).

2.4. Identification of Characteristic Variables for LNC Estimation

The determination of the characteristic variables is a critical step in the development of the LNC estimation method [6]. However, conventional statistical analysis methods often struggle to precisely determine the optimal characteristic variables due to the inherent challenges in capturing weak spectral information. Moreover, the relationship between spectral data and agronomic parameters is typically complex, encompassing a combination of linear and nonlinear associations [20]. In this study, we employ a hybrid approach integrating correlation analysis with XGBoost to screen characteristic variables for estimating LNC and compare it with the equivalent singular method. These methods are briefly described in the following sections:

• Pearson correlation coefficient

A significant correlation has been reported between the spectral reflectance of certain bands and the LNC [6]. In this study, bands exhibiting high correlation coefficients with statistical significance at p ≤ 0.01 were identified as characteristic variables using the Pearson correlation coefficient, ${\mathrm{r}}_i$, determined as [21]:

$${\mathrm{r}}_i=\frac{\sum _{n=1}^N(R_{ni}-\overline{R_i})(y_n-\overline{y})}{\sqrt{\sum _{n=1}^N{(R_{ni}-\overline{R_i})}^2\sum _{n=1}^N{(y_n-\overline{y})}^2}}$$

(1)

where ${\mathrm{r}}_i$ is the correlation coefficient between LNC and spectral reflectance; $R_{ni}$ is the spectral reflectance of the n-th sample in the i-th band; $\overline{R_i}$ is the mean reflectance of samples in the i-th band; $y_n$ is the LNC of n-th sample; and $\overline{y}$ is the mean value of LNC.

• Extreme gradient boosting

We employed the XGBoost algorithm to select the optimal characteristic variables for LNC estimation. XGBoost is a modified gradient-boosting algorithm that averages and subsequently ranks the feature importance (FI) of each tree. It employs the following calculation method [22]:

$$\mathrm{F}\mathrm{I}\left(\mathrm{T},\mathrm{F}\right)=\mathrm{H}\left(\mathrm{T}\right)-\mathrm{H}\left(\mathrm{T}|\mathrm{F}\right)=-{\sum }_{i=1}^j{\mathrm{p}}_i{\mathrm{l}\mathrm{o}\mathrm{g}}_2{\mathrm{p}}_i-{\sum }_{\mathrm{F}}\mathrm{p}\left(\mathrm{F}\right)\times {\sum }_{i=1}^j\mathrm{p}\left(i|\mathrm{F}\right){\mathrm{l}\mathrm{o}\mathrm{g}}_2\mathrm{p}\left(i|\mathrm{F}\right),$$

(2)

where $\mathrm{H}\left(\mathrm{T}\right)$ and $\mathrm{H}\left(\mathrm{T}|\mathrm{F}\right)$ refer to the entropy of the parent and child nodes based on the F-feature segmentation, respectively, and ${\mathrm{p}}_i$ represents the score of the labeled samples at the i-th node.

2.5. Model Construction and Validation

Three algorithms, namely partial least squares regression (PLSR), random forest (RF), and deep neural network (DNN), were employed to develop the LNC estimation model using the selected characteristic variables. These algorithms were created using Python software (version 3.10). In the following, we present a summary of each algorithm.

• Partial least squares regression

Partial least-squares regression (PLSR) integrates the strengths of three analytical techniques: principal component analysis, canonical correlation analysis, and multiple linear regression. It leverages all available data to construct a model, extracting maximal information that reflects data variation. PLSR demonstrates a strong predictive capability and is particularly effective at handling datasets with strong linear correlations among variables [23]. This method includes a principal component analysis of both the spectral data matrix and the matrix of rice leaf content during modeling, resulting in a regression model based on the contributions of the derived variables.

• Random forest

Random forest (RF) is an ensemble learning algorithm proficient in effectively modeling the nonlinear relationship between the characteristic and explanatory variables. It boasts a strong generalization ability and robustness and a rapid training speed, and only requires a minimal number of tuning parameters. The model employs a bootstrap strategy to generate new samples of equal size from the original data. Two-thirds of the original sample are typically utilized to construct decision trees. The remaining one-third serves as out-of-bag data (OOB) for inner cross-validation to evaluate the estimated accuracy of RF. The results from all decision trees are aggregated and their average serves as the final prediction outcome [24].

• Deep neural network

Deep neural networks (DNNs) are advanced neural network architectures characterized by multi-layered structures comprising input layers, multiple hidden layers, and output layers. Each layer is interconnected, with connections existing between nodes in adjacent layers. However, nodes within the same layer and across layers are not interconnected. Through fully connected layers and combinations of activation functions, various nodes are linked to construct a sophisticated multi-layer neural network model [25]. Figure 4 depicts the structural layout of DNNs.

DNNs undergo training via forward and backward propagation algorithms, continually updating each weight (w) and bias (b) within the network. The backward propagation algorithm, pivotal to DNN functionality, retroactively propagates the error between the estimated and measured output values. It iteratively adjusts the w and b across each layer of the network model, aiming to minimize the error until it aligns with specified accuracy criteria [26]. The weight and bias are calculated as follows [25]:

$$w^{\mathrm{*}}=w-\alpha \frac{\alpha j(w,b)}{\alpha w}$$

(3)

$$b^{\mathrm{*}}=b-\alpha \frac{\alpha j(w,b)}{\alpha b}$$

(4)

where $w^{\mathrm{*}}$ and $w$ denote the weights after and before the update, respectively; $b^{\mathrm{*}}$ and $b$ represent the biases after and before the update, respectively; $\alpha$ is the learning rate; and $j(w,b)$ is the loss function of the model. The optimal parameters $w^{\mathrm{*}}$ and $w$ are attained through iterative gradient descent. In this study, a DNN comprising three hidden layers was constructed. The number of nodes in the input layer corresponded to the number of characteristic variables. We adopted tanh as the activation function. The network was trained for 3000 iterations and the mean squared error (MSE) was adopted as the Loss function to assess the performance.

• Accuracy verification

The collected data (90 samples) were divided into a training set (60 samples) and a test set (30 samples). The test set was utilized to evaluate the performance of the LNC estimation models by computing the coefficient of determination (R²) and root mean squared error (RMSE). Models with a higher R² and lower RMSE are considered to have a better estimation effect [27].

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

(5)

$$\mathrm{RMSE}=\sqrt{\frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}({\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}{)}^2}{\mathrm{n}}}$$

(6)

where ${\mathrm{y}}_{\mathrm{i}}$ and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ are the measured and estimated values, respectively; $\stackrel{\mathrm{-}}{\mathrm{y}}$ is the mean of the measured values; and $\mathrm{n}$ signifies the sample number.

3. Results

3.1. Determining the Characteristic Variables for the Estimation of LNC

3.1.1. Spectral Characteristics of Rice at Different Growth Stages

Figure 5 presents the spectral curves of rice at various growth stages under different nitrogen (N) application levels. The rice canopy spectra exhibit low reflectance (r < 0.1) within the 400–670 nm range, followed by a sharp increase in reflectance within the red-edge range (670–760 nm). Notably, marked differences in spectra are observed across varying nitrogen levels within the 750–1000 nm range, with nitrogen fertilizer application demonstrating a positive correlation with rice spectral reflectance. This phenomenon may arise from an optimal nitrogen supply, which enhances rice growth and thereby results in increased reflectance levels.

3.1.2. Screening of Characteristic Variables for LNC Estimation

The PCC and XGBoost algorithms were employed to identify the characteristic variables for estimating LNC. Following numerous experiments, the screening criteria for PCC and XGBoost were established as |r| > 0.76 and FI > 0.01, respectively. Stepwise regression was then utilized to mitigate multicollinearity among the selected characteristic variables (Figure 6). Note that the characteristic variables identified by both algorithms exhibited some degree of overlap (band_750.92, PSSR, and RARS).

3.2. Estimation and Accuracy Assessment of LNC

The characteristic variables obtained solely from PCC (Figure 6b), solely from XGBoost (Figure 6a), and from their combination (band_756.40, band_770.11, band_742.71, band_569.61, band_731.79, band_671.94, band_463.34, band_495.06, band_680.08, band_880.59, band_500.37, band_750.92, PBI, LCI, PSSR, and RARS) were utilized as independent variables, respectively, with LNC serving as the dependent variable to construct the estimation models of LNC using PLSR, RF, and DNN algorithms (Figure 7). For the linear models (PLSR), the model based on the characteristic variables derived from the linear screening algorithm (PCC) exhibits a scatter distribution closer to the 1:1 line compared to models built with characteristic variables from the XGBoost algorithm, indicating a superior estimation performance of the former. Conversely, for nonlinear models (RF, DNN), the model constructed using the characteristic variables from XGBoost demonstrates a better performance. Furthermore, models based on the characteristic variables obtained from the combination of PCC and XGBoost consistently yield improved estimation results (R² > 0.8). Among these, the DNN model exhibits the best performance for estimating the LNC, with a validation R² of 0.89 and an RMSE of 0.17%.

3.3. Mapping LNC at the Field Scale

Previous research has indicated that nitrogen-monitoring models based on the entire growth period are applicable for nitrogen monitoring across all growth stages [8]. Therefore, we adopted the optimal model (PCC-XGBoost-DNN) to map the spatial distribution of LNC during various growth stages at the field scale (Figure 8). LNC was higher during the tillering stage compared to the jointing-to-booting and heading-to-flowering stages. Moreover, LNC exhibited an increasing trend corresponding to fertilizer application, which is consistent with the distribution pattern of the sample data. These findings offer valuable insights for subsequent precision fertilization strategies in rice cultivation.

4. Discussion

As the global population continues to expand, the world faces significant challenges in mitigating the environmental repercussions of excessive nitrogen inputs while striving to maintain high crop yields [28]. China’s annual nitrogen fertilizer input for rice cultivation is reported to reach 6.3 million tons, representing approximately one-third of the world’s total rice nitrogen fertilizer consumption. However, nitrogen utilization efficiency stands at less than 50%, resulting in a loss of over half of the applied nitrogen to the environment. This phenomenon adversely impacts air and water quality, posing threats to environmental safety and public health [29,30]. Given the pivotal role of nitrogen information in assessing rice growth status and yield, real-time access to such data serves as a crucial reference for optimizing nitrogen fertilizer application and enhancing nitrogen utilization efficiency [31]. UAV hyperspectral technology has proven to be a vital tool for obtaining this information [6].

Scholars have achieved successful LNC estimations using both linear [32] and nonlinear [20,33] screening algorithms. This prompted us to speculate on the existence of a blending relationship (linear and nonlinear) between nitrogen and the corresponding spectral information. Thus, we conducted an experiment to validate this hypothesis. We combined a correlation analysis with XGBoost to screen characteristic variables and compared the results with those obtained using a single-method approach. Our findings indicate that the feature variables identified through the combined correlation analysis and XGBoost algorithm significantly enhance rice nitrogen inversion, offering a novel approach to bolstering the accuracy of rapid crop nutrition diagnosis. This hints at the potential of the multi-method ensemble selection of characteristic variables in regression problems, which is consistent with former research [10,11]. Furthermore, among all the tested models, the DNN model emerged as the most effective in terms of LNC estimation. This may be attributed to the DNN’s inherent nonlinear mapping ability and its resilience to noisy data, enabling it to handle complex datasets better than shallow modeling approaches. The study offers a valuable reference for diagnosing the nutritional status and optimizing nitrogen fertilizer management in rice, which plays a crucial role in ensuring the sustainability of crop production.

In this study, we only utilized the cultivar ‘Meixiangzhan 2’ as the test material. Thus, further work is required to ascertain the method’s applicability to other rice varieties. In addition, South China belongs to the typical oceanic subtropical monsoon climate influence zone, with frequent cloudy and rainy weather. Due to constraints such as weather conditions and collection costs, the dataset utilized in this study remains relatively limited, potentially introducing some uncertainty into the results. To address this limitation, future research will explore the integration of data generation algorithms such as generative adversarial networks (GAN) to augment the dataset and enhance model robustness. Furthermore, this study solely focuses on estimating LNC. Subsequent research will integrate spatial nitrogen distribution data, expert insights, and regional soil characteristics to formulate a prescription map for rice fertilizer application. This holistic approach aims to facilitate precise fertilizer application, thereby enhancing fertilizer reductions and efficiency.

5. Conclusions

The precise determination of pertinent characteristic variables is key to developing accurate estimation models for LNC. In this study, we propose a novel approach that combines the strengths of PCC and the XGBoost algorithm to comprehensively screen for the optimal characteristic variables for LNC estimation. Utilizing these selected variables and field observations of LNC, we subsequently developed LNC estimation models. The most accurate estimation model was then adopted to explore the possibility of spatially mapping the LNC at the field scale. The results reveal that the integration of PCC and XGBoost enables a more meticulous screening of characteristic variables than solely adopting single-class methods. Furthermore, based on the RMSE values derived from test datasets, the DNN was determined as the most accurate model at the sample-point level. The proposed method offers the potential to map the LNC at the field scale using UAV hyperspectral imagery. The estimation accuracy of LNC was enhanced by combining PCC and XGBoost for characteristic variables screening, offering insights into optimizing nitrogen fertilizer’s application in rice cultivation.