Automatic Modeling Prediction Method of Nitrogen Content in Maize Leaves Based on Machine Vision and CNN

Abstract

Existing maize production is grappling with the hurdles of not applying nitrogen fertilizer accurately due to subpar detection accuracy and responsiveness. This situation presents a significant challenge, as it has the potential to impact the optimal yield of maize and ultimately, the profit margins associated with its cultivation. In this study, an automatic modeling prediction method for nitrogen content in maize leaves was proposed based on machine vision and convolutional neural network. We developed a program designed to streamline the image preprocessing workflow. This program can process multiple images in batches, automatically carrying out the necessary preprocessing steps. Additionally, it integrates an automated training and modeling system that correlates the images with nitrogen content values. The primary objective of this program is to enhance the accuracy of the models by leveraging a larger dataset of image samples. Secondly, the fully connected layer of the convolutional neural network was reconstructed to transform the optimization goal from classification based on 0–1 tags into regression prediction, so that the model can output numerical values of nitrogen content. Furthermore, the prediction model of nitrogen content in maize leaves was gained by training many samples, and samples were collected in three key additional fertilizing stages throughout the growth period of maize (i.e., jointing stage, bell mouth stage, and tasseling stage). In addition, the proposed method was compared with the spectral detection method under full-wave band and characteristic wavelengths. It was verified that our machine vision and CNN (Convolutional Neural Network)-based method offers a high prediction accuracy rate that is not only consistently better—by approximately 5% to 45%—than spectral detection approaches but also features the benefits of easy operation and low cost. This technology can significantly contribute to the implementation of more precise fertilization practices in maize production, leading to potential yield optimization and increased profitability.

1. Introduction

Nitrogen (N), as a nutrient element with the largest demands, plays a vital role in the growth and development of maize, and is also the key to characterizing growth conditions and photosynthesis of maize plants [1,2]. Nitrogen deficiency leads to short plant height and slow growth of maize, whereas excessive nitrogen causes lodging and heat injuries in maize plants. Leaves can reflect nitrogen content in maize plants effectively. Real-time accurate acquisition of nitrogen content information in maize leaves, as an important condition for scientific management, is conducive to implementing on-demand fertilization and increasing yield and production profits of maize [3,4,5]. Traditional N content detection methods in plants, such as the Kjeldahl method and the Dumas method, are mainly chemical ones [6]. Chemical methods are characterized by high accuracy and reliability. However, these methods have complicated detection processes and require professionals in complex operations. Consequently, they are primarily employed as a benchmark to calibrate other detection methodologies [7]. The most important disadvantage of these methods is that they are invasive and time-consuming (considerable time lag between sampling and result acquisition). In recent years, spectrum technology has advanced, which provides a new way for N content detection in maize plants. The spectroscopy and remote sensing technology are superior to chemical methods in terms of easy operation, analysis speed, and accuracy [8,9,10,11]. For example, based on spectral remote-sensing technology, Dayananda constructed a model for the relationship between N content and the growth condition of maize plants [12]. Wang Lifeng utilized the successive projections algorithm (SPA) to identify the characteristic wavelengths that are representative of the biochemical and biophysical properties of maize leaves specifically during their jointing stage. Under this wavelength, they constructed a mathematical model of N content in maize leaves and spectral reflectivity, which achieved good prediction results, The SPA-based PLS (Partial Least Squares) model showed strong predictive accuracy with R_c² = 0.944 and R_p² = 0.749, greatly simplifying the model by reducing 95.07% of the variables [13]. Some imaging detection methods that are used in leaf detection include hyperspectral imaging, near-infrared imaging, and magnetic resonance imagers. However, it should be noted that the equipment costs for spectroscopy detection methods can be significant, particularly when specialized equipment is required and operated by trained personnel, while subsequent data analysis is very complex. As pointed out by Taheri–Garavand, the cost factor is compensated by employing cameras in the visible portion of the electromagnetic spectrum. However, such methodology requires not only positioning of the leaf at a specific orientation in relation to the camera but also defined illumination conditions. Apart from that, the required specific illumination limits the method applicability to controlled-light environments. Advantages of this method include affordability/cost-effectiveness, portability, and provision of rapid measurements [14]. Over the last decade, Convolutional Neural Networks (CNNs) have been increasingly employed in plant phenotyping community. They have been very effective in modeling complicated concepts, owing to their ability to distinguish patterns and extract regularities from data. Examples include variety identification in seeds [15] as well as the identification of plant varieties based on the morphological analysis of leaves from intact plants [16]. Detection of nutrients including N contents in plants based on machine vision is appreciated increasingly by industrial experts and farmers due to the low cost and easy operation [17,18].

In N content detection, establishing a model of the relationship between detection objects and N content is vital to ensure high detection accuracy [19,20]. Traditional modeling techniques, such as multiple linear regression (MLR) and backpropagation (BP) neural networks, cannot meet the demands of researchers on prediction accuracy increasingly. Recently, deep learning based on big data has provided a strong guarantee of accurate modeling prediction [21,22,23]. Deep learning algorithms, represented by CNN have attracted the extensive attention of experts in various industrial sectors and become widely applied [24,25,26,27]. In the training and processing of many data models, these methods achieve higher efficiency than traditional machine learning algorithms, such as fully connected neural networks. Thus, they are highly appreciated by researchers engaging in maize detection, aiming to improve modeling accuracy through a increase in training sample size [28,29,30]. Nevertheless, deep learning algorithms are generally used for classification and identification, rather than accurate classical prediction. For example, Lu Hao trained images of maize tassels in different states by using a deep learning algorithm to recognize maize development state automatically and provide guidance to irrigation and fertilization [31]. If the deep learning algorithm based on CNN can be applied to construct a model of image and N content data output, the accuracy of such algorithms has the potential to be enhanced, particularly if the model architecture and training processes are carefully optimized. However, the degree of improvement is subject to vary depending on specific dataset characteristics and application contexts. Moreover, these deep learning algorithms tend to require preprocessing of images (e.g., segmentation and normalization) in the process of training and even prediction, increasing operation complexity. The sample size is restricted by the workload of image preprocessing. Thus, modeling accuracy cannot be improved by increasing data size [32].

In order to address the above problems, an automatic modeling prediction method of nitrogen content in maize leaves based on machine vision and CNN was proposed using maize leaf samples which were collected by machine vision equipment. In this method, programs of image preprocessing and data extraction were combined and the model could be constructed quickly by inputting images. Beyond that, prediction results can be gained directly by inputting one image. The proposed method simplifies the modeling process significantly and the goal of improving modeling accuracy through an increase in sample size is realized. This method can predict N content in maize leaves quickly and accurately at a lower equipment cost, laying a good foundation to promote the application of the proposed method in the agricultural field.

2. Materials and Methods

2.1. Experiments and Materials

The experimental field is located in Baoding, Hebei Province, China. Zheng Dan 958 was adopted as a plant cultivar. Our team was responsible for cultivation in 2021. Irrigation and other work were completed with sensor monitoring. The maize planting area is presented in Figure 1. Block-based control variable method was used to perform fertilizing and additional fertilizing. Different doses of nitrogen fertilizer were fertilized in different sections. Moreover, N content in maize leaves was controlled to improve modeling accuracy.

Our experiment, guided by skilled agricultural staff, divided the experimental field into 20 test blocks, each with an area of about 120 square meters (0.2 acres). To simulate varying fertilization conditions, we established a gradient of nitrogen application ranging from low to high. The first test block was treated with 1 kg of nitrogen as a starting point. Subsequently, for each additional block, the amount of nitrogen was increased by 0.2 kg, culminating with a nitrogen application of 4 kg in the final block. This design was intended to investigate the effect of varying nitrogen levels on maize growth and effectively collect maize leaf samples across a spectrum from low to optimal to high nitrogen availability.

In order to ensure that the proposed prediction model can provide accurate detection data support to additional fertilization of maize plants, about 1000 maize leaf samples were collected from blocks with different fertilizer contents in three key stages (jointing stage, bell mouth stage, and tasseling stage), respectively. A total of 3102 maize leaf samples were collected in the laboratory. The ground data of the samples can be seen in

Table 1. Temperature and soil relative humidity record.

	Date	Temperature/°C	Soil Relative Humidity/%
Jointing Stage	4.13	22.5	70
	4.14	23.7	76
	4.15	23.3	83
	…	…	…
Bell mouth Stage	5.07	25.1	75
	5.08	27.7	78
	5.09	25.8	86
	…	…	…
Tasseling Stage	6.19	31.7	76
	6.20	30.6	81
	6.21	33.2	78

.

Each leaf sample was numbered for convenience in the corresponding modeling between leave images and N content values in detection leaves in a professional laboratory. In the whole samples, leaf N content ranged from 17.93 to 26.55 mg·g⁻¹ in the jointing stage, 15.38 to 22.07 mg·g⁻¹ in the bell mouth stage, and 12.83 to 20.75 mg·g⁻¹ in the tasseling stage.

2.2. Image Acquisition and Pre-Processing

Canon EOS70D camera, manufactured by Canon Inc. in Tokyo, Japan, was used as the image acquisition equipment to predict N content. In our research methodology, the primary programming and model establishment are conducted using Python 3.6. Due to its extensive library support and flexibility, Python plays a central role during the data modeling stage, responsible for automating script writing, constructing, and validating models. Furthermore, Python’s various data science libraries are extensively utilized in the data preparation stage to facilitate effective data preprocessing and feature engineering. When it comes to the post-processing steps of model predictive results, we opt to use Matlab 2018a for its robust capabilities in mathematical computations, to delve deeper into the interpretation of model outputs, and to perform necessary statistical tests.

In the collection of maize leaf sample images using a camera on sunny daytime with good lighting conditions, the whole leaf sample was put to fill in the whole field of view as much as possible, in order to prevent interferences of other objects in the image to modeling, such as other vegetation, soil, and artificial objects. Some images collected by the camera are presented in Figure 2.

2.3. Automatic Modeling Prediction Based on CNN

The classification is performed in the fully connected (FC) layer of the CNN, which leverages the characteristics of the input data to categorize the images effectively. Based on classification results, the probability of allocating samples to different classes is calculated. In this study, the objective function in the last layer of the model is based on variable planting maize rather than regression values. The value range of N content in collected leaf samples was used as data classification labels. After processing these data classification labels, the target values of different classes that are divided according to the probability of output vector classification were gained, which are the desired predictions of N content. In this way, a regression prediction was realized. Among the collected 3102 maize leaf samples, 80% were used as the training set and 20% were used as the validation set.

2.3.1. Automatic Data Processing

In this study, we analyzed maize leaf images as the primary data source. From these images, we extracted several key parameters: the average grayscale value, the mean of the RGB (Red, Green, Blue) components, and the standard deviation and coefficient of variation for the RGB values. These parameters served as the input variables for our predictive model. The model was tasked with estimating the nitrogen (N) content in the maize leaves, which we defined as the model’s output variable. The modeling steps are as follows:

• Inputting images that are collected by a camera.
• Image preprocessing
• Inputting images into the program to obtain intermediate data: mean grey value, RGB component mean, RGB standard deviation, and RGB variable coefficient. These data are used as inputs of the model.
• Training the model by using the inputs of models in Step (3) and the laboratory detected N content in maize leaves.

Following the data collection in Step (1), we developed an integrated program that consolidates Steps (2) to (4) into a comprehensive workflow. Consequently, the system is capable of performing three automated procedures—batch preprocessing, extraction of model inputs, and construction of models incorporating nitrogen content—once the images are uploaded into the program. Contrarily, in conventional practices, models typically rely on a single color channel due to constraints in modeling efficiency, which in turn compromises accuracy. The method proposed in this paper can not only decrease manual workload but also realize the goal of improving modeling accuracy by increasing training samples. So we can use RGB images directly to improve modeling accuracy.

One maize leave image was a piece of 200 pixels × 200 RGB pixels image using a 6 × 6 × 3 convolutional kernel (6 × 6 filters with three channels). The convolution kernel size was 3 × 3, indicating that there were convolution kernels of three channels in different colors. A 4 × 4 characteristic pattern was generated. Through gray processing of RGB images, the average gray value of maize leaves was gained. At this moment, it was transformed into a convolution kernel of a single channel, through which the average gray value, RGB mean, RGB component standard deviation, and RGB variable coefficient were obtained.

2.3.2. Training and Modeling

CNNs are a class of deep learning algorithms that have shown remarkable success in areas such as image recognition, classification, and analysis [33]. These networks are specifically designed to process data that come in the form of multiple arrays, such as 1D for time series, 2D for images, or 3D for video data. The architecture of CNNs is inspired by the organization of the animal visual cortex and is particularly adept at capturing spatial hierarchies in data. The fundamental elements of a CNN include convolutional layers, pooling layers, and fully connected layers—each contributing toward the feature extraction and pattern learning capabilities of the network.

In the design of our proposed network, we embraced CNN’s powerful feature extraction and dimensionality reduction mechanisms to handle complex patterns in our dataset. By leveraging proven CNN methodologies, we stand on the shoulders of decades of research that have optimized the processing of high-dimensional data. Therefore, the techniques employed, such as normalization, weight initialization, and pooling, follow a well-established theoretical framework conducive to CNN architecture and function.

The convolutional structure of the proposed method is as follows, to facilitate a coherent evaluation of indicators within the constrained unit system of the credit index, we initially normalized individual groups of numerical data. This process allowed for the alignment of sample distributions across a specific range while maintaining compatibility. Acknowledging the variability in units and scales, standardization of each indicator was imperative before their assessment could commence. Subsequently, to address the varying nitrogen content levels amongst maize leaf samples, we transformed the initial values into a unified numerical range. This normalization was not only crucial to homogenize the data but also a requisite step before constructing the mathematical model. Finally, we applied a comprehensive normalization to the entire dataset, ensuring a consistent scale and range. This pivotal procedure guaranteed that the data comparability and integrity were upheld, setting a solid foundation for the subsequent initialization of the weight matrix. Subsequently, the weight matrix was initialized. Beyond that, normally distributed noises with a standard deviation of 0.1 were added to improve training accuracy. The polarization was initialized and some small positive values were added to avoid dead nodes. “The tf.constant function was used to create a matrix of shape [1, 151]. In this matrix, all elements were set to a numerical value of 0.1.” In order to obtain more image information, one step was chosen at a time in the definition of the pooling layer: strides[1] = strides[2], “strides[1] = strides[2]” means that the stride value is the same for both horizontal and vertical movements of the kernel window during the pooling operation, by setting strides[1] = strides[2], the pooling operation takes the maximum value within a 2 × 2 filter window and moves by 2 pixels both horizontally and vertically across the image. This essentially downsamples the feature map, reducing its size by half. However, because the stride value is the same for both horizontal and vertical movements, the resulting downsampling does not alter the aspect ratio of the image. In other words, the same amount of information is retained in both dimensions, effectively preventing changes in the image size. In order to decrease parameters and thereby reduce the complexity of the system, pooling was applied for parameter sparsification. Here, the maximum pooling was used. The size and step length of the pooling kernel function were set to 2 × 2 and 2, respectively. Through x_image, the original data were transformed into 6 × 6 two-dimensional images.

In convolutional neural networks, each convolution kernel usually only processes single-channel information. Therefore, the original RGB image is grayscale processed and the number of channels is set to 1. The convolution layer is an important hierarchical structure used to extract image features. For the parameter setting of the convolutional layer, selecting the appropriate size of the convolutional kernel, the number of image channels, and the number of convolutional kernels are the key factors. The following reasons are mainly taken into account for the parameter setting of the convolutional layer. Convolution kernel size: The choice of a 2 × 2 convolution kernel size can capture more intricate image features while avoiding high parameter counts caused by excessively large convolution kernel sizes. In addition, a 2 × 2 size is used in another pooling layer, which helps maintain consistency and stability between layers. Image channels: After grayscaling, the original RGB image has been transformed into a redundant three-channel image. Therefore, setting the image channel to 1 can save training and computing resources, and the value of the channel does not affect the model’s classification ability. Number of convolution kernels: The selection of the number of convolution kernels needs to consider the computational capacity of each kernel and the feature differences among different kernels. In this paper, choosing 16 and 32 convolution kernels can guarantee the effectiveness and accuracy of the model while minimizing the complexity of the convolutional layer, and fewer convolution kernels are also beneficial for reducing overfitting problems Convolutional kernel and offset: In the convolutional layer, matrix convolutional operation mainly involves the calculation of convolutional kernels and offsets. Convolutional kernels are used to capture features in the image, while offsets can be used to adjust model bias or increase anti-noise ability. When setting parameters such as convolution kernel size, image channels, and the number of convolution kernels, it is necessary to adjust the value of the offset appropriately to ensure the accuracy and generalization ability of the model. Therefore, the first convolutional layer was added, in which the convolution kernel size, number of image channels, number of convolution kernels, and corresponding offset were set to 2 × 2, 1, 16, and 16, respectively. Later, the second convolutional layer was added, in which the convolution kernel size, number of image channels, number of convolution kernels (multiplied by 16), and corresponding offset were set to 2 × 2, 16, 32, and 32, respectively. The third convolutional layer, namely a FC layer, was added, in which 4 × 4 × 64 (height) three-dimensional images were pulled into 512 long one-dimensional arrays. Finally, the output layer I was added, and a 512-long one-dimensional array was compressed into a 1 long array. The offset was set as 1.

The Rectified Linear Unit (ReLU) layer implemented nonlinear mapping to outputs of convolutional layers. The calculation formula could be expressed as:

$$f(x)=max(0, x)$$

(1)

The above structure of CNN proposed is shown in Figure 3.

The hyperparameters for initializing the CNN model were carefully selected based on preliminary experiments aimed at optimizing model performance. Learning rate and training epochs were set at 0.01 and 10,000, respectively. This particular learning rate was determined after evaluating various values and observing that 0.01 offered a sufficient rate of convergence without causing instability in the learning process. Furthermore, to prevent overfitting, we implemented an iteration stop condition utilizing early stopping based on validation loss. If no improvement in the loss was seen over 100 consecutive epochs on the validation set, training would terminate. One loss value was output every 100 times of training. Sample training results are presented in Figure 4. The loss value decreased continuously and it approached 0 with the increase in training times, while the accuracy increased continuously to 100%. This verifies that the proposed method achieves a good modeling effect.

The finished model was stored and called directly at the prediction step. It outputs N content in maize leaves directly by inputting images of detecting maize leaves.

2.4. Prediction

In the modeling process of this article, a total of 3102 samples were used for training and testing. Deep learning was carried out on the training set, with a train-to-test ratio of 8:2. To measure the performance of the proposed modeling method, 30 new samples were collected from different blocks in the jointing stage, bell mouth stage, and tasseling stage of maize plants. Furthermore, these 90 new samples were used as the test and predicted using different methods. Prediction accuracy was obtained by comparing detection results with prediction results of different methods. The calculation formula for prediction accuracy could be expressed as:

$$\mathit{Accuracy}=(1-\frac{\left|y_p-y_c\right|}{y_c})\times 100\%$$

(2)

where y_p represents the prediction results of different methods and y_c represents laboratory detection results.

If the predicted result y_p is exactly the same as the laboratory test result y_c, the accuracy is 100%. If it is higher or lower than y_c, the accuracy will decrease accordingly.

In order to verify the advantages of the method proposed in this paper, the spectral detection method under full-wave band and under the characteristic wavelength of N content in maize leaves were treated as comparison methods. Prediction accuracies of the proposed method to samples collected in three growth stages are shown in Figure 5a–c. The prediction accuracy of the proposed method ranges from 93% to 99% among the three growth stages. For contrast verification, the above samples were analyzed using the spectrum detection method. Light reflectance information of maize leaves was collected with a PSR-1100F spectrometer with a wavelength ranging from 320 nm to 1100 nm. The prediction accuracy of the light reflectivity and N content model of maize leaves which is constructed with the MLR method under full-wave band to samples in three stages are shown in Figure 6a–c (85–95%). The prediction accuracies of the multiple regression model to samples in three stages under the characteristic wavelength of N content in maize leaves are presented in Figure 7a–c.

In order to analyze the application and convenience of different methods in agricultural practices, 30 samples were selected from 90 samples randomly. The prediction accuracy of the proposed model for these 30 samples is shown in Figure 5d. The prediction accuracy of the spectral detection-MLR model under a full-wave band is presented in Figure 6d. Given that characteristic wavelengths of N content in maize leaves vary in different stages, three stages correspond to MLR models under three different characteristic wavelengths. The characteristic wavelengths were selected according to the introduction in Reference [34]. The characteristic wavelengths are shown in

Table 2. Characteristic wavelengths of N content in maizes leaves vary in different stages.

Stage	Characteristic Wavelength Number	Characteristic Wavelength/nm
Jointing Stage	7	321, 349, 509, 633, 690, 901, 1083
Bell mouth stage	7	321, 510, 603, 684, 821, 894, 1076
Heading stage	7	323, 344, 529, 610, 690, 764, 854

. The prediction accuracies of MLR models under characteristic wavelengths to 30 random samples in the jointing stage, bell mouth stage, and tasseling stage are presented in Figure 7d–f.

The mean accuracies of different models to samples in three stages are shown in

Table 3. Mean prediction accuracy of different models.

Accuracy (%)		Jointing Stage	Bell Mouth Stage	Tasseling Stage	Random Samples
The proposed method		96.46400	95.75625	95.80349	96.25322
MLR model under full wave band		90.59979	89.79837	91.54138	90.85356
MLR model under characteristic wavelength	Jointing stage model	94.16378			48.69417
	Bell mouth stage model		94.33698		46.63486
	Tasseling stage model			94.86209	51.53014

.

3. Results

The proposed modeling method and experiment process can be seen in Figure 8. A total of 3102 maize leaf samples had been collected. We collected maize leaf sample images under sunny conditions to ensure good lighting, angling the camera to capture the entire leaf and avoid background interference. We developed an automatic method for predicting leaf nitrogen content using CNN-based machine vision. Our validation involved 90 new samples from different maize growth stages (jointing, bell mouth, and tasseling), using 30 samples from each to compare various prediction methods.

As shown in Figure 5, the prediction accuracy of the proposed method ranges from 93% to 99% among the three growth stages. It can be observed from Table 3 that the average prediction accuracy of the proposed method to samples in three stages is about 95%. The prediction accuracy of the proposed method to random samples is basically consistent with the prediction accuracy to samples in three stages. However, as revealed in Figure 6, the prediction accuracy ranges from 85% to 95%, while the average prediction accuracy of the spectral detection-MLR model under full-wave band to samples in three stages is about 90%. The prediction accuracy of the spectral detection-MLR model under full-wave band to random samples is basically consistent with the prediction accuracy to samples in three stages. As shown in Figure 7, the prediction accuracies of the spectral detection-MLR model under characteristic wavelength to samples in three stages are relatively high, about 94% on average. However, its prediction accuracy to random samples fluctuates greatly and the average prediction accuracy is about 50%.

4. Discussion

The CNN-based machine vision technique we have developed stands out in comparison to recent research due to its exceptional accuracy and stability across various growth stages of maize leaves. It achieves a predictive accuracy ranging from 93% to 99% within the jointing, tasseling, and silking stages, with an average prediction accuracy of approximately 95%. This is notably superior to methods employed by Silva et al. (2024) [35], which reported only a modest correlation coefficient of around 0.6 and an MAE (Mean Absolute Error) below 0.5 when predicting chlorophyll content using random forest models, failing to match the precision achieved by our CNN approach. Additionally, Cao et al. (2021) [36] demonstrated the effectiveness of combining dimensionality reduction techniques with different regression methods for hyperspectral data analysis, highlighting the EN-PLSR (Elastic Net Partial Least Squares Regression).model as the most accurate option with an R² value of 0.96 and RMSE value of 0.19. Despite their promising results, our CNN model offers comparable or even superior accuracy without requiring complex data preprocessing steps typically associated with traditional spectral detection and dimensionality reduction methods.

Furthermore, our method consistently demonstrates strong predictive capabilities throughout different growth stages while maintaining model stability that promises practical application benefits in agricultural monitoring tasks. In contrast, although the PLSR model used by Cao et al. shows high R² values, its real-world implementation may involve more intricate data preprocessing steps compared to our CNN approach which directly processes image data without complicated spectral data dimensionality reduction requirements; thus holding significant promise for streamlined predictive analysis in agricultural monitoring.

Given that the collected data are within the whole wave band, the spectral detection-MLR model under the full-wave band does not need modeling for different stages. It is applicable to samples in different growth stages and random samples. However, it involves various inputs that are spectral reflectivity under different wavelengths in the full-wave band of the spectrograph, which increases modeling complexity. Moreover, prediction accuracy declines due to multicollinearity. Thus, the prediction accuracy fluctuates around 90%.

Characteristic wavelengths of N content in maize leaves are different in the jointing stage, bell stage, and tasseling stage. Therefore, corresponding models are constructed for these three stages according to the introduction in Reference [34]. In each stage, light reflectivity under only 7 characteristic wavelengths is used as the input, determining low model complexity. Moreover, the constructed models could fully reflect the accuracy of spectrum detection and achieve high prediction accuracies (about 94%) in the corresponding stage. The prediction accuracy of the spectral detection-MLR model under characteristic wavelength is higher than that of the spectral detection-MLR model under full-wave band. Given that characteristic wavelength is different in different stages, the constructed models are only effective for samples in the corresponding stage. Therefore, the prediction accuracies of models for three stages to random samples are relatively low, which is similar to the random distribution prediction effect. Hence, different characteristic wavelength models should be constructed for samples in different stages, restricting their applications to the agricultural field to some extent. Yu Fengh, et al., (2022) [37] proposed remote sensing of nitrogen content in rice leaves to realize precision fertilization. However, it serves as a hyperspectral vegetation index for the rapid inversion method with costly equipment and indirect detection. Apart from that, Hong Bo, et al. [38] proposed a digital imaging detection method for Nitrogen content in cotton leaves, but this method is based on linear regression and the result will be poor when the number of samples is large.

On a commercial scale, a capital investment will be initially required to adopt the employed approach [39]. Nevertheless, maybe the wide-ranging large-scale commercial applications will be able to provide high returns through considerable improvements in process enhancement and cost reduction. The proposed method in this paper will collect images using machine vision. The image information will cover information in the whole wave band. Therefore, it is not necessary to distinguish samples in different stages. In comparison to the other two models, the proposed method achieves relatively higher prediction accuracy (greater than 93% and about 95% on average) for both samples in specific stages and random samples. The proposed method, which can predict N contents in maize leaf samples in different stages accurately, is simpler than the spectral detection-MLR model under characteristic wavelengths.

The accuracy of the proposed method is higher than that of the two spectrum detection methods. One important reason is that it collects abundant samples as the training set, but previous image processing and modeling methods have low degrees of automation. Researchers should preprocess samples before modeling, which requires considerable workloads. The proposed method integrates image preprocessing and data extraction programs, which allow automatic batch preprocessing of sample images and direct modeling by batch input of images. Therefore, it can realize the goal of improving modeling accuracy through the increase in sample size. Moreover, N content can be detected directly by simply inputting images. This is convenient to be extensively applied in agricultural practices.

With another advantage of low equipment cost, the proposed method in this paper only requires a machine vision device. However, the cost of spectrographs presents a significant barrier, which limits the widespread application of spectrum detection methods in agriculture. The proposed method is characterized by high accuracy, simple modeling, low operation complexity, and low construction cost.

For future research, the continuous advancements in machine vision and artificial intelligence offer immense potential. Exploring the application of deep learning, particularly CNNs could enhance the accuracy and efficiency of the proposed method. Furthermore, expanding the model’s applicability to encompass a wider range of plant health indicators remains a crucial area for investigation. Researchers might also evaluate integrating this method with existing agricultural technologies to create a more comprehensive approach to precision farming. From a commercial perspective, transitioning from research to practical applications is pivotal. The potential integration of this machine vision system with drones or portable devices opens up possibilities for on-the-go plant health monitoring and precision fertilization solutions. This technology has the potential to become an essential component of an advanced agricultural ecosystem when combined with user-friendly software interfaces that enable farmers to make data-driven decisions rapidly. Additionally, lower costs associated with machine vision systems compared to traditional spectrograph equipment could significantly reduce barriers to entry, making this technology accessible to a broader range of agricultural businesses. Lastly, incorporating this proposed method into IoT (Internet of Things) -enabled smart farming systems signifies significant progress toward automated and intelligent crop management. By leveraging the interconnectivity provided by IoT, data from machine vision systems can be analyzed and acted upon in real-time, fostering an environment for enhanced productivity and resource optimization.

In conclusion, while the proposed method already exhibits high accuracy, simple modeling, and low operation complexity, its full potential lies in the ease with which it can be scaled up and adapted for future research endeavors, commercial ventures, and practical agricultural applications. The integration of this method could lead to significant improvements in crop management and the overall efficiency of agricultural operations, with substantial economic benefits for the agricultural sector. Therefore, we recommend a concerted effort to explore these avenues to ensure that the proposed method enjoys a broad and impactful implementation.

5. Conclusions

Nitrogen, a nutrient element with the largest demands, plays a vital role in the growth and development of maize. In N content detection, establishing a model of the relationship between detection objects and N content is crucial for protecting detection accuracy. In this study, an automatic modeling prediction method of nitrogen content in maize leaves based on machine vision and a convolutional neural network was proposed. In this method, programs of image preprocessing and data extraction were combined and the model could be constructed quickly by inputting images. Apart from that, prediction results could be obtained directly by inputting one image. The proposed method simplifies the modeling process significantly and realizes the goal of improving modeling accuracy through an increase in sample size.

In order to verify the advantages of the method proposed in this paper, the spectral detection-MLR method under full-wave band and under the characteristic wavelength of N content in maize leaves have been treated as comparison methods. Based on the comparison of the results, it can be seen that the method proposed in this paper is characterized by high accuracy, low equipment cost, and simple modeling. It may lay a solid foundation for N content detection in maize leaves and provide strong support for fast and accurate fertilization as well as high yield and high production profits of maize.