Research Progress in Intelligent Diagnosis Key Technology for Orchard Nutrients

Abstract

The intelligent diagnosis key technology of orchard nutrients provides a decision-making basis for precision fertilization, which has important research significance. This article reviewed the recent research literature, compared and analyzed existing technologies, and summarized solved and unresolved problems. It aimed to find breakthroughs to further improve the level of intelligent diagnosis key technology for orchard nutrients, and promote the implementation and application of the technology. Research had found that the current rapid nutrient detection technologies were mostly based on spectral data, with a focus on preprocessing algorithms and regression models. Hyperspectral technology shows good performance in predicting tree and soil nutrients due to its large number of characteristic variables. Meanwhile, preprocessing algorithms such as filtering, transformation, and feature band selection had also solved the problem of data redundancy. However, there were few studies for small and trace elements, and field applications. Laser breakdown-induced spectroscopy has good prospects for soil nutrient detection, as it can simultaneously detect multiple nutrients. There had been some studies on the technology for generating suitable nutrient standards for orchards in terms of soil and tree nutrients, but it requires a long and extensive experiment, which is time-consuming and laborious. A universal and rapid method needs to be studied to meet the construction needs of suitable nutrient standards for different varieties of fruit trees.

1. Introduction

Unreasonable fertilization has always been present in orchard production management, which has a negative impact on the ecological and economic benefits. Therefore, it is necessary to continue to promote precision fertilization. Nutrient diagnosis is the basis for precise fertilization prescriptions. On the one hand, it should be based on the current status of tree nutrients and soil nutrients [1,2], and on the other hand, it should be based on the appropriate nutrient standards for different varieties, tree ages, and growth periods [3]. Through comparison, it can be used to diagnose nutrient richness and deficiency. Traditional methods are mostly based on experience to obtain appropriate nutrient standards, using chemical methods to determine nutrient status, which is time-consuming and laborious [4], difficult to apply on a large scale, and has poor accuracy. So, currently, nutrient diagnosis is mostly based on the entire orchard as a unit, without considering the differences in soil and fruit trees at different locations inside, which limits the development of precision fertilization technology.

The key technology of the intelligent diagnosis of orchard nutrients uses rapid nutrient detection technology to collect each tree and root soil nutrient information, improve data acquisition convenience, and increase data volume. In addition, it scientifically generates suitable nutrient standards based on the growth laws of fruit trees. In this way, each tree can be diagnosed and personalized fertilization can be guided, which is of great significance.

Many scholars have conducted extensive and in-depth research and achieved good results. We studied a large amount of the literature. Databases and journal websites were used to search for the literature, such as the China National Knowledge Infrastructure (CNKI), Web of Science, Elsevier ScienceDirect, SpringerLink, etc. Fruit nutrients, soil nutrients, the multispectral, hyperspectral, and fertilization model, and others were used as keywords for the search. A total of 83 relevant articles from the past five years were obtained, which were divided into three aspects according to the key technology of intelligent diagnosis. For the rapid detection of tree nutrients, the methods used were all based on optical principles, so they were classified into RGB (red, green, blue) images, and multispectral, hyperspectral, and combined spectra according to the type of sensor. For the rapid detection of soil nutrients, the methods used were not only based on optical principles, but also on electrochemical principles. Therefore, the category of electrochemistry was added. The differences between different methods are mainly reflected in the sample type, data sources, preprocessing algorithm, characteristic variable, predictive variables, regression model and results. Therefore, these data were highlighted for comparison. Techniques for generating suitable nutrient standards only target trees and soil; therefore, this was used as the classification basis.

The purpose of this article is to (1) summarize the advantages and disadvantages of different tree and soil nutrients rapid detection methods, identify methods with better predictive performance and their unresolved issues; (2) provide direction for subsequent research and promote the field application of rapid nutrient detection technology; and (3) analyze the shortcomings of existing techniques for generating suitable nutrient standards.

2. Rapid Detection Technology for Nutrients

2.1. Rapid Detection Technology for Tree Nutrients

At present, research on the rapid detection technology for tree nutrients was mainly based on RGB images, multispectral, hyperspectral, and combined spectra, and focused on preprocessing algorithms and regression models.

Based on RGB images, Salazar-Reque et al. explored different vegetation indices (VIs) extracted from aerial RGB images acquired in different flights to differentiate the nutritional and water statuses of Hass avocado plantations [1]. Zhang et al. applied a color factor combination regression model to Salix suchowensis Cheng images to characterize the canopy distribution of SPAD, achieving the visualization of the chlorophyll content in the distribution of the entire plant [5]. Wang et al. proposed a prediction model of nitrogen content in apple leaves based on combined color features to predict the nitrogen content at the flowering, young fruit, and fruit expansion stages [6]. Xu et al. proposed a model that utilizes image processing technology and machine learning techniques to enhance the accuracy of potassium detection in apple leaves based on MLR-LDA-SVM [7]. A specific comparison was presented in

Table 1. Key information of rapid detection technology for tree nutrients based on RGB images.

Sample Type	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
Canopy Hass avocado [1]	Collected by UAV with camera	CNN Otsu’s method	Vegetation Indices	Nitrogen		Better VIs: MGRVI
Canopy Willow [5]	Collected by Phenotype information collection platform with camera	YOLO V5 SNV	Color factor	Chlorophyll	FTM STM LTM	Best: R, G, B, G/R, G/B + LTM R2: 0.731, RMSE: 2.16
Leaf Apple [6]	Collected by camera	MSRx	R, G, B 14 combination	Nitrogen	PCA SVM BP ELM	Best: PCA-SVM MAE: ≤0.64 g/kg RMSE: ≤0.80 g/kg
Leaf Apple [7]	Collected by camera	Gaussian filter Threshold segmentation Canny operator MSR-color restoration	Color features Shape features	Potassium	MLR LDA SVM KNN DT	Best: MLR-LDA-SVM Average accuracy: 93.5%

.

Based on multispectral data, Costa et al. proposed a novel methodology to determine leaf nutrient concentrations in citrus trees using unmanned aerial vehicle (UAV) multispectral imagery and artificial intelligence [8]. Sun et al. used UAV multispectral indices to perform ordinary linear regression, multivariate stepwise regression, and ridge regression inversion models the on leaf NPK content [9]. Tang et al. designed a portable crop chlorophyll detection device based on ambient light correction [10]. Luo et al. collected citrus canopy spectral data, used a multispectral shadow index to remove canopy shadows and soil background, and predicted the chlorophyll content based on vegetation indices and texture features [11]. To overcome the influence of the environment and canopy structure on reflectivity, Cheng et al. used a 3D radiative transfer model and UAV multispectral imagery to estimate the canopy-scale chlorophyll content in apple orchards [12]. A specific comparison was presented in

Table 2. Key information of rapid detection technology for tree nutrients based on multispectral data.

Sample Type	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
Canopy Citrus [8]	Collected by UAV with multispectral camera		Red, Green, Blue, Red edge, Near-infrared	Macronutrients Micronutrients	GBRT	Average error: For macronutrients: <17% For micronutrients: <30%
Canopy Apple [9]	Collected by UAV with multispectral camera		Vegetation index	Nitrogen Phosphorus Potassium	OLR MSR RR	Better: MSR, RR LNC R2: 0.52–0.76 LPC R2: 0.67, 0.69 LKC R2: 0.76
Non woven fabric Impregnated chlorophyll [10]	Collected by multispectral sensor	LEC	NDVI	Chlorophyll		R2: 0.965
Canopy Citrus [11]	Collected by UAV with multispectral camera	NDCSI	Vegetation index Texture features	Chlorophyll	FSR PLSR DNN	Best: Deep neural network R2: 0.665 RMSE: 9.49 mg/m2
Canopy Apple [12]	Collected by UAV with multispectral camera	3D RTM LESS Linear interpolation	NDVI, Clgreen Clred edge, GNDVI	Chlorophyll	PJI	More robust VIs: GNDVI, CIred edge, CIgreen

.

Based on hyperspectral data, Yue et al. proposed a potassium content prediction method for citrus leaves based on hyperspectral and deep transfer learning [13]. Liu established a non-destructive monitoring model of the functional nitrogen content in citrus leaves during the expansion and transformation stages based on hyperspectral data [14]. Paz-Kagan et al. attempted to quantify soluble carbohydrates and starch in dried and powdered tissues of almond by visible-to-shortwave infrared reflectance spectroscopy [4]. Azadnia et al. established a novel approach for rapidly estimating the status of NPK in apple tree leaves based on visible/near-infrared spectroscopy coupled with machine learning [15]. Xiao et al. proposed a novel method for estimating the leaf chlorophyll content in citrus [16]. Ye et al. presented a rapid and non-destructive approach for the estimation and mapping of nitrogen content in apple trees at the leaf and canopy levels using hyperspectral imaging [17]. A specific comparison was presented in

Table 3. Key information of rapid detection technology for tree nutrients based on hyperspectral data.

Sample Type	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
Leaf Citrus [13]	Collected by spectrometer	SPA Wavelet denoising	Reflectance values	Potassium	SSAE–DLNs	R2: 0.8771 RMSE: 0.5528
Leaf Citrus [14]	Collected by spectrometer	SNV	Reflectance values	Nitrogen	PLSR SVM BP RF	Best: BP fruit expansion period: R2:0.78 fruit color-changed period: R2:0.74
Almond Dried and powdered [4]	Collected by spectrometer		Reflectance values	Carbohydrates	PLSR DA	Overall accuracy: >90% Unique spectral region: SWIR
Leaf Apple [15]	Collected by spectrometer	SNV MSC Spectral derivatives PLSR Random frog (Rfrog) VIP	Multi-band reflectance	Nitrogen Phosphorus Potassium	RF ANN SVM PLSR	MSC + D2-Rfrog-RF: rp = 0.985 SNV + D2-Rfrog-RF: rp = 0.977 SNV + D2-Rfrog-RF: rp = 0.978
Leaf Citrus [16]	Collected by spectrometer	FOD CWT	Dual-band Tri-band	Chlorophyll	HO PLSR	Optimal range: 550 nm, 750 nm Kurtosis: 3.2, skewness: 0.066
Leaf and canopy Apple [17]	Collected by spectrometer	First derivative	Reflectance values	Nitrogen	PLSR MLR	The MLR model based on the raw reflectance was better (4 key wavelengths, less data)

.

Based on the combined spectra, Wang et al. developed a rapid detection method for chlorophyll content and distribution in citrus orchards based on low-altitude remote sensing and bio-sensors [18]. Cao et al. proposed a method for estimating the N content of tea plants under field conditions based on the combination of a multispectral imaging system and hyper-spectral data [19]. In view of the uneven distribution of illumination on broad-leaved banana canopies in the field, An et al. proposed a correction method based on the dark channel prior to enhance the image quality, and improve the estimation accuracy of chlorophyll content [20]. A specific comparison is presented in

Table 4. Key information of rapid detection technology for tree nutrients based on combined spectra.

Sample Type	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
Leaf and canopy Citrus [18]	Multiplex 3.6 sensor FieldSpec4 radiometer Multispectral		Reflectance values RVI NDVI	Chlorophyll	ULR MLR PLSR	Based on RVI: R2: 0.7063 Based on NDVI: R2: 0.7343
Canopy Tea plant [19]	Hyperspectral Multispectral (spectrometer, multispectral camera)	CARS POA, VCPA	1664 nm, 1665 nm H, VOG, BGI	Nitrogen	SVM	R2: 0.9186 RMSE: 0.0560
Canopy Banana [20]	RGB-NIR (RGB camera, and NIR cameras)	CMDCP	Color features Vegetation indices	Chlorophyll	RF	RC2: 0.738, RMSEC: 6.296 RV2: 0.692, RMSEV: 7.357

.

2.2. Rapid Detection Technology for Soil Nutrients

Research on rapid detection technology for soil nutrients has been mainly based on electrochemistry, hyperspectral, laser-induced breakdown spectroscopy, and combined spectra, and has focused on preprocessing algorithms and regression models.

According on the methods of electrochemistry et al., an ion-selective electrode (ISE) is a quick and low-cost method for soil nitrate nitrogen (N) detection. The measurement accuracies of the linear regression, multiple regression and BP neural network models were compared. The results showed that the BP neural network model had the highest accuracy [21,22]. The standard addition method was used to predict nitrate nitrogen. Archbold et al. used an Ion-Sensitive Field-Effect Transistor (ISFET) sensor to estimate the pH and available nutrients in soils [2]. Jia et al. designed a real-time detection system for soil organic matter content based on the high-temperature excitation principle. This system mainly detected the carbon dioxide produced by heating the soil. The experimental results indicated that the prediction accuracy was highest when the heating depth and time were 15 mm and 20 s [23]. Li et al. used a method based on pyrolysis and artificial olfaction to predict soil total nitrogen. The model using the PLSR-BPNN exhibited better performance [24,25,26]. A specific comparison is presented in

Table 5. Key information of rapid detection technology for soil nutrients based on method of electrochemistry and others.

Method	Characteristic Variable	Predictive Variables	Regression Model	Results
Ion-selective electrode [21]	Reading of nitrate ISE	Nitrate nitrogen	BP	The average relative errors: 5.07% and 8.81%
Ion-selective electrode [22]	Reading of nitrate ISE	Nitrate nitrogen	SAM	R2: >0.9 RMSE: 2.21–5.49 mg/L
High-temperature excitation [23]	Carbon dioxide content	Organic matter	MLR	The 15 mm 20 s model had the highest accuracy, >90%,
Pyrolysis and artificial olfaction [24,25,26]	Artificial olfaction feature space	Total nitrogen	PLSR BP	R2: 0.92186 RMSE: 0.21781 RPD: 3.3426

.

Based on hyperspectral data, Munnaf et al. used online collected Vis-NIR spectroscopy to develop a novel soil fertility index. The results revealed that this method had very good accuracy in predicting the soil fertility index [27]. Wang et al. proposed a deep learning-based method for soil total nitrogen characteristic wavelength screening. A range of 8–50 characteristic wavelengths were screened, and the deep learning model incorporating the inception module performed well based on the characteristic wavelengths [28]. Wang et al. used an embedded method to select characteristic wavelengths, and convolution operations were used to predict the soil total nitrogen. Thus, the accuracy improved [29]. Wang et al. used the successive projection algorithm to extract wavelengths that were minimally redundant. The back-propagation neural network model was used to predict soil nitrogen, and a high prediction accuracy was obtained [30]. Guerrero et al. conducted a study to develop an automatic filtering system of very noisy and non-soil spectra. The results indicated that the machine learning algorithms provided high classification accuracies [31]. To solve the limitations generated by massive hyperparameters on CNN for the prediction of organic carbon and total nitrogen, Qiao et al. employed spectral data obtained by Fourier-transform NIR spectroscopy and used SVD-CNN [32]. Wan et al. proposed a spectral enhancement method based on a masked autoencoder to predict soil nutrients. This method can learn highly robust and generic spectral features from public NIR spectral datasets [33]. To overcome the influence of soil moisture, Lin et al. proposed a method based on mixture-based weight learning to predict the soil total nitrogen [34]. Tavakoli et al. preprocessed the visible and near-infrared spectra of soil using the dual-wavelength indices transformations and constructed a soil parameter prediction model based on stacking machine learning approaches [35]. Yang et al. extracted 272 hyperspectral bands using uninformative variable elimination and constructed a soil total nitrogen prediction model based on partial least squares regression [36]. The sensitive bands of soil total nitrogen content were extracted based on the Pearson correlation coefficient, and the random forest model was used to build the inversion model [37]. A specific comparison is presented in

Table 6. Key information of rapid detection technology for soil nutrients based on hyperspectral data.

Reference	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
[27]	Collected by on-lone soil sensing platform	Moving average Savitzky–Golay 1st-order derivative Smoothing	Vis-NIR reflectance values	Soil fertility index	PLSR LOOCV	RPD: 2.01
[28]	LUCAS 2009 TOPSOIL data		8–50 characteristic wavelengths reflectance values	Total nitrogen	MCCNII	R2: 0.93 RMSEP: 0.97 g/kg RPD: 3.85
[29]	LUCAS 2009 TOPSOIL data	Embedded method	16 characteristic wavelengths reflectance values	Total nitrogen	CO	R2: 0.86 RMSEP: 1.98 g/kg RPD: 1.89
[30]	Collected by spectrometer	Successive projections algorithm	Selected wavelengths reflectance values	Nitrogen	PLSR BP	Better: BPNN Wet soil, R2: 0.93, RMSEP: 0.0297%, RPD: 4.00 Dry soil, R2: 0.99, RMSEP: 0.0132%, RPD: 8.76
[31]	Collected by online multisensor platform	Similarity algorithms Machine learning algorithms				Filtering of very noisy and non-soil spectra can be achieved using machine learning algorithms
[32]	Collected by spectrometer LUCAS 2009 TOPSOIL data	Fourier transform	Near-infrared spectroscopy reflectance values	Organic carbon Total nitrogen	SVD-CNN	R2: 0.9304 for organic carbon R2: 0.9319 for total nitrogen
[33]	Collected by spectrometer LUCAS 2009 TOPSOIL data	Spectral enhancement method based on the masked autoencoder	Near-infrared spectroscopy reflectance values	Nitrogen Phosphorus Potassium		R2: Nitrogen 0.941 Phosphorus 0.926 Potassium 0.903
[34]	Collected by spectrometer	First derivative	Vis–NIR spectroscopy reflectance values	Total nitrogen	MBWL RF	R2: 0.757 RMSE: 0.235 g/kg Relative error: 10%
[35]	LUCAS 2009 TOPSOIL data	Dual-wavelength index transformations	Vis–NIR spectroscopy reflectance values	CaCO3 N OC	SML	RMSE/R2: CaCO3 25.71/0.96 N 1.11/0.92 OC 21.34/0.95
[36]	Collected by spectrometer	Factional-order derivative Uninformative variable elimination	272 bands reflectance values	Total nitrogen	PLSR	R2v: 0.7937 RMSEV: 0.1976 g/kg RPDV: 2.1904
[37]	Collected by UAV with hyperspectral camera	First derivative Pearson correlation coefficient	Sensitive bands reflectance values	Total nitrogen	RF	R2: 0.859 RMSE: 0.143 g/kg

.

Based on LIBS, Xu et al. applied a CNN to soil analysis based on laser-induced breakdown spectroscopy. The CNN model decreased the root-mean-square error compared with partial least squares [38]. Tavares et al. developed a prediction model for organic matter, extractable P, K, Ca, and Mg, using a regression model. The iSPA-PLS method was the best [39]. Li et al. used Raman spectroscopy for the rapid analysis of soil components, which has the characteristics of being less affected by moisture interference, less sample preprocessing, and complementary-to-infrared spectroscopy [40]. A specific comparison was presented in

Table 7. Key information of rapid detection technology for soil nutrients based on LIBS.

Reference	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
[38]	Collected by MobiLIBS system	MWPL	Reflectance values	pH SOM TN, TP, TK	CNN	Decreased the RMSEV by 1.48%, 4.97%, 9.56%, 10.05%, and 2.90%, respectively.
[39]	Collected by spectrometer	ISPA	Reflectance values	SOM Extractable P, K, Ca, Mg	MLR	RPD: >1.40

.

Based on the combined spectra, Xing et al. proposed a method to determine the soil organic matter combining FTIR-ATR and Raman spectroscopy. Competitive adaptive reweighted sampling was used to improve prediction accuracy [41]. Wang et al. designed a vehicle-mounted soil total nitrogen prediction system based on the fusion of near-infrared spectroscopy and image information [42]. Zhou et al. chose a catboost prediction model to predict the TN. Seven characteristic wavelengths were selected using uniform variable illumination and adaptive weighted sampling [43]. Nawar et al. evaluated the prediction accuracy of extractable potassium using portable Gamma-rays and X-ray fluorescence spectral data [44]. Wang et al. used a combination of soil spectral information and image features to construct a soil total nitrogen prediction model with spectral fusion [45]. A specific comparison is presented in

Table 8. Key information of rapid detection technology for soil nutrients based on combined spectra.

Reference	Data Sources	Preprocessing Algorithm	Characteristic Variable	Predictive Variables	Regression Model	Results
[41]	FTIR-ATR Raman (Raman spectrometer, infrared spectrometer)	Wavelet transform Fast Fourier Transform AIRPLS CARS	Reflectance values	Organic matter	PLSR	RMSEP: 4.35 g/kg
[42,43]	Near-infrared Image (sensor)	Uniform variable illumination Adaptive weighted sampling	7 characteristic wavelengths reflectance values	Total nitrogen	CPM	R2: 0.8 Relative error: <10%
[44]	Gamma-rays X-ray (gamma-ray system, XRF spectrometer)	Moving average Smoothing with SG Maximum normalization	Reflectance values	Extractable potassium	PLSR	R2: 0.75 RMSE: 31.3 mg/kg RPD: 2.03
[45]	Spectroscopy RGB (spectrometer, camera)	UVE CARS	7 characteristic wavelengths reflectance values Angular second moment Energy, Moment of inertia Gray mean, Entropy	Total nitrogen	CB	R2: 0.8668 RMSE: 0.1602 g/kg

.

3. Techniques for Generating Suitable Nutrient Standards

Research on suitable nutrient standard generation techniques can be divided into two main directions. One direction was to generate suitable nutrient standards for the soil based on the relationship between soil nutrients and yield. Another direction was to generate suitable nutrient standards for leaves based on the relationship between leaf nutrients and yield.

3.1. Techniques for Generating Soil Suitable Nutrient Standards

Ahmed et al. utilized years of soil nutrient and yield data and employed an improved genetic algorithm to construct a mapping relationship between soil nutrient and yield. A neighborhood-based exploration and development strategy was proposed to optimize the soil nutrient level and achieve maximum yield [46]. Gokalp developed an Approximate Dynamic Programming algorithm to obtain the best fertilization policy. The data used mainly included the fertilization amount, yield, and precipitation [47]. Zhang et al. studied the effects of nitrogen fertilizer varieties and dosages on the photosynthetic characteristics and yield of red dates, and obtained the best combination [48].

3.2. Techniques for Generating Leaf Suitable Nutrient Standards

Termin et al. built a spatiotemporal dynamic clustering model based on Fuzzy C-means, and predicted the October best citrus canopy nitrogen level according to the October N-yield envelope curve [49]. Jiang et al. studied the dynamic changes in the N, P, and K requirements of wine grapes over different growth stages using the dissection of roots and the branching stems method. The optimal nutrient levels of the trees at different stages were obtained [3]. Sun et al. used the diagnosis- and recommendation-integrated system to obtain the best nitrogen status of apple leaf [9]. Chen et al. constructed a leaf nutrient suitability standard generation algorithm by continuously tracking the mineral content and fruit quality of citrus leaves at fixed points for two years, which can generate leaf nutrient standard content based on fruit quality [50].

4. Discussion

Tree nutrient rapid detection technology based on RGB images mainly predicts nutrient content based on color feature parameters, and its equipment and computational costs are relatively low. Changes in lighting conditions have a significant impact on nutrient prediction results, which has been validated by experimental results. It had high accuracy in leaf detection [3,4], but low accuracy in canopy detection [1,2]. Preprocessing algorithms were mainly aimed at removing the background and correcting colors, without addressing color differences caused by lighting changes.

The application for the multispectral detection of tree nutrients is based on vegetation indices, and the data processing volume is not large. It is often mounted on drones and has high efficiency. Considering environmental changes is crucial as it is directly detected in the field. Compared with existing research, it can also be found that without data preprocessing, the prediction accuracy was lower [5,6]. Considering the influence of tree structure, background, and lighting intensity, the prediction accuracy was greatly improved [7,9]. A RGB-NIR multispectral method has also been used to address the issue of uneven illumination [18].

Compared to multispectral data, hyperspectral data has more spectral bands, which means there are more characteristic variable parameters that can predict nutrients. So it will have higher accuracy, but it will bring significant challenges to data processing. The application of filtering and transformation algorithms for preprocessing effectively eliminated interference information and improved correlation [14,15]. The preprocessing algorithm for filtering sensitive feature bands reduced data processing and computational costs [10,13]. The studies on the hyperspectral detection of tree nutrients have mostly been at the laboratory level, with limited field applications. The combination of hyperspectral and multispectral did not show significant gain.

Rapid detection technologies for soil nutrients based on electrochemistry, high-temperature excitation and artificial olfaction have the advantage of low cost. However, their characteristic variables are single, easily affected by soil components, and have certain limitations. They also require sample preprocessing, such as preparing solutions [19], heating [22], and so on.

The hyperspectral detection of soil nutrients has been given extensive research, and like detecting tree nutrients, it has also applied filtering, transformation [26,31], and sensitive band screening algorithms [27] to solve the problem of data redundancy. The impact of soil moisture has also been considered [33]. From the results, it can be found that it had high prediction accuracy. Laser-induced breakdown spectroscopy is atomic spectroscopy that is generally superior for analytical applications with multi-element measurements [37]. Its research has just begun, and due to the complexity and diversity of soil components, spectral analysis poses high challenges. From the results, it can be observed that the combination spectrum did not show a significant improvement in predictive performance.

At present, many scholars have obtained soil and leaf suitable nutrient standards through long-term and extensive experiments, providing a basis for the intelligent diagnosis of fruit tree nutrients. However, different varieties of fruit trees have different fertilizer requirements for growth, and suitable nutrient standards are not universally applicable, so experimental research needs to be conducted again.

5. Conclusions

(1)

Rapid detection technologies of tree nutrients based on RGB images and multispectral have the characteristics of a low cost and low computational complexity, but their prediction accuracy is not high and they are susceptible to interference. Hyperspectral technology has high prediction accuracy due to its large number of characteristic variables. The application of filtering, transformation, and sensitive band selection algorithms solves the problem of data redundancy, improves the correlation of feature variables, and reduces computational complexity. However, there are few field applications, and preprocessing algorithms that consider the impact of changes in the field environment should be further studied in depth.

(2)

Hyperspectral technology also plays an important role in the rapid detection of soil nutrients, demonstrating good prediction accuracy. Laser breakdown-induced spectroscopy has good prospects, as it can simultaneously detect multiple nutrients, including those with lower content. However, spectral analysis is still full of challenges.

(3)

The existing rapid detection technologies for nutrients are mostly aimed at nitrogen with high content, and research on the rapid detection of small and trace elements needs to be strengthened.

(4)

Many suitable nutrient standards for soil and leaf have been established, but they are often obtained through long-term and extensive experimentation, which is time-consuming and laborious. A universal and rapid method needs to be studied to meet the construction needs of suitable nutrient standards for different varieties of fruit trees.