Prediction of Soil Water-Soluble Organic Matter by Continuous Use of Corn Biochar Using Three-Dimensional Fluorescence Spectra and Deep Learning

The purpose is to study the soil's water-soluble organic matter and improve the utilization rate of the soil layer. This exploration is based on the theories of three-dimensional fluorescence spectroscopy, deep learning, and biochar. Chernozem in Harbin City, Heilongjiang Province, is taken as the research object. Three-dimensional fluorescence spectra and a deep learning model are used to analyze the content of water-soluble organic matter in the soil layer after continuous application of corn biochar for six years and to calculate different fluorescence indexes in the whole soil depth. Among them, the three-dimensional fluorescence spectrum theory provides the detection standard for the application effect detection of biochar, the deep learning theory provides the technical support for this exploration, and the biochar theory provides the specific research direction. The results show that the application of corn biochar for six consecutive years significantly reduces the average content of water-soluble organic matter in different soil layers. Among them, the highest average content of soil water-soluble organic matter is “nitrogen, potassium, phosphorous” (NPK) and the lowest is “boron, carbon” (BC). Comparing the soil with BC alone, in the topsoil, the second section (330–380 nm/200–250 nm) with BC + NPK increases by 13.3%, the third section (380–550 nm/220–250 nm) increases by 8.4%, and the fourth section (250–380 nm/250–600 nm) increases by 50.1%. The combination of nitrogen (N) + BC has a positive effect of 20.7%, 12.2%, and 28.4% on sections I, II, and IV, respectively. In addition, in the topsoil, the combination of NPK + BC significantly increases the content of acid-like substances compared with the application of BC alone. In the black soil, with or without fertilizer NPK, there is no significant difference in the level of fulvic acid-like components. The prediction of soil water-soluble organic matter after continuous application of corn biochar based on three-dimensional fluorescence spectra and deep learning is carried out, which has reference significance for the rapid identification and early prediction of subsequent soil activity.


Introduction
With the development of society and the continuous increase of population, the social demand for agriculture is higher and higher, so it is imperative to improve agricultural productivity through various means. Among them, agricultural fertilization is the best way to promote the stable improvement of agricultural productivity, but excessive agricultural fertilization will cause irresistible harm to the land. Hence, in recent years, due to intensive agricultural management, people have conducted extensive research on the input of farmland fertilizers, especially organic materials [1]. Water-soluble organic matter can be decomposed by microorganisms and release nutrients for plants to absorb. Terefore, it is usually considered one of agricultural development's most important components. Strengthening the prediction of soil water-soluble organic matter can fnd the possible problems in the soil earlier and improve the yield of crops to a certain extent [2,3].
Bradford et al. efectively integrated image technology and spectral technology in the prediction of soil watersoluble organic matter. Tis technology can image the object to be measured and visualize the image information of the sample. Moreover, it can obtain the spectral information of the sample to be measured under multiple continuous narrow bands and then refect the internal main material composition through the peak and trough information. Terefore, each data acquisition obtains a hypercube generated by stacking hundreds of single-channel black-andwhite (gray scale) images (each image represents the corresponding spectral wavelength band) with each other. It is a noninvasive and noncontact rapid detection technology, which can scan massive samples simultaneously, thus overcoming the limitations of conventional spectral technology that can only measure a single sample [4]. Tilakarathna and Hernandez-Ramirez established a prediction model for the content of organic matter in the soil by analyzing the relationship between indoor chemical composition and outdoor soil spectrum and applied it to hyperspectral remote sensing images to obtain the classifcation map of soil organic matter in this region [5]. Galicia-Andres et al. conducted plot yield and canopied spectral refectance replication tests on three sets of soybean breeding lines at diferent growth stages. Hyperspectral remote sensing technology was used to predict the yield of soybean breeding areas. It was found that multiple linear regression was suitable for plot yield prediction [6]. Man et al. conducted nondestructive testing on internal defects of apples in semitransmission mode. Te classifcation results vary with the direction of tested fber and fruit. When using the average spectrum, better classifcation results can be obtained for apples with slight defects [7]. Blonska et al. studied and compared four multivariate data analysis methods to optimize beef adulteration's rapid nondestructive quantitative detection based on visible near-infrared hyperspectral imaging. Te least-squares support vector machine model has achieved good prediction results, and the root mean squared errors of training and prediction are 5.39% and 6.29%, respectively [8]. Rocci et al. used perennial ryegrass as an experimental crop and combined hyperspectral imaging with machine learning to analyze the fodder value. Te results show that it has a good prediction ability, which is conducive to improving the speed and reducing the cost of the commercial breeding plan [9]. Escalona et al. used hyperspectral imaging technology in the range of 400-1000 nm to predict the quality changes of corn seeds at diferent storage times. It mainly includes hardness, elasticity, and resilience. A partial least square regression algorithm was adopted to integrate the reference values measured by traditional methods with the extracted spectral data. Te deep learning model has produced good results in predicting hardness, elasticity, and resilience. It shows that hyperspectral imaging technology can be used as a fast, accurate, and nondestructive tool to detect the impact of diferent storage times on the quality characteristics of corn [10].
Under this background, this exploration takes chernozem in Heilongjiang Province as the research object and applies three-dimensional fuorescence spectra and a deep learning model to analyze the content of water-soluble organic matter in the soil layer after six years of fertilization and to calculate diferent fuorescence indicators in the whole soil depth. Te research innovation is to integrate the three-dimensional fuorescence spectrum and deep learning model and comprehensively analyze the content of watersoluble organic matter in the soil layer after continuous application of corn biochar for six years, the three-dimensional fuorescence spectrum, and diferent fuorescence indicators in the whole soil depth. Te purpose is to provide a method for improving soil organic matter content and fnding soil problems as soon as possible. It suggests that this exploration provides a reference for the optimization of biochar application and also makes a contribution to the development of agriculture in the future.

Tree-Dimensional Fluorescence Spectroscopy and Deep
Learning. Te three-dimensional fuorescence spectrum is an excitation-emission-matrix spectra characterized by the three-dimensional coordinates of the excitation wavelength (y-axis), emission wavelength (x-axis), and fuorescence intensity (z-axis), also known as the total luminescence spectrum [11]. Te usual fuorescence spectrum is a plan view obtained by scanning the fuorescence intensity against the emission wavelength. Te three-dimensional fuorescence spectroscopy technology can obtain the excitation wavelength, emission wavelength, and fuorescence intensity information when it changes [12]. Tree-dimensional fuorescence spectrograms are generally expressed in threedimensional projection and contour fuorescence spectrograms [13]. Figure 1 shows the characteristics of this spectrum. Figure 1 shows that the fuorescence spectrum represented by three-dimensional projection is more intuitive, and it is easier to observe the position and height of the fuorescence peak and some characteristics of the spectrum. However, it is not easy to directly provide the information of the corresponding fuorescence emission intensity of the excitation-emission wavelength pair. In synchronous scanning with a fxed excitation wavelength and emission wavelength diference, each compound has a single fuorescence peak, and the contour map of each compound is limited to a rectangular box. If the sample contains four fuorescent compounds, there will be four rectangular boxes, but the components can be detected only in the nonoverlapping position. Contour fuorescence spectrograms make it easier to make graphic comparisons. Because the threedimensional fuorescence spectrum has one more coordinate than the two-dimensional plan and the total fuorescence data obtained are much more than the ordinary fuorescence spectrum, it has high selectivity and can be used for the analysis of multicomponent mixtures. Based on the above characteristics, the three-dimensional fuorescence spectrum is commonly used in the following felds: (1) Te contour fuorescence spectrum of blood is used to check the health of the human body. Human blood is composed of multiple components. When human blood has problems, the visible spectrum of total fuorescence will change signifcantly [14].
(2) Te three-dimensional fuorescence spectrum helps solve criminal cases. It has a stronger discrimination ability for relevant samples. Especially, the spectral subtraction method is conducive to solving criminal cases [15].
Te "deep" of deep learning is compared with the shallow machine learning method, which originates from artifcial neural networks [16]. Te deep neural network is the foundation of deep learning. Its structure well refects the characteristics of "multilayer" and "nonlinear." Te neuron structure is the basic structure of the neural network system. Te biological nervous system contains nervous subsystems with a diferent division of labor. After receiving corresponding stimulation as input, they give corresponding feedback information based on the comprehensive input results and jointly drive the normal operation of the whole system [17]. Inspired by biological neurons, artifcial neurons also have a similar structure and mechanism; that is, the receiving vector is used as the input of neurons to give the corresponding results in the form of a weighted sum, and fnally, the delinearization is completed through the activation function. Figure 2 displays its basic structure.
In this model, I i , i � 0, 1, 2 . . . n is the initial input of the model, u is the output of the model after linear change, and T is the fnal output of the neuron. a i , i � 0, 1, 2 . . . n is adopted as the weight parameter of the linear transformation process of the neuron structure. b is the ofset term, and the function f plays the role of delinearization [18]. Tus, equations (1) and (2) are inputs of the structure of neurons.
where f is called the activation function. Besides, based on the neuron structure, multiple neurons are connected in turn according to diferent levels to form the most basic deep neural network [19]. Te network model can be divided into three layers: input, hidden, and output. Figure 3 presents the specifc model.
In Figure 3, the frst layer is the input layer, whose function is to input the data to be calculated into the model to provide materials for the model's operation. Te second level is the hidden layer, which needs to complete the feature Contour fuorescence spectrogram is easy to compare fgures It is intuitive and easy to observe the position and height of the fuorescence peak and some characteristics of the spectrum It has high selectivity and can be used for the analysis of multicomponent mixtures Computational Intelligence and Neuroscience extraction. Each node of the hidden layer will have diferent receiving weights for the input information of the input layer, so it is more inclined to a certain recognition mode. Tat is, the meaning of the hidden layer is to abstract the characteristics of input data to another dimension of space to show its more abstract characteristics, which can be better divided linearly. Multiple hidden layers are multilevel abstractions of input features, and the ultimate goal is to better linearly divide diferent data types. Te third level is the output layer. Te function of the output layer is to classify and output the calculated data through the extracted features. It is set that the input value is I, the weight value of each layer is a, the ofset term is b, and the activation function is f. Te corresponding output of the hidden layer and the output layer is T. Equations (3)-(6) display the deduced calculation process of the forward propagation algorithm of the deep neural network.
Te forward propagation process can be abstracted from the above derivation process. If there are n neuron structures in layer l − 1, equation (7) is the expression of the output of the k-th neuron in layer l.
Te evaluation criteria of the neural network model and the corresponding optimization direction and objective are usually defned by the loss function. When the neural network is adopted to complete the classifcation task, the interaction entropy is usually introduced to judge the difference between the output and expected vector [20]. For given two probability distributions p and q, the interaction entropy of p expressed by q is as follows: Te distance between two probabilities can be described by the interaction entropy. Te closer the two probabilities' distribution is, the smaller the diference between the output vector and the expected vector is, and the better the classifcation efect is. At this time, the value of interaction entropy will be smaller [21].

Biochar and Soil Water-Soluble Organic Carbon.
Biochar is a kind of charcoal as a soil conditioner, which can help plants grow. It can be used for agricultural purposes, carbon collection, and storage. It difers from the traditional charcoal commonly used for fuel [22]. In recent years, scientists have paid attention to the use of biochar due to the impact of climate change caused by the emission of greenhouse gases such as carbon dioxide, nitrous oxide, and methane. Te reason is that biochar helps capture and remove greenhouse gases in the atmosphere by means of biochar storage, convert them into quite stable forms, and store them in the soil for thousands of years. In addition, biochar can increase agricultural productivity by 20%, purify water quality, and help reduce the use of chemical fertilizers [23]. Figure 4 shows its main properties.
Soil carbon pool is a crucial component of the carbon cycle and an important indicator of soil fertility and biodiversity [24]. Soil water-soluble organic carbon is one of the components of soil organic carbon, which mainly comes from fallen leaves and plant residues, microbial degradation products, and root exudates [25]. Organic manure and animal manure under human infuence increase the solubility of soil organic matter, which is also one of the sources of soil water-soluble organic carbon.
Te main factors afecting soil water-soluble organic carbon are as follows: (1) Soil water-soluble organic carbon is afected by multiple factors, such as climate, temperature, and vegetation coverage. Among them, the soil watersoluble organic carbon content in Phyllostachys edulis forest is the highest and that in pinus massoniana forest is the lowest. Tis phenomenon is related to the root system, exudates, and other factors of the vegetation [26].
(2) Te change of seasons will also afect soil watersoluble organic carbon. In the season with less rain, the amount of plant litter is relatively large, and the soil water-soluble organic carbon content is relatively high [27].  Computational Intelligence and Neuroscience large aggregates in the soil decrease signifcantly after 10 years of cultivation, and the water-soluble organic carbon in the surface soil also decreases [28].

Research Data Settings
Chernozem in a small town of Harbin City, Heilongjiang Province, is taken as the research object. Te local long-term average annual rainfall ranges from 486.4 mm to 543.6 mm. Te rainy season is mainly from June to September. Te average altitude is 138 m, and the depth of groundwater level is 80 m [29,30]. Te annual average wind speed is 4.1 m/s, and the maximum wind speed is 18.9 m/s. Chernozem in a small town in Heilongjiang Province is extracted, detected, and counted based on the relevant data released by the department of ecological environment of Heilongjiang Province. Tables 1 and 2 show the local soil properties and the particle composition of biochar. Field trials of soybean/maize rotation were conducted from 2013 to 2018. Tere were mainly fve types of fertilization during this period, namely, carbon phosphorous (CK), nitrogen, potassium, phosphorous (NPK), 3.6 t/ha boron, carbon (BC), NPK + 3.6 t/ha BC, and nitrogen (N) + 3.6 t/ha BC. On May 14, 2013, the spring soybean variety Heinong 58 was sown and harvested on October 8. Te spring maize variety Longdan 42 was sown on May 1, 2014, and harvested on October 8. From 2015 to 2018, soybeans and corn were planted in turn. Te NPK fertilizer of soybean was applied before planting, while the N fertilizer of maize was applied in the form of urea. Te amount of 176 kg·N/ha was used as the base fertilizer, and the remaining 50% was applied at the jointing stage [31]. As a soil conditioner, BC was applied to the ditches near the ridge every year from 2013 to 2018. Te method is to thoroughly mix the soil with a plow and plow to a depth of at least 20 cm.
In 2018, soil samples were collected using a soil auger (diameter of 10 cm) based on a 60 cm soil profle. Simply put, samples were taken from 0-60 cm soil layers every 10 cm and then dried at room temperature for 1 week. After sieving through a 2 mm sieve, the soil sample was kept at 4°C until the next analysis [32].

Extraction and Measurement of Soil Organic Matter.
Soil water-soluble organic carbon was extracted by soil water oscillation. Simply put, each soil sample was mixed with deionized water at a solid water ratio of 1 : 6 (w/v) and incubated for 24 hours under continuous shaking (180 rpm). After centrifuging 10,000 g of ionic water for 6 minutes, the suspension was fltered through a cellulose acetate membrane flter (pore size: 0.45 μm). After fltration, the sample was kept at −20°C for three-dimensional fuorescence spectroscopic analysis. Te total organic carbon analyzer (TOC-VCPH, Shimadzu, Japan) was adopted to detect the organic carbon level in the fltrate.
Te Parafac model is constructed using Matrix and Laboratory (MATLAB) 7.0 and DOMFluor toolbox to characterize the fuorescence components of soil watersoluble organic carbon. An appropriate number of components in the model are measured according to the core consistency diagnosis and the split validation test. By estimating the relative contribution of each component, the diferences between diferent components in each sample are further compared.
Statistical Product and Service Solutions (SPSS) 20.0 (SPSS Company, Chicago, Illinois, USA) is used for data analysis. After checking homogeneity and normality, oneway ANOVA and least signifcant diference test are performed on the data of homogeneity variance and normal distribution. Te purpose is to compare the diferences in water-soluble organic matter content between various land use types and soil depths. Meanwhile, the Pearson correlation coefcient is used to determine the correlation between the parameters of water-soluble organic matter. P values of 0.05 and 0.01 are considered statistically signifcant.

Contents of Water-Soluble Organic Matter in Diferent Soil
Layers after Six Years of Fertilization. Te above-mentioned methods for extracting and measuring soil organic matter are used to analyze the content of water-soluble organic matter in the soil layer after the continuous application of corn biochar for six years. Figure 5 shows the specifc situation.
During the 6-year test, BC will afect the average content of water-soluble organic matter in the soil profle, whether chemical fertilizer is applied or not. In 0∼60 cm layer depth, the average content of water-soluble organic matter in soil under diferent fertilizers is NPK (150. Computational Intelligence and Neuroscience the soil. Moreover, the content of water-soluble organic matter in BC treated soil is lower than that in nonfertilized soil.

Application of Tree-Dimensional Fluorescence Spectra and Deep Learning Model in Prediction of Soil Water-Soluble
Organic Matter. Tree-dimensional fuorescence spectroscopy and a deep learning model are applied to detect soil water-soluble organic matter after 6 years of fertilization. Figure 6 displays the specifc results. According to the emission and excitation wavelengths of the target molecules, the spectrometer can be divided into fve parts. Section 1 (250-330 nm/200-250 nm) contains tyrosine-like substances. Section II (330-380 nm/ 200-250 nm) contains tryptophan-like substances. Section III (380-550 nm/220-250 nm) contains fulvic acid-like substances. Section IV (250-380 nm/250-600 nm) contains soluble microbial by-product samples. Section V (380-600 nm/250-600 nm) contains humic acid-like substances. Figure 6 displays that the fuorescence intensity and the shape of the spectrogram change with the change in soil depth. After 6 years of fertilization, the fuorescence intensity of the third and ffth sections reaches the maximum, especially the fuorescence intensity of the ffth section, which exceeds 60%. It shows that humic acid and fulvic acid are the most important organic matter in black soil, whether or not it is improved. Compared with the BC group, BC + NPK treatment signifcantly increases the fuorescence intensity of section II (0-20 cm soil layer increases by 13.3%), section III (0-20 cm soil layer increases by 8.4%), and section IV (0-20 cm soil layer increases by 50.1%). Te results show that BC and NPK fertilizers are enriched with tryptophan-like substances, fulvic acid-like substances, and soluble microbial by-products. Meanwhile, in the surface layer (0-20 cm), the frst section (increases by 20.7%), the second section (increases by 12.2%), and the fourth section (increases by 28.4%) of BC + N fertilizer showed an increasing trend. In contrast, compared with NPK fertilizer, BC fertilizer signifcantly afects the fuorescence intensity of the second section (0-20 cm decreases by 9.6%) and the fourth section (0-20 cm increases by 8.5%), indicating that tryptophan and humic acid-like substances are reduced and enriched, respectively. In addition, BC + NPK fertilizer signifcantly increases the fourth part of soluble microbial by-productlike substances (the fuorescence intensity increases by 5.9% at 10-20 cm) and signifcantly reduces the frst part of tyrosine-like substances (the fuorescence intensity decreases by 26.1% at 10-20 cm and 13.7% at 20-30 cm).
In addition, through the model established above, different fertilizers are put into the 0-60 cm soil layer to estimate the three fuorescent components. Figure 7 shows the specifc results.
Te data results of C1-C3 in Figure 7 show that the content of fulvic acid-like substance (C2) decreases and the content of protein-like substance (C3) increases with the increase in soil depth. In addition, the content of humic acidlike substance (C1) in each treatment group is higher than   Particle components (%) Soil organic carbon (SOC) * (g/kg) Ca (g/kg) K (g/kg) Mg (g/kg) N (g/kg) O (g/kg) P (g/kg) Si (g/kg) pH <0.    Computational Intelligence and Neuroscience that of fulvic acid-like substance (C2), while the content of protein-like substance is the lowest. Te protein-like substance (C3) content shows an opposite trend, while the content of C2 remains more than 50% in the deep layer (20-60 cm). It reveals that the water-soluble organic matter in the surface layer (0-20 cm) is composed of humic acids.
On the contrary, the water-soluble organic matter in the subsoil layer (20-60 cm) is produced by microbial decomposition, and the level of microbial by-products increases with the increase in soil depth. In addition, after 6 years of fertilization, the level of humic acid-like substances (C1 and C2) in the subsoil layer (20-60 cm) in BC decreases by about Computational Intelligence and Neuroscience 9 5% compared with NPK. Te C1 content of NPK + BC treatment is the lowest and the C2 content is the highest, which is distributed in the surface layer and the lower layer. Figure 8 shows the specifc results of the fuorescence index (FI), humifcation index (HIX) and biological index (BIX) of fve treatment groups in the whole soil depth. Figure 8 shows that FI values in diferent groups remain above 1.59. When BC is applied, the FI value of topsoil (0-10 cm) is the lowest, which is 1.52. When BC and NPK fertilizers are applied, the FI value of subsoil (40-50 cm) is the highest. Terefore, it can be inferred that the water-soluble organic matter imported by BC contains similar land-integrated resources in the topsoil (0-20 cm).
It has moderate neogenesis characteristics and shows strong authigenic characteristics in the subsoil (20-60 cm). Most of the HIX values for diferent group classes are about 0.27. Compared with the control group, in which the HIX value decreases with the increase of soil depth, the average HIX value of BC-improved soil in diferent soil layers is higher than that of NPK fertilized soil. As expected, the HIX value of the NPK + BC group is higher than that of BC. In addition, the average HIX value of the water-soluble organic matter in surface soil and subsoil of the NPK + BC group is lower than that of the N + BC group. It shows that the contribution of NPK is greater than that of N. Besides, BI values for all groups  show the same growth trend. In particular, the average value of BIX distribution in the soil profle is between 0.63 and 0.67. Tese results show that the content of watersoluble organic matter increases with the increase of soil depth through microbial degradation.

Conclusions
Soil quality is an important factor afecting seed growth. Based on the three-dimensional fuorescence spectra and deep learning model, this exploration studies the prediction of soil water-soluble organic matter by the continuous application of corn biochar. Te results reveal the following: (1) Te average content of soil water-soluble organic matter is NPK > NPK + BC > N + BC > CK > BC.
(2) Te application of BC for six consecutive years signifcantly reduces the average content of water-soluble organic matter in diferent soil layers, mainly because BC has a higher adsorption capacity than CK. Te contents of NPK + BC and N + BC treatment groups are similar, which shows that N is the main factor afecting the soil water-soluble organic carbon level. (3) Compared with the application of BC alone, in the 0-20 cm soil layer, the second section using BC + NPK increases by 13.3%, the third section increases by 8.4%, and the fourth section increases by 50.1%. Te combination of N + BC has a positive efect of 20.7%, 12.2%, and 28.4% on the fuctuation of segments I, II, and IV, respectively. (4) In the surface layer (0-20 cm), compared with BC alone, the combination of NPK + BC signifcantly increases the content of acid-like substances. In the black soil, with or without NPK, there is no signifcant diference in the level of fulvic acid-like components, especially in the surface layer (0-20 cm). However, the humifcation degree of T3 (NPK + BC) is slightly higher than that of T4 (N + BC) from the perspective of the surface soil decay value. In addition, the BC addition decrease in the past six years negatively impacts soil humidifcation. Tese fndings are conducive to a better understanding of the dynamics of water-soluble organic matter in diferent soil layers under diferent fertilization types and the comprehensive efects of fertilizer and corn straw BC on the biogeochemical characteristics of water-soluble organic matter in black soil areas.
Although this exploration provides relatively perfect research results, only chernozem in Heilongjiang Province is taken as the research object, and the prediction of water-soluble organic matter in other types of soil has not been carried out. Besides, three-dimensional fuorescence spectroscopy and a deep learning model are used to analyze the water-soluble organic matter in soil, and other spectral techniques can be selected to improve the prediction ability of soil water solubility according to the actual situation. Meanwhile, it is essential to expand the selectivity of research samples and establish a more sound database. Terefore, the scope of research objects will be expanded, and the research methods will be optimized to provide a reference for the agricultural development of more regions in the future.

Data Availability
Te data supporting the fndings of the current study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that there are no conficts of interest.