Cultivated Land Quality Evaluated Using the RNN Algorithm Based on Multisource Data

Abstract

Cultivated land quality (CLQ) is associated with national food security, benign economic development, social harmony, and stability. The scientific evaluation of CLQ provides the basis for achieving the “trinity” protection of cultivated land quantity, and quality, as well as ecology. However, the current research on CLQ evaluation has some limitations, mainly the poor consideration of evaluation indicators, time-consuming and labor-intensive data acquisition, and low precision of evaluation at the regional scale. Therefore, this study introduced multisource data to evaluate CLQ and proposed a new method for CLQ evaluation (natural grade evaluation, utilization grade evaluation, and economic grade evaluation), combining multisource data and the recurrent neural network (RNN) algorithm. Initially, optimal indicators were determined by correlation analysis and generalized linear regression coefficient methods based on factors related to CLQ acquired from multisource data. Then, CLQ evaluation models were constructed with the RNN algorithm on the basis of the aforementioned optimal indicators. Finally, the models were adopted to map CLQ. The present study was carried out in Guangzhou City, Guangdong Province, China. According to the results: (1) CLQ showed close relationship to pH, effective soil layer thickness (EST), chemical fertilizer application rate (CHFE), organic matter content (OMC), annual accumulated temperature (TEMA), 5–15 cm soil depth soil cation exchange capacity (CEC515), 0–5 cm soil depth soil cation exchange capacity (CEC05), 5–15 cm soil depth soil organic carbon content (SOC515), 0–5 cm soil depth soil organic carbon content (SOC05), field slope (FS), groundwater level (GWL), and terrain slope (TS). (2) All modeling accuracies (${\mathrm{R}}^2$) were greater than 0.80 for the CLQ evaluation models constructed based on the RNN algorithm. The area and spatial distribution of each grade of CLQ evaluation were consistent with the actual situation. The best and the worst quality cultivated land occupied a small area, and the area without a gap with the actual CLQ was as high as 76%, indicating that the model results were reliable. The study shows the suitability of the method for evaluating CLQ at the regional scale, offering a scientific foundation for the rational utilization and management of cultivated land resources, as well as a reference for evaluating CLQ in the future.

1. Introduction

Cultivated land is the lifeline and main carrier of grain production, which is of significant importance to guarantee global food security, ecological security, and sustainable human development [1,2,3,4]. Cultivated land quality (CLQ) denotes the soil quality, spatial geography quality, and management quality, showing the overall level of cultivated land production capacity, cultivated land environment, and product quality [5,6]. However, in recent years, factors containing soil pollution and non-agricultural occupation have led to CLQ degradation, which can affect food production and may also threaten regional food security as well as the sustainable use of cultivated land resources [7,8]. Thus, introducing new technical means to obtain regional-scale CLQ information rapidly and accurately is essential.

The research on CLQ evaluation includes determining evaluation indicators and constructing CLQ evaluation models. In terms of determining evaluation indicators, the traditional CLQ evaluation indicators were determined according to the relevant national standards of China, such as the Regulation for Gradation on Agriculture Land Quality (GB/T 28407-2012) and the Cultivated Land Quality Grade (GB/T33469-2016). The evaluation indicators mainly include five aspects: topographic characteristics, soil properties, tillage conditions, health status, and biological characteristics of cultivated land. However, these indicators were obtained through field sampling and laboratory analysis, which was time consuming and labor intensive. Moreover, the quantitative methods of these indicators are not suitable for regional scale CLQ evaluation or cannot reflect the spatial details of CLQ. For example, irrigation guarantee rate and drainage conditions are usually quantified based on simple manual interpretation, which leads to strong subjectivity. In addition, with the continuous enrichment and development of the concept and connotation of CLQ, based on the principle that indicators are easily accessible and quantifiable, aiming at the characteristics of soil quality, natural quality, utilization conditions, and ecological environment in cultivated land, many scholars have proposed a variety of CLQ evaluation indicator systems. For example, Duan et al. defined CLQ from four aspects: soil fertility, natural conditions, construction level, and cultivated land productivity, and then selected eight indicators in combination with multisource remote sensing data to evaluate CLQ [7]. Liu et al. directly used slope, temperature vegetation drought index (TVDI), vegetation index (VIs), road accessibility (RA), and patch fractal dimension (PFD) to construct a pressure-state-response (PSR) framework for CLQ evaluation [9]. However, these studies mainly selected indicators through subjective judgment, and the methods rely excessively on a priori knowledge that is highly subjective and lacking objective data support, which can easily lead to decision bias. It is well known that evaluation indicator selection is one of the most vital steps in CLQ evaluation. This step helps eliminate irrelevant and redundant information, reduce the data dimension and the complexity of the algorithm, lower the error of the model fitting, as well as improve the promotion ability of the model and the accuracy of the evaluation [10]. At present, the generalized linear regression algorithm is commonly used for indicator selection because of its good indicator interpretation and visual display capability, thus overcoming the limitations of subjective judgment methods. The method is a linear model extension that overcomes the shortcomings of the latter (e.g., the ordinary linear algorithm requires input data that satisfy a normal distribution) by finding the exponential family distribution of the dependent variable, constructing the appropriate connection function, and establishing the correlation between the mathematical expectation value of the response variable and the linear combination of the predictive variables [11]. Therefore, this study selected correlation analysis and generalized linear regression algorithm for indicator selection to obtain more accurate CLQ evaluation indicators.

At present, CLQ evaluation models are divided into traditional models based on the relevant national standards of China (GB/T 28407-2012 and GB/T33469-2016) and the multi-indicator synthesis approaches based on artificial intelligence algorithms. In the traditional evaluation method, evaluation indicators are obtained by field sampling and indoor analysis, and then the Delphi method, combined with the analytic hierarchy process (AHP), fuzzy comprehensive evaluation method, and other mathematical statistical models, is used to score the indicators and assign the weight to complete CLQ evaluation [12,13]. Zhuang et al. used the Delphi–AHP method to assess CLQ and quantified the impact of CLQ on crop productivity, and the result showed that the Delphi–AHP method could be adopted for assessing CLQ with the highest accuracy (96.35%) [1]. Zhang et al. used the gray relational analysis (GRA) method and traditional AHP to evaluate farmland quality in Xiuwu County, northwestern Henan Province, and the results showed that the validation of GRA and AHP was close, good, and acceptable, and the CLQ findings of both could show the crop yield [14]. This method accurately evaluates CLQ; however, the evaluation is greatly affected by human subjectiveness and is mostly based on expert cognition and experience [15]. The indicators are poorly considered, lack objective description, and cannot avoid the problems of subjective will and conversion from qualitative to quantitative indicators. Multi-indicator synthesis models for evaluating CLQ based on artificial intelligence algorithms include artificial neural network (ANN), genetic algorithm (GA), support vector machine (SVM), and Random Forest (RF) model [16,17,18,19]. Liu et al. considered Ya’an cultivated land as a research object, constructed a CLQ evaluation system from seven perspectives containing soil organic matter and soil texture, introduced the attention mechanism (AM) into the back propagation neural network (BPNN) model, and designed an AM-BP neural network suitable for Ya’an CLQ evaluation, which yielded satisfactory results [20]. Based on remote sensing spectral data, Xia et al. adopted the GA-BPNN model for assessing the CLQ in Conghua District, Guangzhou, and compared the findings with the results of the partial least square regression (PLSR) model. According to the results, the accuracy of the GA-BPNN model was better when compared with that of the PLSR model [21]. Although this method is capable of enhancing the evaluation efficiency and accuracy of CLQ, it exhibits some limitations such as the poor consideration of indicators, limited types of indicators obtained, and inability to express CLQ fully, and the evaluation algorithm cannot optimize the spatiotemporal information in the data into a model, resulting in the low accuracy of CLQ evaluation at the regional scale [22]. The recurrent neural network (RNN) deep-learning algorithm is a special neural network structure that uses the partial derivative of the loss function to adjust the weight of each unit and shares the parameters in the connection of each step of input, output, and hidden states. This sharing mechanism can significantly lower the complexity of the model, shorten the training time, and maintain high accuracy [23,24,25,26]; hence, it has greater advantages over other ANNs. Additionally, in the current CLQ evaluation research, many studies have focused only on evaluating a single aspect of CLQ (such as the evaluation of natural grades) while ignoring the multifunctional evaluation of CLQ (such as the evaluation of utilization grades and economic grades). A few researchers have constructed multiple evaluation models for the quality of Guangzhou’s cultivated land. Therefore, for improving the accuracy of CLQ evaluation, the RNN deep-learning algorithm was applied with the aim of constructing the CLQ multifunctional evaluation model in this study.

Therefore, the current work introduced multisource data to obtain the evaluation indicators of CLQ. Compared with traditional data, multisource data exhibit 5 V characteristics: huge volume, variety, velocity, redundancy, and high value. It can explore the knowledge law and high-value information in massive data by using the big data technology of the high speed of capture, discovery, and analysis and can better solve the existing problems in CLQ evaluation [27]. Additionally, the deep-learning model can automatically extract the spatiotemporal information in the data and improve the modeling ability of the multi-time scale and long-range spatial correlation [22]. Thus, with the aim of improving the accuracy of CLQ evaluation, the RNN deep-learning algorithm was applied with the aim of constructing the CLQ evaluation model, aiming to realize multifunctional CLQ evaluation.

2. Materials and Methods

2.1. Study Area

Guangzhou City (112°57′–114°03′E, 22°26′–23°56′N) (Figure 1) is the central city in South China and is also an economic, political, cultural, educational, and technological hub. It is situated in the central and southern parts of Guangdong Province and the northern part of Pearl River Delta. Guangzhou covers an area of approximately 7434.4 ${\mathrm{km}}^2$, covering 11 districts under its jurisdiction, as shown in Figure 1. Guangzhou is a hilly area, with an average elevation of 107 m. In general, the terrain is high in the northeast, low in the southwest and south, and tilts northeast to southwest and south. The annual average temperature in Guangzhou is 21.9 °C, with the average annual precipitation being approximately 1800 mm, thus representing the marine subtropical monsoon climate [7]. By the end of 2016, the total area of cultivated land in the jurisdiction reaches 84,269.99 ha. The main utilization methods involve paddy field, irrigated land, and dry land. Rice is the main crop in the study area, and the rice planting area accounts for 88% of the total cultivated land area [7]. According to the 2016 Guangzhou CLQ grade results, the original sample set was generated by randomly selecting cultivated land patches of varying quality grades in Guangzhou based on the stratified sampling method. Totally 2000 samples were chosen, of which 1400 (70%) were used for modeling (yellow plots in Figure 1) and 600 (30%) were adopted for evaluating the accuracy of the estimated CLQ (red plots in Figure 1).

2.2. Data and Preprocessing

The CLQ data involved in Guangzhou in 2016 were derived from Guangdong Provincial Key Laboratory of Land Use and Consolidation (GPKLLUC), South China Agricultural University, Guangzhou, China (accessed on 6 September 2018), comprising 73,498 vector speckle data, and the data space projection was GCS_Xian_1980. The attribute table of the data set mainly includes six evaluation indicators such as effective soil layer thickness (EST), soil organic matter content (OMC), pH, terrain slope (TS), field slope (FS), and groundwater level (GWL), as well as six CLQ attributes such as national natural grade index (NNGI), national natural grade (NNG), national utilization grade index (NUGI), national utilization grade (NUG), national economic grade index (NEGI), and national economic grade (NEG). Here, NNGI is the benchmark crop yield index converted from the yield ratio coefficient for each designated crop under natural quality conditions of agricultural land as determined by the standard farming system. It reflects the natural productive potential of cultivated land or background yield level. NUGI is the sum of the base crop yields converted by the yield ratio coefficient for each designated crop determined according to the standard farming system under the natural quality conditions of cultivated land and the average utilization conditions of the land use zoning where the cultivated land is located. This yield can also be interpreted as the maximum possible yield level achieved by the cultivated land within the gradation unit under the most favorable local economic conditions. Therefore, it can also be called the realistic yield level of cultivated land. NEGI is the sum of the base crop yields converted by the yield ratio coefficient under the natural quality conditions of the cultivated land, the average utilization conditions of the land use zoning where the cultivated land is located, and the average economic conditions of the land economic zoning where the cultivated land is located. This yield can also be interpreted as the maximum economic yield level that can be achieved by the cultivated land in the gradation unit under the current agricultural technical and economic conditions. Therefore, it can also be called the economic yield level of cultivated land. Additionally, the above three CLQ grade indices reflect the quality of cultivated land in combination. With reference to the Regulation for Gradation on Agriculture Land Quality in the National Standard of China (GB/T 28407-2012) and the CLQ grade indices (NNGI, NUGI, and NEGI) were calculated as:

$${\mathrm{R}}_{\mathrm{i}}={\displaystyle \sum }{\mathsf{\alpha }}_{\mathrm{j}}\times \frac{{{\displaystyle \sum }}_1^{\mathrm{m}}{\mathrm{w}}_{\mathrm{k}}\times {\mathrm{f}}_{\mathrm{ijk}}}{100}\times {\mathsf{\beta }}_{\mathrm{j}}\times {\mathrm{K}}_{\mathrm{LC}}+\mathrm{b},$$

(1)

where ${\mathrm{R}}_{\mathrm{i}}$ is the ith sample CLQ grade indices (NNGI, NUGI, and NEGI), ${\mathsf{\alpha }}_{\mathrm{j}}$ is the solar and temperature productivity potential index of the jth crop, ${\mathrm{w}}_{\mathrm{k}}$ is the kth indicator weight, ${\mathrm{f}}_{\mathrm{ijk}}$ is the kth indicator score at the ith sample of the jth appointment crop, the value range is (0–100], ${\mathsf{\beta }}_{\mathrm{j}}$ is the yield ratio coefficient of the jth crop, ${\mathrm{K}}_{\mathrm{LC}}$ is the comprehensive land coefficient, b is a constant. CLQ grade indices (NNGI, NUGI, and NEGI) were calculated using Equation (1) with different coefficients according to GB/T 28407-2012. Then, CLQ grade indices are divided into three CLQ grades (NNG, NUG, and NEG) according to different equidistant methods. Thus, these CLQ data can be identified as field measured values (true values).

Additionally, based on the existing CLQ data, 37 CLQ evaluation indicators were preliminarily determined by referring to previous studies [1,5,7,15]. As shown in

Table 1. Cultivated land quality evaluation indicator system.

Indicators	Indicator Description	Data Source	Data Acquisition Time
EST	Effective soil layer thickness (cm)	GPKLLUC	Accessed on 6 September 2018
OMC	Organic matter content (%)	GPKLLUC	Accessed on 6 September 2018
PH	Soil pH	GPKLLUC	Accessed on 6 September 2018
TS	Terrain Slope (°)	GPKLLUC	Accessed on 6 September 2018
GWL	Groundwater level (cm)	GPKLLUC	Accessed on 6 September 2018
FS	Field slope (°)	GPKLLUC	Accessed on 6 September 2018
NDVI	Normalized vegetation index	GEE platform Landsat 8	Accessed on 15 May 2022
DEM	Digital Elevation Model (m)	GEE platform NASADEM	Accessed on 15 May 2022
SLP	Slope (°)	GEE platform NASADEM	Accessed on 15 May 2022
ASP	Aspect	GEE platform NASADEM	Accessed on 15 May 2022
POP	Population density	GEE platform WorldPop	Accessed on 15 May 2022
PH05	0–5 cm soil depth soil pH value	http://soil.geodata.cn	Accessed on 12 May 2022
PH515	5–15 cm soil depth soil pH value	http://soil.geodata.cn	Accessed on 12 May 2022
CF05	0–5 cm soil depth soil gravel content (%)	http://soil.geodata.cn	Accessed on 12 May 2022
CF515	5–15 cm soil depth soil gravel content (%)	http://soil.geodata.cn	Accessed on 12 May 2022
TN05	0–5 cm soil depth soil total nitrogen content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
TN515	5–15 cm soil depth soil total nitrogen content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
TP05	0–5 cm soil depth soil total phosphorus content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
TP515	5–15 cm soil depth soil total phosphorus content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
TK05	0–5 cm soil depth soil total potassium content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
TK515	5–15 cm soil depth soil total potassium content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
BD05	0–5 cm soil depth soil bulk density (g/cm3)	http://soil.geodata.cn	Accessed on 12 May 2022
BD515	5–15 cm soil depth soil bulk density (g/cm3)	http://soil.geodata.cn	Accessed on 12 May 2022
CEC05	Soil cation exchange capacity of 0–5 cm soil depth (cmol(+)/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
CEC515	5–15 cm soil depth soil cation exchange capacity (cmol(+)/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
SOC05	0–5 cm soil depth soil organic carbon content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
SOC515	5–15 cm soil depth soil organic carbon content (g/kg)	http://soil.geodata.cn	Accessed on 12 May 2022
TKN	Soil thickness (cm)	http://soil.geodata.cn	Accessed on 12 May 2022
TEMM	Annual average temperature (°C)	http://data.cma.cn	Accessed on 18 October 2021
TEMA	>0 °C annual accumulated temperature (°C)	http://data.cma.cn	Accessed on 18 October 2021
PRE	Total annual precipitation (mm)	http://data.cma.cn	Accessed on 18 October 2021
RAD	Annual total solar radiation (MJ/m2)	http://data.cma.cn	Accessed on 18 October 2021
CHFE	Chemical fertilizer application rate of cultivated land per unit area (t/ha)	http://tjj.gz.gov.cn/	Accessed on 13 March 2022
PSTD	Pesticide application rate of cultivated land per unit area (t/ha)	http://tjj.gz.gov.cn/	Accessed on 13 March 2022
PLSH	Film usage per unit area of cultivated land (t/ha)	http://tjj.gz.gov.cn/	Accessed on 13 March 2022
DFRS	Distance from field to rural settlement (m)	https://lbsyun.baidu.com/	Accessed on 18 May 2022
DFRR	Distance from field to rural road (m)	https://download.geofabrik.de/	Accessed on 18 May 2022

, normalized vegetation index (NDVI), digital elevation model (DEM), slope, aspect, and population data were derived from Landsat 8, NASADEM, and WorldPop data, respectively. The spatial resolution of Landsat 8 and NASADEM data was 30 m, and the spatial resolution of WorldPop data was 100 m. Data preprocessing was performed using the GEE platform (https://earthengine.google.com/, accessed on 15 May 2022). Soil data were derived from Soil Sub Center, National Earth System Science Data Center, National Science and Technology Infrastructure (http://soil.geodata.cn, accessed on 12 May 2022) and included soil pH, organic carbon, total nitrogen, total phosphorus, total potassium, soil bulk density, soil cation exchange capacity, gravel content (>2 mm), and soil thickness data at 0–5 cm and 5–15 cm soil depths. Additionally, the data were in the TIFF format, the spatial projection was WGS_1984_Albers, and the spatial resolution was 90 m. The meteorological data (containing temperature, precipitation, and sunshine hours) were obtained from the National Meteorological Science Data Center (http://data.cma.cn, accessed on 18 October 2021), and spatial interpolation was performed using ANUSPLIN software (version 4.2) [28] with the purpose of acquiring the annual average temperature, annual accumulated temperature > 0 °C, annual total precipitation, and annual total solar radiation data in Guangzhou in 2016. The socio-economic data mainly included the chemical fertilizer application rate (CHFE), pesticide application rate (PSTD), film usage (PLSH), distance from field to rural settlement (DFRS), and distance from field to rural road (DFRR). CHFE, PSTD, and PLSH were derived from Guangzhou Statistical Yearbook (http://tjj.gz.gov.cn/, accessed on 13 March 2022). Vectorization was performed by adopting ArcGIS 10.7 software (developed by Environmental Systems Research Institute, Inc., Redlands, CA, USA) to obtain the CHFE, PSTD, and PLSH raster data of cultivated land per unit area in each district of Guangzhou in 2016. DFRS and DFRR were mainly based on the point of interest (POI) and open street map (OSM) data, and were obtained with the use of ArcGIS10.7 software for kernel density and nearest neighbor analyses. The POI data were obtained from the Baidu map development platform (https://lbsyun.baidu.com/, accessed on 18 May 2022), and OSM data were obtained from the Geofabrik’s free download server (https://download.geofabrik.de/, accessed on 18 May 2022).

To facilitate the acquisition of the CLQ evaluation indicator, the cultivated land patch data in Guangzhou were converted into point data, and the geometric center point was used to replace the whole patch. A total of 73,498 data points were obtained, and then, the projections of the above-described remote sensing, soil, meteorological and socio-economic data were converted into GCS_Xian_1980 projections, and their spatial resolutions were resampled to 30 m × 30 m by using ArcGIS 10.7. Finally, the CLQ evaluation indicator of each cultivated land was obtained by extracting and analyzing the data (Table 1).

2.3. Methods

2.3.1. Determining the Optimal Indicators for Evaluating the CLQ

Evaluation indicator selection exerts a vital function in CLQ evaluation. In the current work, first, based on the correlation analysis method, the CLQ evaluation indicator was preliminarily selected by setting the correlation coefficient |r| > 0.1 (p < 0.01), and the correlation analysis formula used is as follows:

$$\mathrm{r}=\frac{{\displaystyle \sum }\left[\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)\right]}{\sqrt{{\displaystyle \sum }{\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)}^2{\displaystyle \sum }{\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2}},$$

(2)

where r refers to the correlation coefficient between the CLQ grade indices (NNGI, NUGI, and NEGI) and the evaluation indicators, ${\mathrm{x}}_{\mathrm{i}}$ is the evaluation indicator value of the ith sample, $\overline{\mathrm{x}}$ represents the average value of the evaluation indicators, ${\mathrm{y}}_{\mathrm{i}}$ refers to the ith sample of the CLQ grade indices, and $\overline{\mathrm{y}}$ indicates the average value of the CLQ grade indices.

Based on the results of correlation analysis, a generalized linear regression model is introduced, and the parameters of the model were performed by adopting the weighted least squares method, and the indicators were selected per the coefficient size of the regression model. The more important the indicator is, the larger the absolute value of the coefficient in the model, whereas the more irrelevant the output variable is, the closer the coefficient is to 0 [29,30]. The generalized linear regression model formula is as follows [31]:

$$\mathrm{y}={\mathrm{g}}^{-1}\left({\mathsf{\omega }}^{\mathrm{T}}\mathrm{x}+\mathrm{b}\right),$$

(3)

where y is CLQ grade indices (NNGI, NUGI, and NEGI), function $\mathrm{g}(·)$ is usually a monotone differentiable coupling function, $\mathsf{\omega }$ is the regression coefficient of the evaluation indicator (setting the regression coefficient of the evaluation indicator as |ω| > 1 to determine final CLQ evaluation indicators [32], x is the different CLQ evaluation indicator corresponding to three different CLQ grade indices, and b is the intercept.

2.3.2. Modeling and Mapping Methods

In this study, a recurrent neural network (RNN) algorithm was used to construct an evaluation model of CLQ. The RNN algorithm mainly includes forward propagation and time backward propagation, as follows:

(1) The forward propagation algorithm of the RNN is shown in Equations (4)–(6) [33]:

$${\mathrm{a}}_{\mathrm{h}}^{\mathrm{t}}={{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{I}}{\mathsf{\omega }}_{\mathrm{ih}}{\mathrm{x}}_{\mathrm{i}}^{\mathrm{t}}+{{\displaystyle \sum }}_{\acute{\mathrm{h}}=1}^{\mathrm{H}}{\mathsf{\omega }}_{{\mathrm{h}}^{\prime }\mathrm{h}}{\mathrm{b}}_{{\mathrm{h}}^{\prime }}^{\mathrm{t}-1},$$

(4)

$${\mathrm{b}}_{\mathrm{h}}^{\mathrm{t}}={\mathrm{f}}_{\mathrm{h}}\left({\mathrm{a}}_{\mathrm{h}}^{\mathrm{t}}\right),$$

(5)

$${\mathrm{a}}_{\mathrm{k}}^{\mathrm{t}}={{\displaystyle \sum }}_{\mathrm{h}=1}^{\mathrm{H}}{\mathsf{\omega }}_{\mathrm{hk}}{\mathrm{b}}_{\mathrm{h}}^{\mathrm{t}},$$

(6)

where ${\mathrm{x}}_{\mathrm{i}}^{\mathrm{t}}$ refers to the value of the ith dimension input at time t, ${\mathsf{\omega }}_{\mathrm{ij}}$ indicates the connection weight of neurons i and j, ${\mathrm{a}}_{\mathrm{j}}^{\mathrm{t}}$ and ${\mathrm{b}}_{\mathrm{j}}^{\mathrm{t}}$ stand for the input value and activation value of neuron j at time t, separately, and ${\mathrm{f}}_{\mathrm{h}}$ signifies the activation function used by neuron h.

(2) The principle of the time backward propagation algorithm of the RNN is consistent with that of the BP algorithm, which includes three steps [33]:

(1)

Calculating the output value of each neuron forward;

(2)

Calculating the error term of each neuron, which refers to the partial derivative of the error function to the weighted input of the neuron, by defining the partial derivative of the loss function to the input value of neuron j at time t, followed by the calculation based on the chain rule. Moreover, the partial derivatives of the loss function to the network weights are shown in Equations (7)–(9):

$${\mathsf{\delta }}_{\mathrm{j}}^{\mathrm{t}}=\frac{\partial \mathrm{L}}{\partial {\mathrm{a}}_{\mathrm{j}}^{\mathrm{t}}},$$

(7)

$${\mathsf{\delta }}_{\mathrm{h}}^{\mathrm{t}}={\mathrm{f}}^{\prime }\left({\mathrm{a}}_{\mathrm{h}}^{\mathrm{t}}\right)\left({{\displaystyle \sum }}_{\mathrm{k}=1}^{\mathrm{K}}{\mathsf{\omega }}_{\mathrm{hk}}{\mathsf{\delta }}_{\mathrm{k}}^{\mathrm{t}}+{{\displaystyle \sum }}_{{\mathrm{h}}^{\prime }=1}^{\mathrm{H}}{\mathsf{\omega }}_{\mathrm{h}{\mathrm{h}}^{\prime }}{\mathsf{\delta }}_{{\mathrm{h}}^{\prime }}^{\mathrm{t}+1}\right),$$

(8)

$$\frac{\partial \mathrm{L}}{\partial {\mathsf{\omega }}_{\mathrm{ij}}}={{\displaystyle \sum }}_{\mathrm{t}=1}^{\mathrm{T}}\frac{\partial \mathrm{L}}{\partial {\mathrm{a}}_{\mathrm{j}}^{\mathrm{t}}}\frac{\partial {\mathrm{a}}_{\mathrm{j}}^{\mathrm{t}}}{\partial {\mathsf{\omega }}_{\mathrm{ij}}}={{\displaystyle \sum }}_{\mathrm{t}=1}^{\mathrm{T}}{\mathsf{\delta }}_{\mathrm{j}}^{\mathrm{t}}{\mathrm{b}}_{\mathrm{i}}^{\mathrm{t}},$$

(9)

(3)

Calculating the gradient of each weight and updating it with an optimization algorithm.

The difference between this algorithm and other ANNs is that the relationship between the loss function and neurons is affected by both the output layer of the current time step (t) as well as the hidden layer of the next time step (t + 1). Using the chain rule, each time step is calculated. Meanwhile, all the findings are supplemented to the time dimension to obtain the partial derivative of the loss function with respect to the weight of the neural network. Then, the weights in the neural network are updated until convergence via gradient descent [33].

Referring to relevant research [23,24,25,26], the CLQ evaluation model constructed by RNN in this study is shown in Figure 2.

The model consists of one input layer, three RNN layers, two fully connected layers, and one output layer. The input layer information is the selected different CLQ evaluation indicators, and the output layer information comprises different CLQ grade indices (NNGI, NUGI, and NEGI), as shown in Figure 2. The model uses the Adam optimizer, and the activation function uses the ReLU function.

2.3.3. Model Accuracy Evaluation

In the current work, three accuracy parameters were chosen with the aim of confirming the performance of the model: coefficient of determination (${\mathrm{R}}^2$), root mean square error (RMSE), and residual prediction deviation (RPD) [10]. Among these parameters, ${\mathrm{R}}^2$ was adopted for measuring the stability of the model; the larger the ${\mathrm{R}}^2$, the more stable the model. RMSE reflected the accuracy of the model; the smaller the RMSE, the better the predictive ability. RPD was applied with the aim of assessing the performance of the model. With RPD > 2, the model had a strong predictive ability; with 1.4 ≤ RPD ≤ 2, the model had a certain predictive ability; and with RPD < 1.4, the model lacked the predictive ability. The expression of the accuracy evaluation index is as follows:

$${\mathrm{R}}^2=1-\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2},$$

(10)

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

(11)

$$\mathrm{RPD}=\frac{\sqrt{\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}})}^2}{\mathrm{n}}}}{\sqrt{\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}{\mathrm{n}}}},$$

(12)

where ${\mathrm{y}}_{\mathrm{i}}$ refers to the measured value of the ith sample, that is, three different CLQ grade indices (NNGI, NUGI, and NEGI), $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ indicates the predicted value of the ith sample, $\overline{\mathrm{y}}$ denotes the measured average value of the sample, and n signifies the number of samples.

3. Results

3.1. Indicator Selection Results

Pearson correlation analysis was carried out on 37 indicators and three different CLQ grade indices (NNGI, NUGI, and NEGI). The correlation analysis findings are presented in Figure 3. The correlation coefficients (r) for the association of these 37 indicators with NNGI, NUGI, and NEGI were −0.41 to 0.40, −0.32 to 0.41, and −0.29 to 0.42, respectively. Positive correlations between PH and NNGI (r = 0.40), EST and NUGI (r = 0.41), and CHFE and NEGI (r = 0.42) were the best. TS had the best negative correlation with NNGI, NUGI, and NEGI (r = −0.41, −0.32, and −0.29, respectively). A total of 13 indicators (PH, EST, CHFE, OMC, TEMA, CEC515, CEC05, SOC515, SOC05, FS, DFRR, GWL, and TS) were significantly related to the correlation coefficient |r| > 0.1 of NNGI. A total of 12 indicators (PH, EST, CHFE, OMC, TEMA, CEC515, CEC05, SOC515, SOC05, FS, GWL and TS) were significantly related to the correlation coefficients |r| > 0.1 of NUGI and NEGI. Thus, PH, EST, CHFE, OMC, TEMA, CEC515, CEC05, SOC515, SOC05, FS, GWL, and TS were closely related to three different CLQ grade indices.

To further verify the relationship between the selected indicators of the correlation analysis and the CLQ grade indices, the generalized linear regression model was introduced. According to the coefficient of the regression model, the indicators were further selected. Figure 4 shows the results. The regression coefficients $\left|\mathsf{\omega }\right|$ of different indicators corresponding to three different CLQ grade indices (NNGI, NUGI, and NEGI) were all greater than 1 (

Table 2. Cultivated land quality evaluation indicators.

NNG Indicators	r	$\mathsf{\omega }$	NUG Indicators	r	$\mathsf{\omega }$	NEG Indicators	r	$\mathsf{\omega }$
PH	0.40	80.57	PH	0.38	89.30	SOC515	−0.12	112.18
SOC515	−0.11	66.15	CEC515	0.11	65.68	SOC05	−0.14	54.91
FS	−0.13	55.39	SOC05	−0.14	49.07	CHFE	0.42	54.81
TS	−0.41	51.15	FS	−0.26	27.56	TEMA	0.17	40.36
CEC515	0.10	32.33	CHFE	0.36	24.66	PH	0.37	36.74
OMC	0.28	30.20	TEMA	0.16	21.30	FS	−0.23	25.09
TEMA	0.17	20.28	CEC05	0.10	16.11	CEC05	0.12	20.33
EST	0.35	−5.60	GWL	−0.24	5.26	GWL	−0.27	−1.60
CHFE	0.32	−7.73	TS	−0.32	−26.04	TS	−0.29	−21.51
DFRR	−0.14	−9.50	OMC	0.20	−34.59	OMC	0.21	−35.68
CEC05	0.10	−21.25	SOC515	−0.12	−77.96	EST	0.41	−45.41
GWL	−0.29	−81.31	EST	0.41	−104.60	CEC515	0.12	−166.58
SOC05	−0.13	−131.43

), indicating that the selected indicators are representative and can be used for the modeling of CLQ evaluation.

3.2. Model Construction and Accuracy Evaluation

Based on previous studies and many experiments [23,24,25,26], the parameters of the CLQ evaluation model constructed in this work were set as follows: the learning rate was determined to 0.001, the epoch was set to 120, batch size was set to 64, the number of neurons in the hidden layer in each RNN layer was set to 1024, and the dense of two full connection layers were set to 64 and 1, respectively. In addition, since the RNN model contains multiple network layers, the problem of overfitting in actual training may exist. Thus, this study used the dropout method to regularize the RNN model, and the regularization process can effectively reduce the complex co-adaptation relationship between neurons, solve the covariance problem among indicators, avoid the model parameters too divergent to converge, and also avoid the phenomenon of overfitting [34]. Dropout was added to each hidden layer and was set to 0.4, that is, 40% of the nodes (and their connected weight connections) were hidden during the training process of the hidden layer with dropout. With the purpose of training the response correlation between different CLQ grade indices (NNGI, NUGI, and NEGI) and the selected CLQ evaluation indicators, the Guangzhou NNGI, NUGI, and NEGI models were constructed. The model accuracy is shown in Figure 5.

As shown in Figure 5, the training set ${\mathrm{R}}^2$ of the NNGI model is 0.91, RMSE is 92.43, and RPD is 3.21; the validation set ${\mathrm{R}}^2$ is 0.91, RMSE is 91.27, and RPD is 3.31; the training set ${\mathrm{R}}^2$ of the NUGI model is 0.82, RMSE is 73.65, and RPD is 2.33; the validation set ${\mathrm{R}}^2$ is 0.83, RMSE is 71.52, and RPD is 2.47; the training set ${\mathrm{R}}^2$ of the NEGI model is 0.84, RMSE is 81.82, and RPD is 2.53; and the validation set ${\mathrm{R}}^2$ is 0.86, RMSE is 77.64, and RPD is 2.71. The results show that the accuracy, stability, and prediction ability of the NNGI, NUGI, and NEGI models in Guangzhou based on the selected CLQ evaluation indicators are better and suitable for constructing the CLQ evaluation model.

3.3. Spatial Distribution and Regional Verification of Cultivated Land Quality

In line with the three CLQ evaluation models, the results of CLQ indices (NNGI, NUGI, and NEGI) of Guangzhou were predicted. Then, the results were linked to the cultivated land patch evaluation unit of Guangzhou by unique coding in ArcGIS 10.7. Finally, the spatial distribution information of different CLQ indices in Guangzhou was obtained. According to the Regulation for Gradation on Agriculture Land Quality in the National Standard of China (GB/T 28407-2012), the CLQ grade was divided. It indicates the larger the CLQ grade index, the better the CLQ grade. The classification criteria are presented in

Table 3. Classification standard of cultivated land quality.

NNG	Range	NUG/NEG	Range
Level 1	NNGI > 5600	Level 4	NUGI/NEGI > 2200
Level 2	5200 < NNGI ≤ 5600	Level 5	2000 < NUGI/NEGI ≤ 2200
Level 3	4800 < NNGI ≤ 5200	Level 6	1800 < NUGI/NEGI ≤ 2000
Level 4	4400 < NNGI ≤ 4800	Level 7	1600 < NUGI/NEGI ≤ 1800
Level 5	4000 < NNGI ≤ 4400	Level 8	1400 < NUGI/NEGI ≤ 1600
Level 6	3600 < NNGI ≤ 4000	Level 9	1200 < NUGI/NEGI ≤ 1400

.

The classification results of three CLQ grades (NNG, NUG, and NEG) in Guangzhou were obtained (Figure 6). NNG in Guangzhou is mainly divided into six grades, among which Level 2 accounts for the largest proportion (up to 50.28%), mainly distributed in Panyu District, Nansha District, northern Baiyun District, and central Huadu District; regions with relatively poor grades (Levels 5 and 6) are mainly distributed in Conghua District. NUG is also mainly divided into six grades, of which Level 6 accounts for the largest proportion (up to 47.68%); the grade is relatively medium, and the distribution is relatively dispersed (it is distributed in all regions except Nansha District). The regions with relatively good grades (Levels 4 and 5) are mostly distributed in Nansha District, the southern region of Panyu District, and the central region of Huadu District, whereas the regions with relatively poor grades (Levels 8 and 9) are mainly scattered in Zengcheng District and Conghua District. NEG is mainly divided into six grades, of which Level 7 accounts for the largest proportion (up to 40.26%); the grade is relatively medium and mainly distributed in Zengcheng District, Huangpu District, and northern Baiyun District. The regions with relatively good grades (Levels 4 and 5) are mainly distributed in Nansha District, whereas the regions with relatively poor grades (Levels 8 and 9) are mainly distributed in the central region of Zengcheng District and the northern region of Conghua District. The reason for this result may be that Conghua District belongs to a mountainous and hilly area, with a large terrain slope, which is prone to soil erosion and relatively poor soil quality. Zengcheng District has a large area of cultivated land but employs extensive management methods and complicated crop rotation. For higher economic benefits, a variety of crops are usually planted in rotation annually, which leads to an overall high utilization intensity of cultivated land in Zengcheng District. The risk of soil acidification and soil heavy metal pollution is higher, resulting in lower NUG [35]. Nansha District and Panyu District are mainly positioned in the alluvial plain of the Pearl River Delta, where the terrain is flat, the thickness of effective soil layer is deep, and the soil is fertile. The soil is mainly loam and sandy soil. There are many water sources, and the water conservancy facilities are good for irrigation and drainage. Simultaneously, these areas (Nansha and Panyu District) are also an important base for the Vegetable Basket Project in Guangzhou. The intensive cultivation of cultivated land in this region generates relatively high quality of cultivated land [36]. Additionally, in line with the Regulation for Gradation on Agriculture Land Quality in the National Standard of China (GB/T 28407-2012), the cultivated land in China is categorized into excellent land, high land, medium land, and low land according to 1–4, 5–8, 9–12, and 13–15 levels. NNG of cultivated land in Guangzhou is categorized as excellent land overall, and NUG and NEG are categorized as high land overall. Compared with other regions in China, the overall quality of cultivated land in Guangzhou reaches the highest.

In addition, with the aim of further confirming the mapping accuracy of the model, the obtained CLQ grades were compared with the actual CLQ grades (NNG, NUG, and NEG) in Guangzhou in 2016, and it was found that the CLQ grade area obtained by the model is consistent with the overall trend of the CLQ grade area in Guangzhou in 2016 (Figure 7). The best and worst cultivated lands occupy a minority area, and the second and third grades of NNG and NUG account for over 80% of the cultivated land area in Guangzhou. The third and fourth grades of NEG account for approximately 70% of the cultivated land area in Guangzhou. Regarding spatial distribution and area proportion of the grade gap (Figure 8 and

Table 4. Area and percentage of cultivated land quality grade gap.

Gap	No Gap (Gap = 0)		Small Gap (Gap = 1)		Medium Gap (Gap = 2)		Big Gap (Gap = 3)
	Area (ha)	Percent (%)	Area (ha)	Percent (%)	Area (ha)	Percent (%)	Area (ha)	Percent (%)
NNG	67,658.204	80.697	16,077.018	19.175	103.654	0.124	3.308	0.004
NUG	69,511.871	82.908	13,907.437	16.588	403.803	0.482	19.072	0.023
NEG	64,069.628	76.417	18,902.307	22.545	838.513	1.000	31.735	0.038

), the areas with no gap (gap = 0) of CLQ NNG, NUG, and NEG in Guangzhou account for 80.697%, 82.908%, and 76.417%, respectively, whereas the areas with a small gap (gap = 1) account for 19.175%, 16.588%, and 22.545%, respectively. CLQ NNG, NUG, and NEG are spatially distributed in regions other than Tianhe District, and the distribution is relatively discrete, which may be because Tianhe District is mainly dominated by urban construction and has less cultivated land resources, which generates little impact on the gap in the quality of cultivated land. Additionally, the area proportions of the medium gap (gap = 2) and large gap (gap = 3) are rarely negligible. The comparative analysis of the gap area proves that the CLQ evaluation model constructed in the current work is reasonable.

4. Discussion

In the existing research, most of the CLQ evaluation indicators refer to the Regulation for Gradation on Agriculture Land Quality in the National Standard of China (GB/T 28407-2012). Data collection of evaluation indicators generally occurs via field sampling and data surveys, which require a lot of time, manpower, and material resources. The obtained data did not meet the current evaluation needs, and the timeliness is poor; hence, achieving rapid CLQ evaluation is challenging [37]. In the present study, multisource data were used to rapidly obtain CLQ evaluation indicators; identify key factors affecting CLQ from four dimensions, namely soil, spatial geography, climatic conditions, and social economy; select the optimal CLQ evaluation indicators; and perform rapid CLQ evaluation.

According to natural factors and social economic factors, the CLQ can be divided into natural, utilization, and economic grades. The CLQ natural grade is mainly the quality of cultivated land in the natural state, reflecting the natural conditions such as soil, geographical conditions, and climatic conditions of cultivated land; the CLQ utilization grade is the result of the CLQ natural grade revised by land use coefficient, which reflects the potential (or theoretical) regional natural quality and average utilization level of cultivated land; the CLQ economic grade is the result of the CLQ utilization grade revised by land economic coefficient, which reflects the potential (or theoretical) regional natural quality, average utilization level, and average investment benefit level of cultivated land. The above three grades reflect the quality of cultivated land in combination, so it is necessary to consider these three grades simultaneously. The results showed that although the evaluation indicators of the three different CLQs are the same, the influence of each indicator on the three types of CLQs are not exactly the same (Table 2). For example, the correlation of OMC, GWL, and TEMA with nature grade is greater than that with utilization and economy grade. The reason is that natural quality mainly reflects the quality of cultivated land soil itself, which shows close relationship to the light, temperature, water, and heat conditions of the environment [38]. The utilization quality of cultivated land mainly reflects the difference in cultivated land suitability and yield, which is closely related to humans’ willingness to cultivate. Therefore, the utilization grade has the strongest correlation with EST. The reason is that the depth of effective soil layer thickness is related to whether the cultivated land is suitable for cultivation, and its correlation with FS and TS is significantly greater than that with the nature and economy grade, mainly due to the unfavorable cultivation in areas with larger FS and TS [38,39]. The economy quality of cultivated land reflects that the input and output ratios of human beings during the process of cultivated land utilization can better reflect the economic benefits of cultivated land; thus, it has the strongest correlation with CHFE. The reason is that the use of chemical fertilizer per unit area can well reflect the economic input of cultivated land [40]. In addition, there is a DFRR indicator in the natural grade, which is most likely due to the long-term infrastructure construction in Guangzhou and the developed traffic road network; when the cultivated land is relatively close to the road, it may be affected by passing vehicles and pedestrians, making the soil surface firm, reducing the original ventilation and water permeability of the cultivated land, and making the cultivation of crops difficult, eventually reducing the quality [41].

Currently, in the CLQ evaluation model, AHP is mostly used to determine the weight, which is a highly subjective parameter [1,14]. With advances in computer technology, machine learning algorithms (such as ANN and SVM) have been used to construct models [18,20], but the spatial information present in the data has not been optimally modeled [22], resulting in the low accuracy of CLQ evaluation at the regional scale. With the aim of improving the accuracy of CLQ evaluation, in the current work, three different CLQ evaluation models (natural, utilization, and economic grade evaluation) were constructed based on the deep-learning RNN algorithm using the existing CLQ evaluation results. The optimal selection of parameters can efficiently and simply complete the evaluation of cultivated land, which can greatly reduce the influence of individual subjective factors in the evaluation process. The model accuracy ${\mathrm{R}}^2$ of the three CLQ evaluation models constructed is greater than 0.80, which meets the high-precision requirements of CLQ evaluation. Accordingly, this study also divides the CLQ grades according to the Regulation for Gradation on Agriculture Land Quality in the National Standard of China (GB/T 28407-2012) and compares them with the actual CLQ grades in Guangzhou. The area and spatial distribution of the three CLQ evaluation grades obtained by the model are consistent with the actual trends; cultivated land with the best and worst quality occupies a minority area, and there is no difference with the actual quality of cultivated land area (above 76%). Compared with previous studies on CLQ evaluation in Guangzhou based on artificial intelligence algorithms [9,15,42,43], the present study is the first to construct a multifunctional CLQ evaluation model in Guangzhou, covering natural, utilization, and economic grades of cultivated land, and the accuracy of this model was higher than that of previous studies. Compared with the actual CLQ of Guangzhou, the results further verified that the RNN algorithm is suitable for CLQ evaluation, which provides a novel academic reference and practical guidance for CLQ evaluation. Moreover, the three CLQ evaluation models can be used in areas where CLQ data are not available, and in addition, it can be used for future land surveys, which can greatly improve work efficiency while ensuring accuracy.

Although this research evaluated the natural, utilization, and economic grades of CLQ in Guangzhou based on multisource data and the RNN algorithm and achieved good evaluation results, there are still some limitations. Only one method—that of combining correlation analysis and the generalized linear regression coefficient—is used in the selection of evaluation indicators. However, some studies have shown that the correlation coefficient can only show the strength of the linear relationship between different variables, and it is difficult to express a nonlinear correlation between variables [10]. As a result, in future research, more selecting of algorithms (such as XGboost and Boruta methods) should be presented and compared to acquire more accurate evaluation indicators of CLQ [44,45], as well as compare the effect of availability of all indicators on the accuracy of the results. In addition, the evaluation indicator system of CLQ is mainly in line with the principle of data availability and quantifiableness, and the representativeness is not strong and cannot reflect the CLQ more scientifically and reasonably. In the follow-up study, we aim to use high-precision data to establish a better CLQ evaluation system to ensure that the evaluation results of CLQ are more objective and authentic.

5. Conclusions

To conclude, a novel method of CLQ evaluation involving multisource data and the RNN algorithm was proposed. First, based on correlation analysis and the generalized linear regression coefficient, evaluation indicators suitable for three different CLQ evaluations (natural, utilization, and economic grade evaluation) were selected from 37 indicators. Subsequently, three different CLQ evaluation models were constructed based on the RNN algorithm, and the multifunctional evaluation of CLQ at the regional scale was performed. On the basis of the findings, the model accuracy ${\mathrm{R}}^2$ of the three CLQ evaluation models constructed in this study is greater than 0.80, and RPD is greater than 2 (NNGI model: ${\mathrm{R}}^2$ = 0.91, RPD = 3.27; NUGI model: ${\mathrm{R}}^2$ = 0.81, RPD = 2.28; NEGI model: ${\mathrm{R}}^2$ = 0.84, RPD = 2.47), which meets the high-precision requirements for CLQ evaluation. The area and spatial distribution of the three types of CLQ evaluation obtained using the model are consistent with the actual trends. Cultivated lands with the best and worst quality occupy a minority area, and the area without a gap with the actual CLQ is more than 76%. Moreover, this study is the first to construct a multifunctional evaluation model of CLQ in Guangzhou, covering the natural, utilization, and economic grades of cultivated land and comparing it with actual CLQ grades. The model results are reliable, indicating that the CLQ evaluation method put forward in the present study is appropriate for macro-regional scale research, and the study findings can offer data support and a decision-making reference for cultivated land protection in Guangzhou.