**Crop Growth Stage GPP-Driven Spectral Model for Evaluation of Cultivated Land Quality Using GA-BPNN**

**Mingbang Zhu <sup>1</sup> , Shanshan Liu <sup>1</sup> , Ziqing Xia <sup>1</sup> , Guangxing Wang <sup>2</sup> , Yueming Hu 1,3,4,5,\* and Zhenhua Liu <sup>1</sup>**


Received: 28 June 2020; Accepted: 30 July 2020; Published: 1 August 2020

**Abstract:** Rapid and accurate evaluation of cultivated land quality (CLQ) using remotely sensed images plays an important role for national food security and social stability. Current approaches for evaluating CLQ do not consider spectral response relationships between CLQ and spectral indicators based on crop growth stages. This study aimed to propose an accurate spectral model to evaluate CLQ based on late rice phenology. In order to increase the accuracy of evaluation, the Empirical Bayes Kriging (EBK) interpolation was first performed to scale down gross primary production (GPP) products from a 500 m spatial resolution to 30 m. As an indicator, the ability of MODIS-GPPs from critical growth stages (tillering, jointing, heading, and maturity stages) was then investigated by combining Pearson correlation analysis and variance inflation factor (VIF) to select the phases of CLQ evaluation. Finally, a linear Partial Least Squares Regression (PLSR) and two nonlinear models, including Support Vector Regression (SVR) and Genetic Algorithm-Based Back Propagation Neural Network (GA-BPNN), were driven to develop an accurate spectral model of evaluating CLQ based on MODIS-GPPs. The models were tested and compared in the Conghua and Zengcheng districts of Guangzhou City, Guangdong, China. The results showed that based on field measured GPP data, the validation accuracy of 30 m spatial resolution MODIS GPP products with a root mean square error (RMSE) of 7.43 and normalized RMSE (NRMSE) of 1.59% was higher than that of the 500 m MODIS GPP products, indicating that the downscaled 30 m MODIS GPP products by EBK were more appropriate than the 500 m products. Compared with PLSR (R<sup>2</sup> = 0.38 and RMSE = 87.97) and SVR (R<sup>2</sup> = 0.64 and RMSE = 64.38), the GA-BPNN model (R<sup>2</sup> = 0.69 and RMSE = 60.12) was more accurate to evaluate CLQ, implying a non-linear relationship of CLQ with the GPP spectral indicator. This is the first study to improve the accuracy of estimating CLQ using the rice growth stage GPP-driven spectral model by GA-BPNN and can thus advance the literature in this field.

**Keywords:** CLQ; GA-BPNN; GPP-driven spectral model; rice phenology; EBK

## **1. Introduction**

Cultivated land quality (CLQ) has significant influence on agricultural production and resident living [1–3]. The CLQ often changes dramatically under conditions of human disturbances or environments [4]. Thus, rapid and accurate quantification of CLQ is critical. Traditionally, the assessment of CLQ is usually conducted using field measurements, which is time-consuming and costly. More importantly, this method lacks the ability to generate spatial distributions of CLQ [5–7]. Using remotely sensed data offers the potential of obtaining accurate and spatially explicit estimates of CLQ with low cost and has attracted the attention of scholars [8–12].

Current studies on CLQ evaluation using satellite imagery can be divided into two categories: traditional CLQ evaluation methods and pressure-state-response (PSR) based approaches. In the former group of CLQ evaluation methods, remote sensing data were only utilized to obtain some traditional indicators of CLQ, such as soil properties. One typical example is the study of Yang et al. (2012) in which Landsat TM images were used to derive soil organic matter, soil acidity, soil texture, and then generate the estimates of CLQ based on gradation regulations on the quality of farmland in China [13]. Instead of soil fertilizer variables, Zhao et al. (2012) utilized normalized difference vegetation index from Landsat TM imagery to evaluate CLQ [14]. However, the evaluation efficiency of CLQ using satellite image-driven evaluation methods is limited because the use of the methods is dependent on field measurements.

While, of the PSR based methods, CLQ is directly evaluated using remote sensing spectral indicators. For example, Liu et al. (2010) developed a linear model for evaluating CLQ based on predictors, including slope, sandy area ratio in a pixel, and modified soil-adjusted vegetation index [4,15]. Liu et al. (2019) generated the spatial distribution of CLQ estimates based on the Genetic Algorithm-Based Back Propagation Neural Network (GA-BPNN) model. The authors utilized five remote sensing data derived predictors, including Slope, Vegetation Index, Temperature Vegetation Dryness Index, Road Accessibility, and Patch Fractal Dimension, and found that CLQ was significantly and nonlinearly correlated with the spectral predictors [16]. Xie et al. (2018) developed a frequent pattern-growth algorithm for improving the evaluation efficiency of CLQ [17]. At present, there have been no reports about CLQ evaluation using gross primary production (GPP). Some scholars have used GPP to evaluate cultivated land productivity. Ma et al. (2018) explored the estimation of cultivated land productivity using the mean GPP from 2000 to 2018 and analyzed the change trend and amplitude of cultivated land productivity, implying that GPP provided the potential to evaluate CLQ [18]. Although the studies demonstrated the possibility of rapidly evaluating CLQ, the spectral responses between CLQ and remote sensing indicators were ignored. Moreover, the prediction accuracy of CLQ is affected by the selected spectral indicators in the PSR framework due to the limitations of image spatial and spectral resolutions.

This study aimed to propose an accurate spectral response model of CLQ based on GPP spectral indicators from the MODIS-GPPs from 2011 to 2015 at different growth stages of late rice and the corresponding temporal CLQ for mapping CLQ. Here, CLQ is defined as the farmland utilization quality grade and represents the degree of anthropogenic use and natural conditions of cultivated land [19]. The Empirical Bayes Kriging (EBK) interpolation was first employed to perform spatial downscaling transformation of the MODIS GPP images from 500 m spatial resolution to 30 m. The accurate spectral model based on GA-BPNN was then developed and validated by comparing it with a linear partial least squares regression (PLSR) and a non-linear support vector regression (SVR). The comparison of the models was made in the study area mentioned next. It is expected that the study can offer a powerful tool to rapidly and accurately estimate CLQ.

### **2. Materials and Methods**

### *2.1. Study Areas*

The study area (Figure 1) is situated in the Conghua and Zengcheng District of Guangzhou, Guangdong of China (22◦260–23◦560 N, 112◦570–114◦030 E). The annual average temperature is 21 ◦C and the annual average precipitation is about 1900 mm, concentrating between April and September [20–22]. The cultivated land area of Guangzhou in 2015 was 13,485.99 hm<sup>2</sup> with an annual crop yield of 440,900 tons (referencing the Statistical Communiqué of Guangzhou on the 2015 National Economic and Social Development), of which the paddy field was 11,885.16 hm<sup>2</sup> with the average paddy stand size of about 0.25 hm<sup>2</sup> , accounting for 87.91% of the total cultivated land area. The cultivated land is mainly concentrated in Conghua and Zengcheng Districts. Rice is the principal crop in the study area, with an annual double rotation system (early rice: March–June and late rice: August–November). *Agriculture* **2020**, *10*, x FOR PEER REVIEW 3 of 16 [20–22]. The cultivated land area of Guangzhou in 2015 was 13,485.99 hm2 with an annual crop yield of 440,900 tons (referencing the Statistical Communiqué of Guangzhou on the 2015 National Economic and Social Development), of which the paddy field was 11,885.16 hm2 with the average paddy stand size of about 0.25 hm2, accounting for 87.91% of the total cultivated land area. The cultivated land is mainly concentrated in Conghua and Zengcheng Districts. Rice is the principal crop in the study area, with an annual double rotation system (early rice: March–June and late rice: August–November).

**Figure 1.** The study area location in Guangzhou City (**a**), and Conghua and Zengcheng District (**b**), respectively, with a total of 420 sample plots for cultivated land quality (CLQ) (training sample plots are in yellow and validation sample plots for the model in purple) and another set of 240 sample plots in red for validating multi-scale Moderate-resolution Imaging Spectroradiometer (MODIS) gross primary production (GPP) products; (**c**,**d**) the validation areas for mapping at Aotou Town of Conghua District and Zhongxin Town of Zengcheng District. **Figure 1.** The study area location in Guangzhou City (**a**), and Conghua and Zengcheng District (**b**), respectively, with a total of 420 sample plots for cultivated land quality (CLQ) (training sample plots are in yellow and validation sample plots for the model in purple) and another set of 240 sample plots in red for validating multi-scale Moderate-resolution Imaging Spectroradiometer (MODIS) gross primary production (GPP) products; (**c**,**d**) the validation areas for mapping at Aotou Town of Conghua District and Zhongxin Town of Zengcheng District.

### *2.2. Data 2.2. Data*

In this study, to match the 500 m spatial resolution of MODIS GPP products, field sampling plots of 500 m by 500 m were designed and within each plot, biomass was taken at five locations with one placed at the plot center and the other four at the middle points from the plot center to the corners along the diagonal lines. The GPP was estimated with an empirical regression model [23]:G = ே .ହଶସ, In this study, to match the 500 m spatial resolution of MODIS GPP products, field sampling plots of 500 m by 500 m were designed and within each plot, biomass was taken at five locations with one placed at the plot center and the other four at the middle points from the plot center to the corners

where NPP was acquired from the following equation.

along the diagonal lines. The GPP was estimated with an empirical regression model [23]: G*PP* = *NPP* 0.524 , where NPP was acquired from the following equation.

$$NPP = \frac{B \times \beta}{\alpha} \tag{1}$$

where *B* is the dry biomass obtained from each of the sampled plots at the heading stage and dried in a constant temperature drying oven at 110 ◦C. The α is the ratio of aboveground biomass to total biomass, and β is the percentage of carbon in the biomass. For cultivated land, the value of α is usually 0.8 [24], and β is 45% [25]. The rice samples were first destructively collected in the field sample plots and oven-dried at 110 ◦C for 50 min in the laboratory, and the temperature was then adjusted to 85 ◦C for 10 h until the sample weights did not change. The samples were finally taken out and weighed.

Moreover, a total of 660 sample data were extracted from the CLQ map obtained from the Guangzhou National Land Department and the CLQ values were derived using gradation regulations on farmland quality in China (Regulation for gradation on agriculture land quality GB/T 28407-2012) [26]. The plot sampled area is 30 m × 30 m, matching the 30 m spatial resolution of the downscaled MODIS products. The 660 samples were randomly divided into three groups: 294 samples in yellow for modeling (Figure 1b), 126 sample in purple for validation of the estimated CLQ (Figure 1b), and another dataset of 240 sample plots in red for assessing the accuracy of mapping CLQ at the regional scale (Figure 1c,d). In addition, 2011–2015 MODIS/Terra 8-day GPP products (MOD17A2H Version 6) at a spatial resolution of 500 m × 500 m were acquired from the Land Processes Distributed Active Archive Center (LP DAAC/NASA). The MODIS Re-projection Tool (MRT) was employed to convert the sinusoidal projection into the Albers Equal Area projection for the MODIS GPP products. The scaling factor of 0.1 was utilized to obtain the standard MODIS-GPP products [27].

According to the rice growth phases recommended by Ricepedia [28], the rice growth process can be characterized by five stages: seedling, tillering, jointing, heading, and maturity. In fact, the seedling stage was not taken into account because of the too small values of MODIS-GPPs detected and the impact of water on the spectral reflectance of rice. Thus, four growth stages were considered. Moreover, the dates of the acquired MODIS-GPP images with cloud cover less than 5% were consistent with the times of the four rice growth stages (Table 1) and the synchronous satellite-field experiment was carried out.


**Table 1.** Acquisition dates of MODIS-GPPs and corresponding with rice growth stages.
