Xiuliang Jin*, Zhenhai Li, Haikuan Feng, Zhiin Ren, Shaokun Li*

aInstitute of Crop Sciences,Chinese Academy of Agricultural Sciences,Key Laboratory of Crop Physiology and Ecology,Ministry of Agriculture,Beijing 100081,China

bBeijing Research Center for Information Technology in Agriculture,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China

cNortheast Institute of Geography and Agroecology, Chinese Academy of Sciences,Changchun 130102,Jilin,China

Keywords:Biomass estimation Maize Vegetation indices Deep neural network algorithm LAI

ABSTRACT Accurate estimation of biomass is necessary for evaluating crop growth and predicting crop yield.Biomass is also a key trait in increasing grain yield by crop breeding.The aims of this study were (i) to identify the best vegetation indices for estimating maize biomass, (ii) to investigate the relationship between biomass and leaf area index (LAI) at several growth stages, and (iii) to evaluate a biomass model using measured vegetation indices or simulated vegetation indices of Sentinel 2A and LAI using a deep neural network (DNN)algorithm.The results showed that biomass was associated with all vegetation indices.The three-band water index (TBWI) was the best vegetation index for estimating biomass and the corresponding R2,RMSE,and RRMSE were 0.76,2.84 t ha-1,and 38.22%respectively.LAI was highly correlated with biomass (R2 = 0.89, RMSE = 2.27 t ha-1, and RRMSE = 30.55%).Estimated biomass based on 15 hyperspectral vegetation indices was in a high agreement with measured biomass using the DNN algorithm (R2 = 0.83, RMSE = 1.96 t ha-1, and RRMSE = 26.43%). Biomass estimation accuracy was further increased when LAI was combined with the 15 vegetation indices (R2 = 0.91, RMSE = 1.49 t ha-1, and RRMSE =20.05%). Relationships between the hyperspectral vegetation indices and biomass differed from relationships between simulated Sentinel 2A vegetation indices and biomass.Biomass estimation from the hyperspectral vegetation indices was more accurate than that from the simulated Sentinel 2A vegetation indices (R2 = 0.87, RMSE = 1.84 t ha-1, and RRMSE =24.76%).The DNN algorithm was effective in improving the estimation accuracy of biomass.It provides a guideline for estimating biomass of maize using remote sensing technology and the DNN algorithm in this region.

1. Introduction

Biomass is one of the most important crop biophysical indicators during the crop growing season [1-3]. Efficient and accurate regional crop biomass estimation at different stages of crop growth supports agricultural field management and decision making [4-6]. Estimation of biomass during the crop growing season not only leads to more effective crop water irrigation[7,8],fertilizer application[9],and weed and disease management[10,11],but also plays a key role in crop yield prediction[12,13].Biomass is not only a target trait for crop breeding [14-17], but also a key indicator in ecological and climate change studies [18,19].As biomass varies with growth conditions, it is desirable to estimate its value accurately and rapidly.

Biomass is conventionally measured using timeconsuming and destructive methods that are difficult to apply on a regional scale. Remote sensing technologies are a powerful tool for estimating biomass on regional scales.Crop canopy spectral reflectance in the visible and near-infrared regions is influenced mainly by crop canopy structure changes. In recent studies [20-26], crop biomass was highly correlated with several vegetation indices (VIs, different integrations of visible and NIR reflectance) based on multispectral and hyperspectral reflectance. The normalized difference vegetation index (NDVI) was very sensitive to low biomass when the leaf area index (LAI) was less than 2 [22].Compared with the NDVI,the cumulative modified triangular vegetation index 2 showed higher sensitivity at medium to high biomass [25]. VIs based on the reflectance of red-edge bands showed high potential for estimating biomass [27].Synthetic aperture radars (SARs) and light imaging detection and ranging (LiDAR) have been used to monitor biomass for several crops [28-30]. Unmanned aerial vehicles (UAVs) have been increasingly used for their relatively low platform cost,high temporal and spatial resolution, and autonomy [31].Biomass estimation using UAV images has been performed for several crops. Biomass has been accurately estimated in barley with crop surface models (CSMs) and NDVI [32], in maize with CSMs and normalized green-red difference index[33], in winter wheat with wide dynamic range vegetation indices and crop height [34], and in onion with canopy size and crop height [35]. However, although these studies demonstrated the estimation of crop biomass using a variety of remote sensing platforms, they did not consider the relationship between crop LAI and biomass in this process.

Deep neural networks(DNNs)have been shown to achieve excellent results in machine learning,especially in classification. DNNs can efficiently learn from training samples to generate models that permit them subsequently to classify unknown samples with high accuracy. The application of DNNs to regression problems has been further developed with the exception of classification issues. Qui et al. [36] was the first to use DNNs for regression models, obtaining promising results for predicting three electric load demands.Later other studies used DNNs for regression in wind forecasting[37]and for forecasting electric load and natural gas [38,39]. These results indicated that DNNs could be used to solve regression issues. To date, however, little use has been made of regression models for time series forecasting using DNNs,and DNNs have not been applied to crop biomass estimation.

The objectives of this study were (i) to identify the best vegetation indices for estimating maize biomass, (ii) to investigate the relationship between biomass and LAI at several growth stages,and(iii)to evaluate the performance of a biomass model using measured vegetation indices or simulated vegetation indices of Sentinel 2A and LAI based on the DNN algorithm.

2. Material and methods

2.1. Experimental site and setup

A field experiment was conducted in 2012 and 2013, at the Xiaotangshan precision agriculture experimental site(40°10′31″N, 116°26′10″E), Beijing, China. The summer maize cultivars were Nongda 108 and Jinghua 8 in 2012 and Xianyu 335 and Zhengdan 958 in 2013. The planting dates were June 19, 2012 and June 5,2013.The area of each microplot was 100 m2in 2012 and 80 m2in 2013. In 2012, three nitrogen levels of 0, 337, and 675 kg ha-1were applied to produce differing biomass levels.The plant density was 67,500 plants ha-1.In 2013,four nitrogen levels were applied:0,262.5,525.0,and 787.5 kg ha-1and three plant density levels were used:52,500,67,500,and 82,500 plants ha-1. A randomized complete block design with three replications was used. Pests and weeds were controlled according to local practice for summer maize production.

2.2. Field measurement

2.2.1.Canopy hyperspectral reflectance measurement

Hyperspectral reflectance measurements of crops were made at several growth stages (Table 1). Spectral reflectance wasmeasured 1.0 m above the canopy, in a nadir orientation,under clear-sky conditions between 10:00-14:00 local time,using an ASD spectrometer (Analytical Spectral Devices,Boulder, CO, USA) in the spectral region 350-2500 nm. The field of view for the optical fiber of the spectrometer was 25°and the corresponding scanned area was 0.70 m2. Black and baseline reflectances were obtained with a 40 × 40 cm BaSO4calibration panel.The spectral reflectance measured with four replicates in each microplot was used to represent the canopy reflectance of each microplot.For each microplot,a total of 40 spectral datasets were measured. More information about spectral measurement may be found in Jin et al.[5].

Table 1-Hyperspectral reflectance measurement in 2012 and 2013.

2.2.2. LAI and biomass measurement

At the corresponding spectral reflectance measurement positions, LAI was measured four times (at jointing, big trumpet, silking, and grain filling, Table 1) using a LAI-2000 Plant Canopy Analyzer (Li-Cor, Inc., Lincoln, NE, USA) with four replicates in each microplot.More information about LAI measurement may be found in Jin et al. [4]. The mean LAI value was used to represent each microplot.

Biomass was determined at the same four stages and the same positions using a representative series of four consecutive plants, with four replicates in each microplot. All samples were briefly heated to 105 °C and then dried at 70 °C to constant weight.

2.2.3. Selection of vegetation indices and biomass estimation

Fifteen hyperspectral vegetation indices from the literature[23,40-52] were used to estimate biomass during the summer maize growing season. More details of these indices appear are in Table S1. To estimate the relationships between biomass and the vegetation indices, linear or nonlinear regression models were fitted using field data of 2013 (n =132, calibration dataset). The estimation accuracy of the calibration models was evaluated using field data of 2012(n = 72,validation dataset).

2.3. Spectral convolution

The mean orbital altitude of Sentinel 2 is 786 km. The swath width of Sentinel 2 is 290 km.It includes Sentinel 2A and 2B satellites. The Sentinel 2A and 2B satellites were launched on June 23, 2015 and March 7, 2017, respectively.The two satellites revisit the same regions every five days under the same viewing conditions. Sentinel 2 satellite band information is presented in Table S2. To assess the estimation accuracy of Sentinel 2A satellite imagery for summer maize biomass, the canopy ASD spectral reflectance measurement was convolved into the simulated canopy spectral data of Sentinel 2A using the corresponding Sentinel 2A spectral response function (Fig. S1) with MATLAB software (version 2017a, MathWorks, Natick,Massachusetts,USA).

Fourteen simulated Sentinel 2A vegetation indices (Table 2)were determined based on the available bands of Sentinel 2 and the corresponding vegetation index formulas of Table S1.The 14 indices were also used in linear or nonlinear regressions to predict biomass using calibration and validation datasets.

Table 2-Summary of vegetation indices based on available bands of Sentinel 2.

2.4. Deep neural networks

Deep neural networks (DNNs) have shown great success in visual classification tasks [53,54], object localization [55], and speech recognition [56]. In general, a DNN is a feed-forward artificial neural network that has more than one layer of hidden units between its input and output variables [57]. In our study,a DNN was used as a multivariate regression model to build relationships between input and output variables.The training procedure of DNNs includes pre-training and fine tuning [58]. The DNNRegressor function from TensorFlow(https://www.tensorflow.org/) was invoked with a Python(version 3.6,Wilmington, DE, USA)program using four steps:

(1) A two-layer fully connected DNN model with 20, 20 hidden units was built;

(2) A biomass model was trained and tuned using the regressor.fit function based on a calibration dataset; the input_fn was set as lambda and the number of iterations as 10,000;

(3) The regressor.evaluate function was used to evaluate the estimation accuracy of the biomass model based on the validation dataset;

(4) Prediction results were output using the regressor.predict function based on the calibration and validation datasets.

3. Results

3.1. Biomass estimation based on hyperspectral vegetation indices

Linear and nonlinear regressions were fitted to relate biomass with hyperspectral vegetation indices.These relationships are presented in Table 3. The highest and lowest coefficients of determination (R2) (0.76 and 0.13) were found for the threeband water index(TBWI)and wide dynamic range vegetation index (WDRVI), respectively. In descending order of R2, the indices were TBWI, NDII, NDMI, WI II, TCARI, MTCI, TCARI/OSAVI, DCNI I, CIrededge, OSAVI_CIrededge, WI I, OSAVI, NDVI,EVI,and WDRVI.Among the R2values,one was above 0.70,six were above 0.40, and nine were below 0.4. All vegetation indices were fitted to power regression models except WI I,NDVI, and WDRVI (Table 3). The results demonstrated that these vegetation indices were used to estimate biomass.

Table 3-Regressions of maize biomass on hyperspectral vegetation indices(n = 204).

Compared with other vegetation indices, the association between TBWI and biomass was the highest, with root mean square error (RMSE) and relative root mean square error(RRMSE) values of respectively 2.84 t ha-1and 38.22% (Fig. 1,Table 3). Thus, TBWI was the best vegetation index for estimating maize biomass.

3.2. Relationship between biomass and LAI

Dynamic changes of biomass and LAI were investigated to further characterize the relationship between LAI and biomass in summer maize. These changes at several growth stages during 2012 and 2013 are shown in Fig.2.In general,LAI increased to a peak before decreasing (Fig. 2-a, b) while biomass increased steadily with crop development (Fig. 2-c,d). With two summer maize cultivars, four nitrogen levels,and three plant density levels,the growth stages of 2013 were more representative than in 2012. The results indicated that the growth stages in 2013 were very consistent with those in 2012.

To better determine the relationship between LAI and biomass at the stages before grain filling and afterward, the dataset was divided into two parts(Fig.3):the first three data acquisition times (jointing, big trumpet, and silking, Fig. 3-a)and the fourth (grain filling, Fig. 3-b) based on LAI and biomass at four growth stages in 2013. The LAI from the first three data acquisition time and the fourth data acquisition time had a significant correlation with biomass.In this study,the RRMSE was given highest priority and was selected to evaluate the different methods when the three statistics (R2,RMSE, and RRMSE) disagreed. Compared with R2and RMSE,the RRMSE can be better used to fit a regression model of biomass on LAI using normalized relative values(%)[4].Thus,the relationship between LAI and biomass was worse at the first three data acquisition times (Fig. 3-a, RRMSE = 38.87%)than at the fourth(Fig.3-b,RRMSE = 29.83%).

A plot of measured against estimated biomass appears in Fig.4.The calibration result from the two different regression models at the first three data acquisition times (Fig. 3-a) and the fourth data acquisition time (Fig. 3-b), respectively. The two regression models were further validated using the 2012 dataset. The result of calibration was slightly better than the result of validation. It indicated that biomass was better estimated based on the relationship of biomass with LAI at different growth stages. Results demonstrated that the estimated biomass corresponded closely to the measured biomass,in agreement with Ballesteros et al.[35]and Kross et al.[59],who reported a close association between biomass and LAI.

Fig. 1 - Regression of biomass on the three-band-water index (TBWI).

Fig.2- LAI(a, b) and biomass(c,d) at four growth stages.

3.3. Deep neural networks for biomass estimation with and without LAI

Fig.3-Regression of measured biomass on measured LAI:(a)at the first three data acquisition times(jointing,big trumpet,and silking)and (b)at the fourth data acquisition time(grain filling)in 2013.

Fig. 4 - Regressions of measured on estimated biomass in 2012 and 2013.

The 15 biomass vegetation indices were used in the DNN algorithm to better estimate biomass of summer maize.Biomass estimation by the DNN algorithm (Fig. 5-a, R2= 0.83,RMSE = 1.96 t ha-1, and RRMSE = 26.43%) was more accurate than using TBWI (Fig. 1, R2= 0.76, RMSE = 2.84 t ha-1, and RRMSE = 38.22%), and slightly more accurate than using LAI(Fig.4).In addition,the fifteen biomass vegetation indices plus measured LAI were also used into the DNNs algorithm to further increase the estimation accuracy of biomass. The agreement between the estimated and the measured biomass was better based on the fifteen biomass vegetation indices plus LAI (Fig. 5-b) than based on the fifteen biomass vegetation indices. The results demonstrated that the addition of LAI into the fifteen biomass vegetation indices increased the estimation accuracy of the biomass. The estimated biomass agreed very well with the measured biomass using the fifteen biomass vegetation indices, with R2, and RMSE and RRMSE values of 0.83, 1.96 t ha-1, and 26.43%,respectively(Fig.5-a),and with the addition of LAI,the R2, and RMSE and RRMSE became of 0.91, 1.49 t ha-1, and 20.05%, respectively (Fig. 5-b). No matter if with or without inclusion of LAI,the DNN estimated biomass accurately.

3.4. Simulating Sentinel 2 dataset for biomass estimation

Fig.5- Plots of measured biomass against biomass estimated using a deep neural network:(a) without LAI and(b) with LAI.Probability level of 0.01 is indicated by**.

The spectral response function of Sentinel 2A was used to simulate vegetation indices based on available bands of Sentinel 2.Fourteen simulated Sentinel 2A vegetation indices were obtained using the corresponding vegetation index formulas of Table 3. Table 4 presents the regressions of biomass on simulated Sentinel 2A vegetation indices. The lowest and highest R2values(0.13 and 0.73)were found for SWI I and S-TBWI,respectively.Compared with the order of R2values in Table 3,there were some differences.In descending order of R2,the indices were ranked S-TBWI,S-NDII,S-WI II,STCARI, S-MTCI, S-TCARI/OSAVI, S-DCNI I, S-CIrededge, SOSAVI_CIrededge, S-EVI, S-OSAVI, S-WDRVI, S-NDVI, and SWI I.Of the R2values,one was above 0.70,four above 0.40,and ten below 0.4. All simulated Sentinel 2A vegetation indices were fitted to power regression models with the exception of S-WI I, S-NDVI, and S-WDRVI (Table 4). In short, the association of biomass with simulated Sentinel 2A vegetation indices (Table 4) was slightly less predictive than that with vegetation indices(Table 3).Results suggested that simulated Sentinel 2A vegetation indices were able to estimate biomass.

The 14 simulated Sentinel 2A vegetation indices were used with the DNN algorithm.The results showed that the estimated biomass using DNNs algorithm and the fourteen simulated Sentinel 2A vegetation indices was very consistent with the measured biomass (Fig. 6-a, R2= 0.82, RMSE = 2.47 t ha-1, and RRMSE = 33.24%). The agreement between measured and estimated biomass was better using DNNs algorithm (Fig. 6-a)than using the S-TBWI(Table 4),but it was slightly worse than the result using LAI (Fig. 4). LAI added into the fourteen simulated Sentinel 2A vegetation indices further increased the estimation accuracy of biomass (Fig. 6-b, R2= 0.87, RMSE =1.84 t ha-1, and RRMSE = 24.76%). This shows that LAI and simulated Sentinel 2A vegetation indices were able to improve the estimation accuracy of biomass using DNNs algorithm.

4. Discussion

Vegetation indices and concurrent biomass were measured during two maize growing seasons. Fifteen vegetationindices were tested for association with biomass (Table 3),given that vegetation indices from the red edge and nearinfrared regions contain useful information for prediction of vegetation biomass [40,44,51,52]. Specifically, TBWI was strongly associated with biomass with corresponding R2,RMSE, and RRMSE values of 0.76, 2.84 t ha-1, and 38.22%,respectively. The TBWI does not include red edge spectral region because spectral absorption at these wavelengths is strongly affected by chlorophyll content,which decreases the signal compared with that of dry biomass. But the TBWI includes spectral wavelengths at 1447 nm and 1720 nm,which are more sensitive to changes in biomass [23]. These spectral wavelengths are integrated to establish TBWI,which contains signals from dry biomass. In addition, TBWI has spectral wavelengths at 973 nm, which is more sensitive to vegetation water content [23,40]. For these reasons, TBWI is more correlated to biomass than other vegetation indices and is more accurate for estimating biomass. In this study,regression relationships between vegetation indices and biomass were explored to determine the best-fitting regression models using the linear and nonlinear regression method. The results showed that some vegetation index models were suitable for power regression and others were suitable for exponential regression. The difference between two regression models may have a very close relationship with biomass and each vegetation index dataset. Relationships between simulated Sentinel 2A vegetation indices and biomass were consistent with the relationships between the hyperspectral vegetation indices and biomass, but the R2ranking in Table 4 is different from that in Table 3. The difference between the simulated Sentinel 2A vegetation indices and the hyperspectral vegetation indices may originate from the spectral response function of Sentinel 2A (Fig.S1). Although the simulated Sentinel 2A vegetation indices showed some difference from hyperspectral vegetation indices, the results identified strong relationships between the measured biomass and the estimated biomass.

Table 4-Simulated Sentinel 2A vegetation indices based on bands available for estimating biomass of maize (n =204).

Fig.6-Plots of between measured biomass against biomass estimated using a deep neural network and simulated Sentinel 2A vegetation indices:(a) without LAI and(b) with LAI.Probability level of 0.01 is indicated by**.

To better estimate maize biomass, 15 biomass vegetation indices or 15 biomass vegetation indices plus LAI were used in the DNN algorithm. The biomass estimated based on the 15 biomass vegetation indices was consistent with the measured biomass using the DNN algorithm(Fig.5-a,R2= 0.83,RMSE =1.96 t ha-1, and RRMSE = 26.43%). This may be because the DNN algorithm comprehensively considers the relationship between each vegetation index and biomass. It showed enhanced sensitivity to variation in biomass.Previous studies[36-39] indicated that the DNN algorithm could be used to better mine effective information from datasets to accurately estimate a parameter of interest. Thus, the DNN algorithm gave a more accurate estimation of biomass. In addition,combination of the LAI with the 15 biomass vegetation indices further improved the estimation accuracy of biomass(Fig.5-b,R2= 0.91, RMSE = 1.49 t ha-1, and RRMSE = 20.05%). This is because LAI is very important canopy structure information in crops,and LAI is highly correlated with biomass(Fig.3,Fig.4).Our results are consistent with those of Ballesteros et al. [35]and Kross et al. [59]. Thus, added LAI information can be obtained using the DNN algorithm and then used to further increase biomass estimation accuracy. In summary, the DNN algorithm was a useful tool for increasing biomass estimation accuracy. Similarly, the DNN algorithm was used to estimate biomass using 14 simulated Sentinel 2A vegetation indices or the 14 simulated Sentinel 2A vegetation indices plus LAI.The agreement between the measured and estimated biomass in Fig.6 was slightly weaker than that in Fig.5.The reason is that(i) the 14 simulated Sentinel 2A vegetation indices did not include the normalized difference matter index(NDMI)which is very sensitive to changes in biomass because of its inclusion of spectral wavelengths at 1649 nm and 1722 nm[43], (ii) the sensitivity of simulated Sentinel 2A vegetation indices to biomass is influenced by the Sentinel 2A spectral response function because the reflectance of a specific wavelength is changed into the reflectance of a spectral range, (iii) and the 15 hyperspectral vegetation indices contained more effective spectral information for biomass than the 14 simulated Sentinel 2A vegetation indices. However, the results suggested that the simulated vegetation indices from Sentinel 2A can be used to estimate biomass and obtain good results using the DNN algorithm. When the respective RMSE and RRMSE values of estimated biomass are lower than 2 t ha-1and 25%,based on our results,this method can be used to estimate biomass in practical applications. In this study, for consistency with the 15 hyperspectral vegetation indices, the 14 simulated Sentinel 2A vegetation indices were calculated based on the limited Sentinel 2 satellite bands. Because previous studies [40-52] showed that the 14 simulated Sentinel 2A vegetation indices were correlated with biomass,the 14 simulated Sentinel 2A vegetation indices were selected to model without use of a feature selection method.We plan to use feature selection methods to simplify the model based on the relationship between vegetation indices and biomass in the future.

In this study, field hyperspectral data were measured to improve the estimation accuracy of biomass. Simulated Sentinel 2A vegetation indices were used to evaluate the available Sentinel 2A satellite for biomass estimation using the spectral response function of Sentinel 2A. Satisfactory results for biomass estimation were achieved (Fig. 5, Fig. 6).This study focused only on the simulated Sentinel 2A vegetation indices for estimating biomass. Sentinel-2 satellites can provide 10-60 m spatial resolution imagery with revisit periods of five days based on different spectral bands.So far we have not considered the impact of spatial resolution of Sentinel-2 satellite imagery on biomass estimation, and future research should explore this issue. The simulated Sentinel 2A vegetation indices and the hyperspectral vegetation indices differ, in that the actual Sentinel 2A vegetation indices are affected by imagery requisition environments(such as atmospheric environment and illumination environment). Therefore, actual Sentinel 2A satellite data must be applied in future research to evaluate the biomass estimation model at regional scales. In the present study, LAI was obtained using LAI-2000, which is expensive for most scientists.The cost of LAI measurement will be sharply reduced by conventional RGB digital camera and image processing algorithms. It will be better matched with hyperspectral data measurement position to maintain the operability and stability of our method using DNNs algorithm in the future.With the fast development of UAV, the integration of UAV platforms and hyperspectral imagery allows us to estimate crop biomass rapidly with very high spatial resolution imagery at the farm or field scale. In order to make our methods more practical,future UAV hyperspectral imagery or hyperspectral satellite imagery could be adapted to better analyze the stability and accuracy of the biomass model. In addition, SARs, LiDAR, and CSMs can be combined with hyperspectral imagery to better estimate and monitor biomass for different crops.We performed only a single-site experiment over two years, and gained good results. To transfer this approach to other sites in the future, we will carefully test these results in different ecological regions based on sufficient independent datasets and improve the reliability and spatial transferability of biomass model, given that the present study was limited to summer maize in Beijing,China.

5. Conclusions

In our study, the deep neural network (DNN) algorithm was used to improve the estimation accuracy of maize biomass of several cultivars under several plant densities and nitrogen levels over two years. The results showed that some vegetation indices were highly correlated with maize biomass. The three-band water index(TBWI) was the best biomass estimation model with corresponding R2,RMSE and RRMSE values of 0.76, 2.84 t ha-1, and 38.22% respectively. LAI was highly significantly associated with biomass (R2= 0.89, RMSE =2.27 t ha-1, and RRMSE = 30.55%). Estimated biomass based on 15 biomass vegetation indices was consistent with the measured biomass using the DNN algorithm (R2= 0.83,RMSE = 1.96 t ha-1, and RRMSE = 26.43%). Adding LAI to the 15 biomass vegetation indices further increased biomass estimation accuracy (R2= 0.91, RMSE = 1.49 t ha-1, and RRMSE = 20.05%). Relationships between biomass and simulated Sentinel 2A vegetation indices showed some differences from the relationships between biomass and hyperspectral vegetation indices. Biomass estimation results from the simulated Sentinel 2A vegetation indices were slightly worse than those from hyperspectral vegetation indices. The DNN algorithm is an effective approach for increasing biomass estimation accuracy. Results of this study could be used as a good guideline for estimating biomass of summer maize based on remote sensing technology and the DNN algorithm.

Supplementary data for this article can be found online at https://doi.org/10.1016/j.cj.2019.06.005.

Declaration of Competing Interest

The authors declared that they have no conflicts of interest to this manuscript. We declare that we do not have any commercial or associative interest that represents a conflict of interest in connection with the work submitted.

Acknowledgments

This study was supported by the National Natural Science Foundation of China (41601369) and the Young Talents Program of Institute of Crop Sciences, Chinese Academy of Agricultural Sciences(S2019YC04).We are grateful to staff for the collection of field data.