Tea cultivar classification and biochemical parameter estimation from hyperspectral imagery obtained by UAV

Study area

The study site is located in the Southlake Tea Plantation (latitude 30°28′N, longitude 114°21′E) on the eastern Jianghan Plain, Wuhan, China. The 30-year (1986–2016) average annual precipitation and sunlight are 1,000 mm and 1,850 h, respectively (China Meteorological Administration). An experimental field containing eight tea cultivars, namely Fuding Dabai, Tai Cha 12#, Huang Dan, Mei Zhan, Tie Guanyin, Wu Niuzao, Ying Shuang, and Fuan Dabai, was selected as the study area (Fig. 1).

Hyperspectral data

Hyperspectral images were acquired using an octocopter UAV (S1000+; DJI, Inc., Shenzhen, China) equipped with a Cubert UHD185 hyperspectral imager, (Cubert, Inc., Ulm, Germany) (Fig. 2). The imager consists of a Sony hyperspectral camera (ICX285, Sony, Inc., Tokyo, Japan), a mini-PC server, a battery power system, and the Cubert suite software.

The hyperspectral camera provided hyperspectral images in the visible and near infrared wave range. The spectral resolution of the camera was 8 nm (@532 nm) and the spectral sampling interval was 4 nm, resulting in 138 spectral channels over wavelength of 450–998 nm for each pixel. Moreover, three bands at wavelengths of 450, 550, and 650 nm share a spatial resolution of 1,000 × 1,000 pixels per frame. The other 135 bands have a spatial resolution of 50 × 50 pixels per frame. These bands were later fused to give 1,000 × 1,000 pixels through a multi-resolution image fusion algorithm embedded in the software accompanying the UHD camera.

A flight with the UAV mapping system was carried out between 10:30 and 12:00 a.m. under clear sky conditions on March 23, 2016. Before the flight, two flight lines were designed and ground control points were set over rows of tea trees to mark sampling sites for foliage collection (Fig. 1). The sensor was fixed so that the imagery was collected at nadir to enable collection of ortho-images. As the camera’s field of view angle was 29.4°, the flight height was set to 100 m, which resulted in each image frame covering an area of 54 × 54 m, with a nadir on-the-ground spatial resolution of 0.054 m. For stability of camera, the errors within the actual flight height and set height was not more than 0.1 m. The exposure time of the camera was set to 5 ms and the ISO was set to 100; therefore, the ground targets could be captured clearly while the reference of white panel in the image were not overexposed. The forward overlap and the side overlap of any two adjacent images were more than 65% and 40% respectively, and therefore most of the ground pixels were covered by at least two images.

Foliage concentration measurement

Tea foliar chemistries of 85 samples were measured on the same day as the flight. Acquired samples were sent to a wet chemistry laboratory to measure the concentration of TP and AA. These samples were collected from 85 locations marked in Fig. 1. At each sampling point, 60–70 leaves from the canopy were collected to meet the minimum requirement with regard to foliar quantity for laboratory measurements. Statistical descriptions of TP, AA, and P/A in the study site are shown in Table 1.

All leaf samples from the canopy were dried for 1 h at 80 °C and then ground into fine tea powder. The concentration of TP was measured by the Folin–Ciocalteu colorimetry method (Singleton, Orthofer & Lamuela-Raventos, 1999; Li & Wang, 2009; Büyükbalci & El, 2008; Schulz et al., 1999; Soultani et al., 2014; Iwasa & Torii, 1962). The AA concentration was measured using the ninhydrin colorimetry method (Lee & Takahashi, 1966; Rosen, 1957).

Image pre-processing

Pre-processing was subsequently applied to the acquired hyperspectral images. The pre-processing was composed by the following steps:

Radiometric calibration and image selection: in this step, raw radiance spectra were converted into reflectance spectra. The UAV hyperspectral image was radiometrically corrected with reference measurements on a white board and dark measurements by covering the black plastic lens cap. The dark measurements were subtracted from the reference measurements and the actual measured values. The white board was measured twice before and after the flight. If the difference between the two measurements was larger than 5%, which means the light condition during the flight was changed significantly, the data collected during the flight was abandoned. This procedure allows us to make sure that the radiance collected are correctly translated to reflectance during the image acquisition process. After that, blurred and redundant images were discarded on the basis of visual interpretation, leaving 12 clear images covering the whole study area.

Image Fusion: due to the different spatial resolutions in different bands, image fusion was applied using the Pan-Sharpening algorithm integrated in the Cube-Pilot software (Cubert GmbH, 2014). After performing this image fusion technique, 12 image cubes with 1,000 × 1,000 pixels and 138 spectra ranging from 400–950 nm in each cube were available.

Mosaic and crop: Fused images were processed (images were aligned, a dense point cloud was built, and a mesh was built) and combined with a larger ortho-mosaic image using Agisoft PhotoScan (v1.2.4; Agisoft, Inc., St. Petersburg, Russia) (Agisoft, 2016). The large ortho-mosaic image was cropped to the study region in Envi (5.1; Harris Geospatial Solutions, Inc., Broomfield, CO, USA).

Tea canopy extraction: Although the average leave area index (LAI) of tea plants is relatively high (LAI > 3). There is still a possibility that the background soil spectra mix with the spectra of tea canopy. In order to distinguish tea plants from the soil and dry branches in the hyperspectral images, the optimized soil-adjusted vegetation index (OSAVI) was utilized as an effective index for extracting the plant canopy from the soil background. OSAVI is a soil-adjusted vegetation index optimized for agricultural monitoring, in Eq. (1).

Rondeau (Rondeaux, Steven & Baret, 1996) summarized the formulation of OSAVI and noted that the proper value of Coe is 0.16. R800 and R670 represent the reflectance at wavelengths of 800 and 670 nm. In this study, all pixels with OSAVI larger than 0.65 were considered as tea canopy.

Dimensionality reduction: hyperspectral information is often affected by the “Hughes phenomenon,” where the classification or prediction accuracy is reduced due to the vast spectral dimension (Hughes, 1968). To avoid this effect, the raw spectra were compressed into a lower dimension by three to-be-determined preprocessing methods: minimum noise fraction (MNF) transformation, principal component analysis (PCA) and independent component analysis (ICA). These methods can decompose the spectra into statistical independent components.

Tea cultivar classification

Supervised classifications were applied in this research. For each tea cultivar, 10–15 regions of interest were used to extract the spectral signatures for classification.

A total of four classification approaches, including both statistical model and machine learning based methods, were compared to find the possibility of tea cultivar classification and to explore the optimal methods. A total of four selected methods were maximum likelihood classification (MLC), minimum distance classification (MDC), artificial neural network (ANN) classification, and support vector machine (SVM) classification.

Validation of tea cultivar classification

To validate the classification accuracy, tea cultivar types at 500 randomly distributed points in the image were identified during a fieldwork. The confusion matrix (Stehman, 1997) was then calculated by comparing the ground truth data with the classification results. The overall accuracy (OA) and Kappa coefficient (Kappa) in the confusion matrix indicate the percentage of correctly classified pixels and the effectiveness of the overall classification. OA and Kappa are described by Eqs. (2) and (3) as follows: (2) $\text{OA}=\frac{{\displaystyle \sum _{i=1}^T}{X}_{ii}}{n}$ (3) $\text{Kappa}=\frac{n\ast {{\displaystyle \sum }}_{i=1}^{T}Xii-{{\displaystyle \sum }}_{i=1}^T{{\displaystyle \sum }}_{j=1}^{T}Xij}{n^2-{{\displaystyle \sum }}_{i=1}^T{{\displaystyle \sum }}_{j=1}^{T}Xij}$

In Eqs. (2) and (3), T is the number of classes; Xii represent the correctly classified pixels in class i; Xij represent the incorrectly classified pixels in class i; and n represents the number of pixels participating in the classification.

Leaf biochemistry prediction from canopy

In order to establish a robust biochemical prediction model, several pre-processing methods were applied to the spectra, including the Wavelet De-noising method, continuum removal, standard normal variant (SNV), first derivative, and second derivative.

The spectral bands centered at 450 nm and between 954 and 998 nm were excluded, because these bands were considered noisy with a low SNR (36). Thus, spectra between 450 and 954 nm were left for modeling.

Partial least squares regression (PLSR) was performed to establish the relationship between the reflectance and the foliar chemistry of tea leaves, because it has proven accuracy and effectiveness when solving small samples and high dimensional regression problems (Shi et al., 2014; Liu et al., 2014). Besides, the predictive ability and stability of the PLSR model are often based on the selection of a number of factors (Asyraf & Afthanorhan, 2013).

Artificial neural network has shown good performances in solving non-linear regression and has high error tolerance (Keiner & Yan, 1998). To tackle the possible nonlinearity in the biochemical parameter estimation, a back-propagation ANN (BP-ANN) was tried with the same prepressing methods, and the best results were obtained by using the PLS method. The modeling work was coded with Matlab® (R2014a; MathWorks, Inc., Natick, MA, USA).

The leave-one-out cross-validation method was employed to estimate the accuracy of the model to prevent over-fitting during training (Kohavi, 1995). All samples were divided into several parts: 84 samples were used as the calibration data set for regression and one sample was used as the validation data set to validate the results. The regression did not stop until every sample was assigned to the validation set once. The models were evaluated by comparing the average correlation coefficient of cross-validation (R_CV) and the averaged root mean squared error of cross-validation (RMSE_CV). R_CV and RMSE_CV measure how well the model fits the data: the larger the value of R_CV and the smaller the value of RMSE_CV, the greater the precision and accuracy of the model. R_CV and RMSE_CV are computed by Eqs. (4) and (5), (4) $R_{\text{cv}}=\frac{{{\displaystyle \sum }}_{i=1}^N\left({\widehat{y}}_i-\text{mean}\left(\widehat{y_i}\right)\right)\left(y_i-\text{mean}\left(y_i\right)\right)}{\sqrt{{{\displaystyle \sum }}_i^N{\left[{\widehat{y}}_i-\text{mean}\left({\widehat{y}}_i\right)\right]}^2\ast {\left[y_i-\text{mean}\left(y_i\right)\right]}^2}}$ (5) ${\text{RMSE}}_{\text{CV}}=\sqrt{\frac{1}{N}\ast {\displaystyle \sum _{i=1}^N}{\left({\widehat{y}}_i-y_i\right)}^2}$where N is the number of samples, $\widehat{y_i}$ and y_i and are the predicted and measured values of biochemical components of each sample. The mean ($\widehat{y_i}$) and mean (y_i) represent the average values of $\widehat{y_i}$ and y_i.

Classification of tea cultivars

The tea cultivar classification maps were generated from the hyperspectral data by using full wave bands with different models of MLC, MDC, ANN, and SVM, as shown in Figs. 3A–3D. The results obtained by employing the dimension reduction methods of MNF transformation (e–h), PCA (i–l), and ICA (m–p) are shown in the second to the fourth rows in the same figure.

The confusion matrix is calculated based on the classification results against the 500 validation points. The OA and Kappa accuracy evaluation indices are shown in Table 2. With the full-band spectra, a relatively low classification accuracy is achieved by using MLC and MDC, which result in OAs of 48.4 and 35.2%, and values of the Kappa coefficient of 41.5% and 26.0%, respectively (Table 2). MLC and MDC do not clearly distinguish between the eight cultivars of tea. The MLC algorithm classifies almost all of the tea plants as Fuding Dabai cultivar, and MDC classifies most of them as Ying Shuang. ANN and SVM give better classification results than MLC and MDC according to the OA. Cultivar boundaries are obvious under ANN and SVM classification, and the spatial distribution of tea cultivars is basically correct (Figs. 3C and 3D). ANN shows poor ability to identify the cultivar Huang Dan, while SVM performs better. However, it is time consuming to carry out ANN and SVM classification on the hyperspectral image cube on a PC (220 and 420 min on a 3-GHZ 24-core workstation).

Minimum noise fraction analysis revealed that the top 10 bands possessed more than 90% of the information in the full-band spectra; therefore, these 10 bands were selected as the new input spectra. As shown in Figs. 3E–3G, the classification results of MLC and MDC are much better than those without MNF transformation. The values of OA and Kappa are twice those obtained without dimensionality reduction (Table 2). As shown in Table 2, worse values of OA and Kappa were obtained with the ANN and SVM methods compared to cases in which the full spectra were used. With dimension reduction of the spectra, the processing time decreased to less than 60 min when using the ANN and SVM algorithms. In brief, MNF can improve the classification accuracy dramatically in statistical models like MLC and MDC but it fails in machine learning algorithms like ANN and SVM.

Through PCA analysis, it was found that the first three principle components contained nearly 90% of the full-band information; therefore, the first three components were used for classification. With PCA, the classification results were similar to the results of MNF, especially for the statistical models such as MLC and MDC, but significantly better results were obtained by ANN and SVM classification. As only three instead of 138 bands were used, the classification efficiency is much higher than without PCA and MNF transformation and the computational expense is reasonable: less than 30 min when using the ANN and SVM algorithms.

When we performed the forward ICA transformation, 10 independent components were used to replace the full 138 bands. After ICA transformation, SVM still obtains the best classification result among the modeling methods. In general, MLC, MDC, ANN, and SVM show similar classification accuracy after PCA transformation.

Among the various classification methods, SVM classification always shows the best result, while ANN classification shows slightly poorer performance. The results of MLC and MDC are worse than those of the above machine learning algorithms. By reducing the dimension of hyperspectral data, the classification accuracy of statistical classification improves considerably, while the improvement in classification accuracy is not obvious for machine learning algorithms.

Prediction of foliar biochemistry

The results indicate that the pre-processing methods have a significant impact on the prediction accuracy (Table 3). The best calibration results for TP, AA, and P/A were achieved by using the SNV pre-processing method (R_CV = 0.58, 0.62, 0.66 and RMSE_CV = 10.01, 1.16, 13.27). By using PLS regression with the SNV pre-processing method, the best estimation results were obtained.

In order to model the potential non-linear relationship between the spectral reflectance and foliar chemistry, a BP-ANN was established with the same inputs as the PLS regression with the best pre-processing methods (SNV). The result (Table 4) shows no significant improvement when using ANN regression, besides the computational expense is much higher using ANN regression.

Figure 4 shows the one-one plot between the predicted and measured biochemical concentrations using PLSR and ANN regression. The PLS regression obtains better results in terms of R_cv and RMSE_CV.