The target tea stresses of this study (i.e., the tea leafhoppers, anthrax and sunburns) often occur in tea gardens in Zhejiang province. The three stresses have relatively similar symptoms such as the grayish-brown lesions on leaves. However, given that the prevention of the different stresses require different procedures, it is important to differentiate them for precise management. This experiment was carried out in the experimental tea field of the Chinese Academy of Agricultural Sciences (Zhejiang) in June 2019. The disease, pest and sunburn stresses occurred naturally in the studied field. Based on a field survey, three locations corresponding to the three stresses were picked by an experienced plant pathologist to conduct in-situ canopy hyperspectral imaging measurements.

The hyperspectral images were obtained using an UHD185 imaging spectrometer developed by Cubert Company in Germany (http://cubert-gmbh.com/). The imaging spectrometer is a full-frame, snap-shot devise, which have 126 bands within 450–950 nm. The spectrometer has a spectral sampling interval of 4 nm, and an image resolution of 1000 × 1000 pixels. The spectral measurements were made between 10:00 AM and 2:00 PM. The spectral images were taken at a distance of 50 cm above from the top of the tea canopy. The image radiometric correction was performed to convert the reflectance using the reference whiteboard and blackboard. For each type of stress, 40–50 spectral images were taken. After quality control, 30 hyperspectral images were used for subsequent analysis for each stress (Fig. 1).

In this study, all hyperspectral images were randomly divided into two parts: 50% of data were used for training, and the other 50% for verification. According to the confusion matrix, the overall Accuracy (OA) and kappa coefficient were calculated as accuracy indicators. In this study, hyperspectral image data processing was completed in ENVI 5.3 (Exelis Visual Information Solutions, USA), and data analysis and modeling was completed in Matlab 2014 (MathWorks, USA).

By observing and comparing the spectral curves that were extracted from the ROIs of GL, AH, BR stressed hyperspectral images, it was found that the spectral characteristic follow the same pattern across different types of ROIs. In all types of stress images, the overexposed region (OR) and the background region (BR) have the highest and lowest spectral reflectance in most bands, respectively. The reflectance of highly illuminated leaf region (HLR) and shadow leaf region (SLR) for the healthy leaf area (HR) were slightly higher and lower, respectively, than those of the normal-level illuminated leaf region (NLR) (Fig. 3).

In comparing the spectral reflectance between healthy and damaged leaf regions, it was found that the reflectance significantly increased in the red band (620–660 nm) and decreased in the near infrared platform (780–900 nm) for the damaged regions, and the shape of the entire spectral curves tend to be relatively flat. Such spectral changes might be associated with the damage to the pigment system and destruction of cellular structure [30]. Among the three stresses, GL and AH showed similar spectral curves, whereas the spectral curve of BU deviated from that of the healthy leaf regions at a larger extent, which indicates the different extent between biological and physical destruction. As a physical damage, BU usually causes relatively wide and severe damage to plant tissues, while the damage caused by disease infection and pest infestation usually presented a gradual damage process with a relatively limited extent. The spectral difference among the different ROIs forms an important basis for subsequent detection of canopy damaged areas and discrimination of stresses.

For separating tea plants from the background in the canopy hyperspectral images, a total of seven spectral bands were selected through the SPA algorithm, including: 460 nm, 550 nm, 670 nm, 706 nm, 742 nm, 893 nm and 920 nm. These bands were distributed at some important positions in the spectral curve of plants (Fig. 4). For example, 460 nm, 550 nm and 670 nm were located around the blue edge, green peak, and red valley regions. These locations are sensitive to changes of some important biochemical components such as the content of chlorophyll, which may have a significant difference between healthy and stressed plant regions. The 706 nm and 742 nm located around the red-edge region, which are sensitive to the stress status of plants. As the SPA algorithm adopted the combinatorial optimization strategy, the optimized bands have a relatively low level of correlation among each other, which is carried out in the background removal process.

In selecting features for detecting damaged areas, the t-test analysis and the cross-correlation analysis were applied on the three types of candidate spectral features (i.e., spectral derivative features, continuum removal features and vegetation indexes). This yielded a total of 5 spectral features: YSV, Area, GI, CARI and NBNDVI. Among them, the spectral derivative feature YSV captures the degree of spectral variation within 540–620 nm, while the continuum removal feature Area indicates the intensity of spectral absorption within 530–770 nm. These two spectral shape-based features mainly reflect the variation of pigment contents. The other three vegetation indexes indicated the greenness and the chlorophyll content of plants. Given that the content of these pigments will significantly decrease when plants are under stress, the identified spectral features can (1) Serve as indicators of the health status of plants, and (2) Be used for detecting the damaged areas in the hyperspectral images (Tab. 2).

In addition, the same strategy was applied on the selection of spectral features for discriminating different stresses, which yielded 6 optimal spectral features: MCARI, CI, LCI, RARS, TCI and VOG. These features are all vegetation indexes that associated with the contents of chlorophyll and carotenoid (Tab. 2). Given that different stresses can cause subtle differences of damage to the pigment system in plant leaves, the spectral responses are able to be captured by the features and used for discriminating these stresses.

Firstly, and based on the optimal bands generated by the SPA algorithm, the non-plant background needs to be removed from the tea canopy hyperspectral images. The K-means algorithm was applied on the images of the selected bands to conduct the clustering on the image pixels. The pixels belonging to each cluster were then projected to the two-dimensional feature space that consisted in the reflectance of the RED (630–690 nm) and near-infrared (NIR; 760–900 nm) bands. It was observed that the background area was located in the part with relatively low RED and NIR values, whereas the overexposed area was located in the part with relatively high RED and NIR values. Through an iterative-optimized thresholding approach, the background and overexposed area were removed so that the subsequent analysis was focused on the plant area in the images (Fig. 5). Comparing with the traditional pixel-based classification methods, the clustering-based method effectively mitigated the edge irregularities and the “salt and pepper” phenomenon in the classification results [8].

Within the plant area of the image, based on the selected optimal spectral features that were suitable for damage detection, both KNN and RF were applied for constructing the model for detecting the damaged areas on tea plants (Tab. 3). In this process, unlike the usual pixel-based analysis, the modeling process also took the K-means clustering results as basic units. The detecting models were validated with the validation image data, and the results showed that the KNN outperformed the RF model in general. The KNN achieved relatively high accuracy for all three stresses, with an OA over 95%. While for RF, the performance differed significantly among the stresses, with a relatively high accuracy for GL (OA = 98%), and moderate accuracies for AH (OA = 86%) and BU (OA = 83%). The difference of the model performance can also be observed from a couple of examples as shown in Fig. 6. It is noted that the results of the RF models included several over-classified regions, which was the main reason for their relatively low accuracy. The significant spectral difference between the stressed and healthy samples may provide a solid basis for the classification using the KNN model, which adopted the straightforward feature distance criteria. The simple structure and low level of computational complexity makes it a promising solution in practice.

For stress discrimination, after comparing with the visual recognition results, the Fisher discriminant model achieved relatively high accuracies for all the three types of stresses. In them, the discrimination accuracy reached 98.53% and 98.29% for GL and AH, respectively. The accuracy was slightly lower for BU, with an OA of 93.94%. Despite the three stresses were similar on symptoms, it is encouraging that the subtle differences were effectively captured by the spectral signals (Fig. 7). When comparing the present study with some studies that were conducted at a leaf level, the complicated background scenarios at the canopy level posed a big challenge in detecting and differentiating the diseases and pest. However, by adopting the conception of stepwise detection, the final discriminant analysis can be carried out in a relatively pure scenario (i.e., the damaged area), which is beneficial for improving the performance and stability of the model. It is of great importance to incorporate this canopy-level tea stress detecting and discriminating method in some autonomous driving platforms such as Unmanned Aerial Vehicle (UAV) that are equipped with hyperspectral or low-cost customized-bands cameras. More studies and experiments need to be carried out to verify the effectiveness of corresponding methods in large-scale detection of stresses in tea plantations.