1. Introduction
Potato, having the high nutritional value and yield, is the third largest staple food in the world [
1,
2]. Chlorophyll, which can capture and transform light energy into chemical energy for the conversion of inorganic matter to organic matter, is an important raw material in potato photosynthesis [
3]. In addition, the chlorophyll content is significantly correlated with the content of nitrogen [
4], which is one of the most important nutrients for the potato plant’s healthy growth [
5], development [
6], and production [
7]. Therefore, the accurate detection of the chlorophyll content of potato plants is significant for precision agriculture management. In recent years, many destructive and non-destructive methods have been developed to estimate the chlorophyll content of leaves. Destructive methods for detecting chlorophyll content through leaf sampling and laboratory chemical analysis [
8]—for example, the spectrophotometer method—are very accurate but time consuming and laborious [
9]. These methods cannot rapidly estimate the chlorophyll level of a crop and cannot meet the development of precision agriculture. As a non-destructive and rapid monitoring method, modern spectral analysis technology has shown outstanding advantages for collecting data on a crop’s physical or chemical components, such as the chlorophyll content of plants.
The soil plant analyzer development (SPAD) meter is a rapid, non-destructive spectral device, which is widely employed to measure in situ the chlorophyll content of crops in the field [
10,
11]. The most commonly used SPAD chlorophyll meter measures transmittance of light through the leaf at 650 and 940 nm to estimate leaf greenness [
12]. There are many reports in the literature that the chlorophyll content is high relevant to the SPAD value at leaf scale. For instance, the square of the correlation coefficient (
R2) between SPAD value and the chlorophyll content measured using solvent extraction method in wheat [
13], rice [
14] and Arabidopsis Thaliana [
15] leaves was 0.92, 0.93, 0.93, and 0.98. In addition, the
R2 of 0.95 was found between the chlorophyll content of potato leaves measured in laboratory and SPAD readings [
16]. However, the SPAD can only detect the chlorophyll content at a point on of single leaf with low detection efficiency [
17,
18], it is limited to rapidly estimate the chlorophyll status of whole canopy of crops. Moreover, the SPAD values were different in different measurement spots [
19], which is not conducive to the practice of precision agriculture. To measure the spectral data of a canopy area or whole potato plant without any contact or damage, some researchers developed spectral imaging sensors [
20,
21,
22]. For instance, Borhan and Panigrah [
21] developed a CCD sensor to collect indoor images at 550 and 700 nm for detecting the SPAD value of potato. And the determination coefficient of calibration set was 0.80. However, the abovementioned study only collected spectral data in the light region, which cannot comprehensively reflect the leaf information to accurately detect the SAPD value. Some studies indicate that the information of some bands located in the near-infrared region is related to SAPD value, such as normalized difference red edge (NDRE
710, 750 nm) index [
23], the MERIS terrestrial chlorophyll index (MTCI
681,708, 751 nm), and red edge positions ranging 702 to 706 nm [
24]. On one hand, these results demonstrated the potential to present the SPAD value by high-throughput imaging sensor. On the other hand, they explored the feasibility to estimate the SPAD transmittance of light through the leaf at 650 and 940 nm by the spectral reflectance of canopy within 681–750 nm.
Based on the above problems, in order to obtain the SPAD value and distribution by high-throughput near the ground, we intended to develop a detection system based on a 25-wavelength spectral sensor. The detection system can collect the spectral images of potato plants; process online the spectral data, including image segmentation, reflectance extraction, SPAD value calculation; and draw the visualization distribution map of SPAD value of potato plants in real time. For data processing, the core work in the study consisted of the segmentation of potato plant images and the establishment of a SPAD value detection model. Herein, the segmentation is referred to the process of extracting the target potato plant pixels from spectral images. In the field, the spectral data of potato are disturbed by various noises, such as canopy geometry, soil, weeds, and residual straw, which greatly affect the accuracy of SPAD value detection. Therefore, the accurate segmentation of potato plants from the image is crucial to ensure data validity and establish a high-precision detection model [
25]. The Otsu algorithm (OTSU) is widely used for image segmentation, which can automatically obtain the segmentation threshold [
26]. However, the OTSU cannot effectively complete the segmentation task if the grayscale values of target object are similar to that of background noises [
27]. Combining OTSU with spectral property, many spectral index segmentation methods such as the normalized difference index (NDI-OTSU) [
28], excess red index (ExR-OTSU) [
29], vegetative index (VEG-OTSU) [
30], combined indices (COM-OTSU) [
31] have been proposed. These spectral indices have been able to reduce the background information to a certain extent to improve the segmentation performance. However, the segmentation accuracy, which ranged from 53% to 88% for the above methods, was still low [
32]. Some learning-based approach segmentation methods have also been proposed, such as a learning approach based on decision trees [
33] or on support vector machines [
34], the supervised mean-shift method with back propagation neural network [
35], or with a Fisher linear discriminant [
36] algorithm with high segmentation accuracy for the calibration data. However, these methods rely on a good deal of data to train the segmentation model, and suffer from a long run time, lower stability, and problems with overfitting. Facing complex background noise in the field environment, the modified difference vegetation index (MDVI) along with OTSU and combined with connected domain-labeling (CDL) segmentation method (MDVI–OTSU–CDL) is proposed in this study to extract spectral data of potato plants online.
The partial least squares (PLS) algorithm is widely used to establish a detection model to describe the relationship between the agronomic parameters and spectral reflectance [
37]. The PLS algorithm can reduce the variable collinearity by principal component analysis, which can enhance the model stability [
38]. Occasionally, the PLS regression possesses poor prediction accuracy because the spectral variables contain irrelevant information. Some studies have reported that the uninformative variable elimination (UVE), serving as a sensitive wavelength selection algorithm, can improve the performance of the PLS regression model by eliminating invalid variables [
39,
40]. We attempt herein to establish a high-performance SPAD value model using the UVE–PLS method.
Therefore, this study aims to develop a detection system based on a novel spectral imaging sensor for the real-time detection of the SPAD value of potato plants in the field. The main data processing steps of the developed sensors system are shown in
Figure 1. The system has the following functions:
- (1)
Automatically segment potato plants using the MDVI–OTSU method,
- (2)
Automatically extract the reflectance of potato plants using segmented mask images,
- (3)
Detect the SPAD value of potato plants in real time using the established model,
- (4)
Generate a visualization distribution map of SPAD value of potato plant based on the spectral images and PLS model.
4. Conclusions
This study developed a SPAD value real-time detection system based on a spectral imaging sensor. The detection system had two core functions, which were potato plant segmentation and SPAD value detection. To accurately extract the potato plant from the spectral images, the MDVI–OTSU–CDL segmentation method was proposed, and the six modified coefficients were discussed. The results showed that the segmentation accuracy was best when the modified coefficient was 2.5 (), so the MDVI (2.5)–OTSU–CDL was applied to segment the potato plant. Then, the reflectance at 25 wavelengths was extracted by the segmented mask images. The UVE–PLS model was established to detect accurately the SPAD value of potato plants, and the , RMSEV, and RPD of the UVE–PLS were 0.850, 3.31, 2.461, respectively, which resulted from the sensitive variables containing the work wavelengths of SPAD meter. The pseudo-colored map could visualize the distribution of SPAD value. Finally, the performance of the developed detection system was measured using the testing dataset; the R2, RMSE, and RPD were 0.776, 2.478, and 1.891, respectively, which demonstrated that the developed detection system possessed good stability and excellent applicability.
There were seven sensitive wavelengths selected by UVE, the number of which was more than the SPAD meter (two wavelengths) but less than the PLS model. The model results showed that the UVE–PLS possessed better prediction ability, which resulted from the UVE having eliminated the uninformative variables. The differences of sensitive wavelengths may be due to that the developed sensor system collects the reflectance data and the SPAD meter collects the transmittance data. SPAD is a non-destructive device to measure the chlorophyll content and its value has a strong correlation with the leaf chlorophyll of the potato plants. In the paper, to continuously detect chlorophyll level distribution of whole potato plant in different growth stages, the SPAD device instead of the laboratory extraction method was employed to measure chlorophyll level of potato plants for establishing the detection model. The distribution of SPAD can be obtained using the sensor system, but the relationship between SPAD value and chlorophyll content need to be calibrated, and then the results can represent chlorophyll content.
Overall, the new proposed system is helpful for detection chlorophyll level of potato plants. However, the method in this study is based on specific spectral data for potato crops. The restrictions are based on the existence of other datasets or potato varieties. Therefore, more datasets from a wide range of potato varieties, planting patterns, and experimental fields should be collected and analyzed to improve the detection performance of the developed sensor system. In addition, this proposed system can be employed to detect the chlorophyll content of others crops after improving the segmentation method and the detection model and adjusting the imaging setups.