Figure 1 is the block diagram of the hardware system for the handheld chlorophyll content detector of wheat and maize leaves. The detector comprises a microcontroller, light source, RGB sensor, displayer, Bluetooth module, and power management module. The microcontroller undertakes the task of managing each module. The light source is employed to provide stable light, while the RGB sensor collects the RGB color information of leaves and forwards the data to the microcontroller. The displayer exhibits the detection results. The Bluetooth module enables data transmission to a mobile phone, and the power management module ensures power provision for the entire detector, managing charging functions and offering protection. Lastly, the Type-C interface facilitates serial communication for the microcontroller, and the power charging interface is also used.

The detector's software, developed in Keil μVision5, consists of a main function and several sub-functions, such as the leaf color collection sub-function, data processing sub-function, key sub-function, and display sub-function. The software's running process is shown in Figure 6 . After the detector is turned on, each module is initialized, and the OLED displays the welcome interface. After 1 s, the message about measurement is displayed until the detection head is pressed. Then, the RGB sensor collects the color information of the plant leaf and the data is transmitted to the microcontroller for calculation and analysis. If the detection value is abnormal, a detection error is displayed on the OLED screen. Otherwise, the detection results are displayed on the screen. If the transmission button is pressed, the detection results are sent to a mobile phone, and the OLED indicates that the transmission is successful.

The wheat and maize leaf samples used in this study were obtained from different experimental fields at Northwest A&F University, China. The wheat leaves for modeling were collected from 12 wheat cultivars at different experimental plots from May 6 to 10, 2023. There were 71 leaf samples, including 5–6 samples of each cultivar. The maize leaves used for modeling were collected from July 27 to August 1, 2023, and the samples were obtained from four cultivars in different experimental fields, with 15–16 samples of each variety and a total of 61 samples. The samples used in detector performance evaluation were evenly collected from the same experimental fields. The wheat leaves were collected from May 15 to 18, 2023, and the maize leaves were collected from August 5 to 11, 2023. The total samples of wheat and maize were 34 and 31, respectively.

Table 1 presents the statistics of the obtained SPAD values, and the CL-a, CL-b, and TCL contents of 71 wheat and 61 maize leaf samples in modelling and 34 wheat and 31 maize leaf samples in the detector's performance validation.

It shows that the ranges of TCL contents of wheat and maize leaves used in modeling were 1.340-4.415 mg/g and 1.342-3.161 mg/g, respectively, and were 1.052-4.859 mg/g and 1.329-3.220 mg/g in validation. The coefficient of variation (CV) of the investigated chlorophyll contents of the used leaves in modeling changed from 22.20% to 32.12% and from 22.07 to 39.49% in validation. Generally, the wheat leaves had a wider change range in chlorophyll content than maize leaves. The large CV indicates that the samples used had good representation.

The 20 color features obtained in Section 3.2.1 were applied as inputs to regress the relationship with CL-a, CL-b, and TCL contents using the PLSR and MLR. The modeling results are listed in Table 2 . It shows that the MLR model had higher R ² and smaller RMSE than PLSR for each chlorophyll content. Therefore, MLR was used to predict chlorophyll content further.

The significance analysis of the built MLR models for predicting CL-a, CL-b, and TCL contents shows that B and B′ had very significant relationships (p< 0.01) with each chlorophyll content, H was correlated with CL-a (p< 0.05), S and V with CL-b (p< 0.05), and H and S with TCL (p< 0.05) significantly, shown in Table 3 . The relationships of other color features with these chlorophyll contents weren't significant (p > 0.05). To eliminate redundant parameters, reduce the complexity of the models, and avoid overfitting, the MLR models for predicting CL-a ( W a ), CL-b ( W b ), and TCL content ( W t ) were built using the 5 color features, that is, B, B′, H, S, and V. The models are shown in Equations 5-7. The R ² of the built models for CL-a, CL-b, and TCL contents were 0.796, 0.795, and 0.801, respectively, and the RMSEs were 0.248 mg/g, 0.080 mg/g, and 0.320 mg/g. Although the modeling performances are a little poorer than those of the built models using the 20 color features ( Table 2 ), the models are much simpler.

The repeatability of the developed detector in obtaining color features was evaluated by randomly selecting three wheat leaves and three maize leaves with different chlorophyll contents. The R′, G′, B′, and C of each leaf were measured 10 times on the same point. The mean, standard deviation, and CV of each color index for each sample are listed in Table 4 . It shows that the CV ranges from 0.000% to 2.083% with a mean of 0.516%, indicating that the detector has good repeatability in obtaining color features.

As chlorophyll content may exhibit slight diurnal variation even within the same leaf and pot, the stability of the developed detector in capturing color features was evaluated using several color cards of different hues. Measurements were conducted in a maize field on a sunny day at 8:00, 12:00, 16:00, and 20:00. The data are listed in Table 5 . The CV of the mean R′, G′, B′, and C values at different times were 0.911%, 1.103%, 0.941%, and 0.771%, respectively, indicating that the detector has good stability.

After the software with the MLR model for predicting chlorophyll contents was downloaded to the explored detector, the precision of the explored detector was evaluated using the 34 wheat and 31 maize leaves in the validation set ( Table 1 ). It is noted that the used samples had a wide range of chlorophyll contents. After the chlorophyll content of each leaf was measured using the detector, the actual chlorophyll contents of each leaf were measured according to the method given in Section 3.2.2.

Figure 7 shows the detected CL-a, CL-b, and TCL contents using the explored detector against the measured values using spectrophotometry. The symmetrical distribution of points around the 45° line suggests that the detector demonstrates high precision. The RMSEs of the explored detector were 0.269 mg/g for CL-a, 0.089 mg/g for CL-b, and 0.350 mg/g for TCL contents of wheat and maize leaves, respectively, and had the mean absolute errors of 0.104 mg/g, 0.031 mg/g, and 0.087 mg/g.

In order to evaluate the predictive performance across crop species, separate analyses were conducted for wheat and maize leaves. The RMSE values for CL-a, CL-b, and TCL in wheat were 0.264 mg/g, 0.090 mg/g, and 0.346 mg/g, respectively, while those for maize were 0.271 mg/g, 0.083 mg/g, and 0.347 mg/g. These results demonstrate that the predictive performance was generally consistent between the two crops. This similarity may lie in their shared phylogenetic traits as Poaceae plants, which exhibit analogous chlorophyll spectral absorption characteristics. Additionally, their parallel leaf venation and similar epidermal structures likely lead to nearly identical light scattering and absorption pathways within leaves.

Li et al. (2021) used an auxiliary device to assist smartphones in shooting RGB images of sugarcane leaves to detect their chlorophyll content. They established a support vector machine regression model to predict the chlorophyll content with an average RMSE of 0.3288 mg/g, a little lower than the obtained 0.350 mg/g for chlorophyll content here. However, the image color is affected by many factors, such as light, camera used in different smartphones, shooting angles, etc., restricting the application of this method. Moreover, the contents of CL-a and CL-b cannot be given.

The TCL content detector developed by Yang et al. (2019) had an absolute measurement error of -0.32-0.20 mg/g and a mean absolute error of 0.14 mg/g for spinach, big green vegetables, and lettuce leaves. However, only TCL can be measured. In contrast with their works, the obtained RMSEs here are higher. The reason might lie in several aspects. One reason is that the wheat leaf is very narrow, which causes the obtained color features to contain the information of leaf veins and make the color features, especially B′ (which has a significant correlation with each chlorophyll content), to have a wide variation, lowering the precision. Another is that the wheat and maize leaves had much higher chlorophyll content and wider change ranges than the leafy vegetables they used. The third is that the samples used herein validation had a wider change range in chlorophyll content than those used in modeling, also decreasing the precision. In addition, although the cheap RGB sensor (about 2 US dollars) decreases the detector's cost, it also reduces the precision. Using high-performance RGB sensors would help improve the precision determination of CL-a, CL-b, and TCL contents.

To evaluate whether the color indices are more efficient in determining chlorophyll content than SPAD values, the linear relationship between SPAD values and the chlorophyll content was analyzed. The RMSE of predicting TCL content in wheat and maize leaves using SPAD values was 0.428 mg/g. The results indicate that the device developed in this study exhibits better performance in predicting chlorophyll content in wheat and maize leaves than the SPAD-502 chlorophyll detector. The developed detector in this study has a mass of only 88 g. It can give the test results within 2 s without destruction to crop leaves, meeting the requirements of real-time and non-destructive detection of chlorophyll contents of crop leaves. However, since the current model is specifically established for wheat and maize, further research is required to evaluate its applicability to other crop species.