A Vis/NIRS device for evaluating leaf nitrogen content using K-means algorithm and feature extraction methods

Abstract

Accurate assessing leaf nitrogen content (LNC) is crucial for actual production and fertilizer management. In this research, a portable device was designed to rapidly and non-destructively evaluate LNC with precision. Using hydroponically grown eggplants exposed to different nitrogen content nutrient solutions as experimental samples, we conducted various measurements, including chlorophyll fluorescence (ChlF) induction curves, hyperspectral images, and LNC values. Correlations between LNC and ChlF parameters were calculated, and the parameter qN obtained the highest correlation with LNC. False color images of qN were segmented using the K-Means algorithm to obtain three regions. The spectral data and the measured LNC of the corresponding region in the leaf were matched, and a LNC prediction model was developed using the partial least square regression (PLSR) algorithm with the processed spectral data as input and the measured LNC as output. The results showed that the model using standard normal variate-iteratively retains informative variables- successive projections algorithm (SNV-IRIV-SPA-PLSR) yielded the best performance, with a correlation coefficient of prediction (R2) of 0.9332, a root mean square error (RMSE) of 2.6890 mg/g, a residual prediction deviation (RPD) of 3.97 and a ratio of performance to interquartile distance (RPIQ) of 7.28. Based on the selected wavelengths from the SNV-IRIV-SPA-PLSR-VIP model, six narrow-band light emitting diodes (LEDs) were chosen as the light source for the designed device. Inexpensive modules were employed to assemble the device, and accuracy tests were conducted. The PLSR algorithm was employed to develop the device’s LNC evaluation model with the reflectance of the leaf under 6 LEDs as input (resulting in R2, RMSE, RPD, and RPIQ values of 0.8075 6.6242 mg/g, 2.30 and 4.26, respectively). The model was then embedded in the core processor. To validate the device’s performance, an independent set was used, resulting in R2 of 0.7559, RMSE of 7.4771 mg/g, RPD of 2.07, and RPIQ of 3.57, respectively. The proposed device could rapidly and accurately determine LNC in plants, surpassing other devices in terms of portability and cost. This research offers a potential solution for plant fertilizer management.