In situ nondestructive identification of citrus fruit ripeness via hyperspectral imaging technology

Abstract

Rapid and accurate assessment of the citrus ripening stage in the field is important for determining harvest timing and improving industrial economic efficiency; however, the lack of effective nondestructive detection methods in the current orchard leads to flaws in ripening stage assessment, which affects harvesting decisions. To solve this problem, this study utilized hyperspectral technology to collect data from 22 fruit trees in an orchard (in the range of 400-1000 nm) and explored the effectiveness of five regions of interest selection methods (x-axis, y-axis, four-quadrant, threshold segmentation, and raw) for the delineation of the citrus ripening stage. The data quality was enhanced via wavelet transform (WT)-multiple scattering correction (MSC) preprocessing, and the effective wavelengths were extracted via the successive projections algorithm (SPA). On the basis of these wavelengths, backpropagation neural network (BP) and convolutional neural network (CNN) models were built for maturity prediction. The results show that the x-axis region of interest selection method outperforms the other methods, and the SPA-BP model based on this method performs best. An accuracy of 99.19% for the correction set and 100% for the prediction set was achieved when only 0.03% of the wavelength was used. This groundbreaking study highlights the significant potential of hyperspectral technology for in situ assessment of citrus ripening stages. Furthermore, it offers crucial technical support and serves as a valuable reference for the advancement of precision agriculture.

Keywords: Hyperspectral; Region of interest; Ripening stage; Successive projections algorithm; ‘Shiranui’ mandarin.

Fig. 1

Schematic diagram of the three stages of maturation of ‘Shiranui’ mandarin. (a) CLASS A (b) CLASS B (c) CLASS C

Fig. 2

Placement of the spectrometer and the whiteboard (a) Gaze-type visible near-infrared hyperspectral imager (b) Placement of the whiteboard

Fig. 3

Schematic selection of areas of interest for citrus (a) x-axis (b) y-axis (c) four-quadrant (d) threshold segmentation (e) raw

Fig. 4

X-axis raw spectral curve (a) Raw Spectral Curve (b) spectral curves of average ± S.D. of Unripe stage (CLASS A), ripe stage (CLASS B), and overripe stage (CLASS C) samples, respectively

Fig. 5

Combined WT and MSC preprocessed image and its mean image (a) Combined WT and MSC preprocessed images (b) Preprocessed images of mean values at each maturity stage

Fig. 6

SPA algorithm for dimensionality reduction of preprocessed spectral data (a) Curve of Root Mean Square Error as a function of the number of effective wavelengths extracted by SPA (b) Distribution of characteristic wavelengths selected by the SPA algorithm

Fig. 7

Confusion matrix and prediction result plots for correction and prediction sets based on x-axis region of interest selection (a) Calibration set prediction confusion matrix (b) Calibration set prediction results (c) Prediction Confusion Matrix for Prediction Sets (d) Prediction set prediction results