Tomato (Solanum lycopersicum L.) is a climacteric fruit exhibiting the ripening pattern with change in colour from green to red. Normalized difference vegetation index (NDVI) derived from spectral-optical analysis has been previously employed in ripeness analysis of various fruits based on its strong correlation to ripening-related chlorophyll content. In this study, light detection and ranging (LiDAR) laser scanner was applied to gain 3D tomato point clouds aimed at estimating the spatially resolved chlorophyll content and ripeness class of tomato fruit. Freshly harvested tomatoes, capturing six ripeness classes (mature green, breaker, turning, pink, light red, and red) according to USDA colour standard, were analysed using a linear conveyor mounted sensor system consisting of two LiDAR units measuring position and return signal strength intensity at 660 nm and 905 nm. Fruit point clouds were pre-processed including geometric correction considering the curvature of each tomato sample, removing highest intensity areas at the specular highlighted spots, calibration of intensity values using standard black and white colour coated boards, and calibration on attenuation of the opaque samples. Particularly, for gaining the attenuation coefficient (μ), 9 opaque synthetic models with known attenuation were used. Obtained μ660 and μ905 were merged to gain NDVILiDAR. Chemically analysed chlorophyll content of tomato samples was correlated to μ660 and NDVILiDAR, whereas low correlation appeared for μ905 with a coefficient of determination (R2) of 0.58, 0.60, and 0.06, respectively. Regression models were used to estimate the total tomato chlorophyll content and the lowest root mean square error (RMSE) was found for μ660 (RMSE = 4.97 mg (100 g dry mass)−1) followed by NDVILiDAR (RMSE = 5.22 mg (100 g dry mass)−1), and μ905 (RMSE = 53.50 mg (100 g dry mass)−1). Histograms of NDVILiDAR and μ660 were extracted from each tomato point cloud and utilized to construct PLS-DA model for tomato ripening class prediction considering the spatial chlorophyll distribution. The overall accuracies for NDVILiDAR and μ660 were 70 % and 68 %, respectively, in leave-one-out cross-validation for visually defined colour classes. The LiDAR-based approach could support selective detection of ripe fruit in robots for harvesting and postharvest handling.
Read full abstract