Feasibility study on fruit parameter estimation based on hyperspectral LiDAR point cloud

Fruit parameter estimation is beneficial to intelligent management of orchards. This study proposed a compound method to estimate the fruit ripeness, number and size based on hyperspectral three-dimensional (3D) point cloud data collected by 101-channel hyperspectral LiDAR (HSL). Firstly, the spectrum-space dual classification method was employed to classify the components of fruit trees, including leaf, trunk, ripe and unripe fruit. The spectral parameters were used for random forest (RF) classification at first, and then combined with spatial information, points’ classes were reclassified by majority voting. Secondly, point cloud of fruits were clustered using the density-based spatial clustering of applications with noise (DBSCAN), which were separated for further fruit number counting. Finally, a novel fruit size estimate method, namely plane projection fitting method was proposed, which included fruit points projection, outline extraction, and shape fitting. The first two principal components were extracted by principal component analysis (PCA) to construct a projected plane. The outline of the projected results was obtained using convex hull algorithm, which was fitted with the least square method (LS) to estimate the fruit length and equatorial diameter. The results showed that the average classification accuracy of ripe and unripe fruits was 94.76% and 77.92%, respectively, and fruit counting accuracy was 96.30%. The estimated values were in an acceptable range compared to the manual measurements (root mean squared error for length and equatorial diameter were 9.55 mm and 4.12 mm, respectively). These promising results demonstrate the potential of HSL-based fruit parameters estimation in orchard environment.