Postharvest moisture-level classification using LiDAR point clouds: Integrating 3D structure and spectral intensity

Moisture content plays a decisive role in the postharvest quality and preservation of horticultural produce. Most non-invasive sensing methods capture either spectral or spatial features, failing to account for structural heterogeneity in fresh tissue and introduced during moisture loss. In this work, a dual wavelengths 3D LiDAR sensor was employed for moisture classification in broccoli and mushrooms. The dataset is comprising spatial coordinates and reflectance intensities at 1320 nm, 1450 nm, and a combined moisture index. Samples include paired ground-truth labels for core and periphery regions, with moisture content categorized into high, medium, and low levels. Results show that by using only reflectance intensity features, the baseline models achieve moderate moisture-level classification performance, with accuracies not exceeding 0.82. Combining geometry with reflectance, particularly at 1450 nm and moisture index, PointNet++ yields the most accurate predictions, achieving up to 0.93 accuracy. Geometric features alone offer limited class discriminability, but contribute to prediction robustness across structurally diverse samples. Spatial generalization experiments reveal an asymmetry: models trained on dense, homogeneous core segments generalize effectively to peripheral regions, whereas those trained on periphery segments exhibit reduced accuracy when evaluated on core data. Modality perturbation shows that intensity provides the dominant moisture-sensitive signal, particularly at 1450 nm, while geometry contributes complementary cues that stabilize predictions. At 1320 nm, showing low spectral sensitivity to moisture, reliable classification requires joint use of both modalities. These findings establish LiDAR as an efficient non-invasive moisture monitoring tool in heterogeneous agricultural products. Code: github.com/majharB/vegPoint .