L-3DCNN, lightweight 3D-CNN for spectral-spatial–elevation fusion: a reproducible baseline for urban land-cover classification

This paper presents a simple yet effective three-dimensional convolutional neural network (L-3DCNN) for urban landcover classification using co-registered hyperspectral imagery (HSI) and LiDAR data. The proposed network treats the full HSI–LiDAR data cube as a three-dimensional volume and employs a series of 3×3×3 convolutions to capture spectral and spatial context simultaneously. LiDAR-derived elevation is incorporated as an additional channel alongside the spectral bands. During preprocessing, each hyperspectral band and the LiDAR digital surface model (DSM) are normalized to zero mean and unit variance. Model training is patch-based: small local patches (P×P) centered on labeled pixels are extracted, and random flips and 90° rotations (dihedral transformations) are applied uniformly across all channels for data augmentation. To address class imbalance, the cross-entropy loss is weighted by the inverse class frequencies and augmented with label smoothing. experiments on Houston 2013 show that the L-3DCNN attains high, stable accuracy. The results confirm that combining spectral information with LiDAR-based structural cues enhances discrimination among urban land-cover classes.