Label-Free Land-Cover Mapping From Airborne LiDAR via SAM Distillation With Convolutional LoRA Adapters
Year: 2026
Type: article
Abstract: Airborne laser scanning supports land-cover mapping with illumination invariance and strong geometric cues, but supervised learning is often constrained by scarce labels and a modality mismatch with natural images. We propose a label-free adaptation recipe that reuses a promptable segmentation foundation model on laser-derived raster composites through a tiny convolutional low-rank adapter while freezing the backbone. We construct three-channel inputs from a digital surface model, slope, and either intensity or local relief, and enforce a strict missing-data convention throughout preprocessing, prompting, training, and evaluation. Pseudo-supervision for adapter distillation is generated from NoData-aware grid prompting on the frozen model; the adapter is distilled from the resulting union masks using a Dice-style overlap loss, cross-entropy, and total-variation regularization, requiring no human annotations. During inference, missing-data–aware prompting and overlap-add feathering produce seam-free mosaics, and unsupervised instance prototypes combined with simple height-based priors map object masks into three land-cover classes (ground, vegetation, building). On the held-out annotated AOIs with polygon reference labels and differing environmental characteristics, the pipeline achieves a mean Macro IoU of 0.568, a mean Micro IoU of 0.687, an overall accuracy of 0.810, and a mean Cohen’skof 0.599. In addition, we report preprocessing and data-coverage statistics for the broader 30-AOI corpus to make the diversity of the study setting explicit. The paper is positioned as an applicability and engineering study that documents a transparent, parameter-efficient pathway for transferring promptable foundation models to airborne laser-scanning composites without full fine-tuning.
Source: IEEE Access
Institution: Innopolis University
Language: English
Subfield: Forestry
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7160169268