Rock Detection and Arability Mapping Using Airborne LiDAR-RGB Fusion and Machine Learning

Agricultural businesses manage 51% of Australia’s land mass, yet sub-metre arability mapping at farm scale remains largely manual. This study presents an automated LiDAR–RGB fusion framework to detect rock outcrops, the dominant non-arable constraint, across a study farm in New South Wales, Australia. Three detection approaches were implemented at 10 cm spatial resolution: (i) a heuristic HSV-LiDAR prominence method (F1 = 0.493); (ii) an eight-feature Random Forest classifier (F1 = 0.704); and (iii) deep learning with a six-band LiDAR-Optical composite. Among three deep learning architectures evaluated, DeepLabV3+ (F1 = 0.721), Mask R-CNN instance segmentation (F1 = 0.834) and U-Net semantic segmentation (F1 = 0.756; F1 = 0.828 after targeted fine-tuning), Mask R-CNN achieved the highest F1 score but consistently under-mapped rock area (2.66 ha vs. 3.25 ha for U-Net and 3.14 ha for RF), reflecting conservative per-polygon confidence gating inherent in instance segmentation. U-Net is recommended for operational arability mapping due to its balanced precision–recall profile and area estimate consistent with Random Forest. Rock outcrops, tree cover, and slope were integrated into standardised arability layers at 0.5 m and 5 m resolution, enabling reliable paddock-scale farm planning. This study demonstrates that airborne LiDAR–optical fusion at 10 cm spatial resolution enables reliable automated rock outcrop detection and farm-scale arability mapping in Australian rocky landscapes. This sub-metre capability enables precision cultivation planning and improved land utilisation at resolutions previously unachievable through conventional survey methods.