Deep learning-based canopy gap detection using a cross-technological approach with airborne laser scanning and aerial imagery data

Canopy gaps are crucial structural elements of forests, supporting biodiversity and influencing forest dynamics and ecosystem health. Airborne laser scanning (ALS) is commonly used for forest gap analysis and typically outperforms digital aerial photogrammetry (DAP), especially in detecting smaller gaps. However, ALS data availability remains limited compared to DAP. Given the broader availability and cost-effectiveness of DAP, this study aimed to overcome its technical drawbacks in canopy gap detection by applying a cross-technological approach with multiple data sources. This involves ALS-derived reference data fused with spectral and height information from DAP. We developed a deep learning-based method, employing a convolutional neural network (CNN), specifically the U-Net architecture, for detecting canopy gaps. The U-Net was trained using gap polygons automatically generated from ALS-derived canopy height models (CHMs), combined with true digital orthophotos (TDOPs) and DAP-based CHMs. Adding spectral information from TDOPs was intended to help detect shadows typically associated with smaller canopy gaps, which are often missed in DAP-based CHMs. The model was tested in the Solling, a forest area in a low mountain range in Central Germany. Performance was evaluated in independent test areas representing a gradient of structural heterogeneity. Overall, our model achieved moderate to high segmentation performance (IoU: 0.67–0.77; F1-score: 0.56–0.74). Once trained, it can be applied to image-derived inputs, improving canopy gap detection F1-score by on average 0.08 compared to using DAP-based CHMs alone. Our results demonstrate a novel approach for detecting canopy gaps without ALS data, suggesting applications across broader spatial and temporal scales. • U-Net using spectral data and DAP-based heights for canopy gap detection. • ALS as an objective data source to automatically map canopy gaps. • Forest structure impacts segmentation results more than DAP-CHM quality. • Approach enables gap mapping in absence of ALS, using image-derived inputs only.