Separating foliage and woody components can effectively improve the accuracy of simulating the forest eco-hydrological processes. It is still challenging to use deep learning models to classify canopy components from the point cloud data collected in forests by terrestrial laser scanning (TLS). In this study, we developed a convolution neural network (CNN)-based model to separate foliage and woody components (FWCNN) by combing the geometrical and laser return intensity (LRI) information of local point sets in TLS datasets. Meanwhile, we corrected the LRI information and proposed a contribution score evaluation method to objectively determine hyper-parameters (learning rate, batch size, and validation split rate) in the FWCNN model. Our results show that: (1) Correcting the LRI information could improve the overall classification accuracy (OA) of foliage and woody points in tested broadleaf (from 95.05% to 96.20%) and coniferous (from 93.46% to 94.98%) TLS datasets (Kappa ≥ 0.86). (2) Optimizing hyper-parameters was essential to enhance the running efficiency of the FWCNN model, and the determined hyper-parameter set was suitable to classify all tested TLS data. (3) The FWCNN model has great potential to classify TLS data in mixed forests with OA > 84.26% (Kappa ≥ 0.67). This work provides a foundation for retrieving the structural features of woody materials within the forest canopy.

Laser return intensity (LRI) information obtained from airborne laser scanning (ALS) data has been used to classify land cover types and to reveal canopy physiological features. However, the sensor-related and environmental parameters may introduce noise. In this study, we developed a local median filtering (LMF) method to point-by-point correct the LRI information. For each point, we deduced the reference variation range for its LRI. Then, we replaced the outliers of LRI with their local median values. To evaluate the LMF method, we assessed the discrepancy of LRI information from the same and diverse land cover types. Moreover, we used the corrected LRI to distinguish points from grass, road, and bare land, which were classified as ground type in ALS data. The results show that using the LMF method could increase the similarity of pointwise LRI from the same land cover type and the discrepancy of those from different kinds of targets. Using the LMF-corrected LRI could improve the overall classification accuracy of three land cover types by about 3% (all over 81%, κ ≥ 0.73, p < 0.05), compared to those using the original and range-normalized LRI. The sensor-related metrics brought more noise to the original LRI information than the environmental factors. Using the LMF method could effectively correct LRI information from historical ALS datasets.