Jinhu Wang

Recent publications

1. Cross-platform forest understanding: A multi-platform synergistic training framework for generalized forest point cloud segmentation

   Jundi Jiang, Yueqian Shen, Jinhu Wang, W. Daniel Kissling, Markus Hollaus, Hongjun Su, Jinguo Wang, Vagner Ferreira, Norbert Pfeifer

   Remote Sensing of Environment 2026

   Abs Bib HTML PDF Code

   Precise forest inventories are fundamental for sustainable ecosystem management, biodiversity conservation, and assessment of forest carbon stocks. Light Detection and Ranging (LiDAR) has emerged as a dominant remote sensing technique capable of accurately characterizing the vertical canopy profile over large areas. While the growing availability of LiDAR datasets from diverse acquisition platforms has driven significant advances in forest structural component and individual tree segmentation tasks, inherent heterogeneity in data characteristics across multiple platforms poses a critical challenge for cross-platform generalization. Traditional data-driven algorithms often fail to generalize across multi-platform forest datasets due to substantial heterogeneity in data characteristics from different acquisition platforms. Furthermore, conventional multi-platform mixed training strategies induce negative transfer, ultimately undermining segmentation performance. To address the inherent heterogeneity and negative transfer challenges arising from multi-platform datasets, we introduce the Multi-platform Synergistic Training (MST) framework, a unified, data- and model-driven representation learning framework. This framework initially pretrains with virtual synthetic forest datasets to extract generalized feature representations, followed by platform-specific fine-tuning utilizing real-world datasets. Extensive experiments were conducted to evaluate the proposed MST on semantic and instance segmentation tasks using forest benchmark datasets from multiple platforms. The results reveal that the MST achieves consistently high segmentation performance across multi-platform datasets (from airborne, unmanned aerial vehicle, mobile and terrestrial laser scanning). Additionally, leveraging MST pretraining enables the use of only 20% of labeled real-world data to match the segmentation accuracy achieved by training on fully annotated datasets. The MST framework therefore represents a powerful and effective representation learning framework with the potential to support downstream forest inventory and ecological applications.

   @article{RSE_2026,
   title = {Cross-platform forest understanding: A multi-platform synergistic training framework for generalized forest point cloud segmentation},
   journal = {Remote Sensing of Environment},
   volume = {342},
   pages = {115467},
   year = {2026},
   issn = {0034-4257},
   doi = {https://doi.org/10.1016/j.rse.2026.115467},
   url = {https://www.sciencedirect.com/science/article/pii/S0034425726002373},
   author = {Jundi Jiang and Yueqian Shen and Jinhu Wang and W. Daniel Kissling and Markus Hollaus and Hongjun Su and Jinguo Wang and Vagner Ferreira and Norbert Pfeifer},
   keywords = {Forest scene, Semantic segmentation, Instance segmentation, Cross-platform, Negative transfer, Point cloud}
   }
   

2. Boosted bagging: A hybrid ensemble deep learning framework for point cloud semantic segmentation of shield tunnel leakage

   Jundi Jiang, Yueqian Shen, Jinhu Wang, Yufu Zang, Weitong Wu, Jinguo Wang, Junxi Li, and Vagner Ferreira

   Tunnelling and Underground Space Technology 2025

   Abs Bib HTML PDF

   Accurate leakage detection in subway shield tunnels remains challenging due to the complex geometry and subtle moisture signatures in structural point clouds. Single-model architectures present limited adaptability to heterogeneous leakage and different tunnel scenarios, leading to substantial performance and compromised generalization capacity. To address these challenges, this research proposes Boosted Bagging, a hybrid ensemble deep learning framework for precise semantic segmentation of leakage in tunnel point clouds. In this method, we propose a boosted learner that integrates three specialized base classifiers to learn complex feature representations of diverse leakage patterns. The boosted learner further synergizes with the bagging strategy to enhance generalization capacity by parallel training on randomized data subsets and voting strategy. Integrating boosting and bagging strategies results in a more precise and robust tunnel leakage segmentation. Moreover, the Lovasz Hinge Loss is introduced to address severe sample imbalance between the leakage and background classes. The experimental results demonstrate the effectiveness of Boosted Bagging in terms of segmentation accuracy and robustness. Comparative experiments with state-of-the-art segmentation methods reveal a notable enhancement in segmentation accuracy. Moreover, its superior performance on test datasets highlights the strong generalization capability of the method.

   @article{jiang2025boosted,
   title={Boosted bagging: A hybrid ensemble deep learning framework for point cloud semantic segmentation of shield tunnel leakage},
   author={Jiang, Jundi and Shen, Yueqian and Wang, Jinhu and Zang, Yufu and Wu, Weitong and Wang, Jinguo and Li, Junxi and Ferreira, Vagner},
   journal={Tunnelling and Underground Space Technology},
   year={2025},
   publisher={Elsevier}
   }
   

3. A workflow for extracting ungulate trails in wetlands using 3D point clouds obtained from airborne laser scanning

   Jinhu Wang, Perry Cornelissen, and W. Daniel Kissling

   Frontiers in Remote Sensing 2025

   Abs Bib HTML PDF Data Code

   Ungulates and other mammalian herbivores can create trails in dense vegetation by trampling and browsing. This can affect vegetation structure and result in the fragmentation of closed, high vegetation, with subsequent impacts on biodiversity. Manually mapping trails in the field or from aerial photographs can be challenging and time-consuming, especially in inaccessible or difficult-to-access habitats such as wetlands and if trails occur beneath the canopy of woody plants (i.e., trees and shrubs) or in other tall vegetation (e.g., reed). Airborne laser scanning (ALS) provides an alternative method because Light Detection and Ranging (LiDAR) can record returns from both the canopy and the ground, as some laser pulses pass through gaps in the vegetation, resulting in highly accurate and dense three-dimensional (3D) point clouds. Here, we present a workflow for extracting ungulate trails using 3D point clouds obtained from country-wide ALS surveys, illustrated by red deer trampling in reedbeds within a 36 km2 marsh area of a Dutch nature reserve. The workflow starts by pre-processing to retile the LiDAR point clouds to designated tiles and removes outliers from the raw data. The (near-)terrain points are then segmented using an iterative filtering algorithm, and digital terrain models are generated with a user-defined resolution. Finally, trail cells are extracted by thresholding the residuals from iterative Laplacian smoothing and then refined by sparse 3D structure tensor voting. The parameters of the workflow were optimized with comprehensive sensitivity analyses. Applying the workflow resulted in a classification of trail and non-trail grid cells at 10 cm resolution across the study area. Compared to manually labeled ground truths, the results showed an overall accuracy of 93% and 90% in regions of red deer grazing only and both geese and red deer grazing, respectively. To test its transferability, the workflow could be applied to other LiDAR data (e.g., ALS surveys from another flight campaign in the same study area or in a different country), to other nature areas (e.g., other rewilding sites or other wetlands), and to other ungulate species (e.g., domesticated livestock or other native large herbivores).

   @article{wang2025workflow,
   title={A workflow for extracting ungulate trails in wetlands using 3D point clouds obtained from airborne laser scanning},
   author={Wang, Jinhu and Cornelissen, Perry and Kissling, W. Daniel},
   journal={Frontiers in Remote Sensing},
   year={2025},
   publisher={Frontiers}
   }
   

4. Multi-temporal high-resolution data products of ecosystem structure derived from country-wide airborne laser scanning surveys of the Netherlands

   Yifang Shi, Jinhu Wang, and W. Daniel Kissling

   Earth System Science Data 2025

   Abs Bib HTML PDF Data

   Recent years have seen a rapid surge in the use of light detection and ranging (lidar) technology for characterizing the structure of ecosystems. Even though repeated airborne laser scanning (ALS) surveys are becoming increasingly available across several European countries, so far, only a few studies have derived data products of ecosystem structure at a national scale, possibly due to a lack of free and open-source tools and the computational challenges involved in handling the large volumes of data. Nevertheless, high-resolution data products of ecosystem structure generated from multi-temporal country-wide ALS datasets are urgently needed if we are to integrate such information into biodiversity and ecosystem science. By employing a recently developed, open-source, high-throughput workflow (named “Laserfarm”), we processed around 70 TB of raw point clouds collected from four national ALS surveys of the Netherlands (AHN1–AHN4, 1996–2022). This resulted in ∼59 GB raster layers in GeoTIFF format constituting ready-to-use multi-temporal data products of ecosystem structure at a national scale. For each AHN (Actueel Hoogtebestand Nederland) dataset, we generated 25 lidar-derived vegetation metrics at 10 m spatial resolution, representing vegetation height, vegetation cover, and vegetation structural variability, together with auxiliary data (∼12 GB) such as raster layers of point density; pulse density; flight line timestamp information; terrain and surface elevation; and masks of water areas, roads, buildings, power lines, and NA values (areas with no points available). The data enable an in-depth understanding of ecosystem structure at a fine resolution across the Netherlands and provide opportunities for exploring ecosystem structural dynamics over time. To illustrate the utility of these data products, we present ecological use cases that monitor forest structural change and analyse vegetation structure differences across various Natura 2000 habitat types, including dunes, marshes, grasslands, shrublands, and woodlands. The provided data products and the employed workflow can facilitate a wide use and uptake of ecosystem structure information in biodiversity and carbon modelling, conservation science, and ecosystem management. The full data products are publicly available on Zenodo (https://doi.org/10.5281/zenodo.13940846, Shi et al., 2025a).

   @article{shi2025multi,
   title={Multi-temporal high-resolution data products of ecosystem structure derived from country-wide airborne laser scanning surveys of the Netherlands},
   author={Shi, Yifang and Wang, Jinhu and Kissling, W. Daniel},
   journal={Earth System Science Data},
   year={2025},
   publisher={Copernicus Publications}
   }