Beyond Weak Labels: Prompt-Guided Local Refinement for Weakly Supervised Water Segmentation in High-Resolution Multispectral Imagery
本文针对高分辨率多光谱图像中水体分割标签难以获取的问题，提出了一种两阶段框架，通过伪标签学习和局部细化提升分割精度。
0 citationsRead paper
Selected work
Representative Papers
Multispectral airborne laser scanning dataset for tree species classification: MS-ALS-SPECIES
This study addresses the scarcity of publicly available multispectral airborne laser scanning (MS-ALS) datasets with high-quality field validation, which has hindered individual tree species classification research. We present the first open MS-ALS dataset, comprising 6,326 individually delineated trees across nine species in southern Finland, acquired using the HeliALS and Optech Titan dual-system sensors to capture three-wavelength point clouds. Ground truth was collected via an efficient and scalable field protocol. Leveraging deep learning models—including point cloud segmentation and Point Transformer architectures—we achieve high-accuracy species classification, demonstrating particularly strong performance for small-sized and rare tree species. Our results validate the efficacy of multispectral ALS data for fine-grained species discrimination and establish a benchmark platform to advance future research in this domain.
0 citationsRead paper
Riverine Land Cover Mapping through Semantic Segmentation of Multispectral Point Clouds
This study addresses the need for high-precision land cover mapping in riparian zones by introducing Point Transformer v2 to semantic segmentation of multispectral LiDAR point clouds. The proposed approach effectively integrates geometric structure with spectral features—specifically intensity and reflectance—to accurately distinguish land cover classes such as sand, gravel, low and high vegetation, forest, and water bodies. By employing a multi-dataset joint training strategy, the model achieves significantly enhanced generalization capability in scenarios with sparse annotations. Evaluated on the Oulanka River dataset, the method attains a mean Intersection over Union (mIoU) of 0.950, substantially outperforming baseline approaches that rely solely on geometric features, thereby demonstrating its high accuracy and robustness for riparian zone mapping.
0 citationsRead paper
Learning Image-based Tree Crown Segmentation from Enhanced Lidar-based Pseudo-labels
Mapping individual tree crowns is essential for tasks such as maintaining urban tree inventories and monitoring forest health, which help us understand and care for our environment. However, automatically separating the crowns from each other in aerial imagery is challenging due to factors such as the texture and partial tree crown overlaps. In this study, we present a method to train deep learning models that segment and separate individual trees from RGB and multispectral images, using pseudo-labels derived from aerial laser scanning (ALS) data. Our study shows that the ALS-derived pseudo-labels can be enhanced using a zero-shot instance segmentation model, Segment Anything Model 2 (SAM 2). Our method offers a way to obtain domain-specific training annotations for optical image-based models without any manual annotation cost, leading to segmentation models which outperform any available models which have been targeted for general domain deployment on the same task.
0 citationsRead paper
Deep Learning-based Robust Autonomous Navigation of Aerial Robots in Dense Forests
To address autonomous UAV navigation challenges in dense forests—characterized by GNSS denial, low visibility, slender and irregular obstacles, and degraded perception—this paper proposes a semantic-enhanced end-to-end navigation framework. Methodologically, it integrates stereo visual–inertial tightly coupled odometry, semantics-guided deep feature encoding, neural motion primitive evaluation, and introduces two novelties: a lateral maneuver control module and a temporal consistency-aware planning suppression mechanism; additionally, a real-time safety action filtering layer ensures flight stability. Evaluated across three real-world northern forest sites, the framework achieves 100% task completion in medium- and high-density scenes and 80% in extremely dense shrubland. Compared to baseline methods, it improves success rate by 23%, reduces trajectory jitter by 41%, and lowers collision rate by 67%, significantly enhancing robustness and safety in complex forest environments.
0 citationsRead paper
Recent publications
Latest Papers
Beyond Weak Labels: Prompt-Guided Local Refinement for Weakly Supervised Water Segmentation in High-Resolution Multispectral Imagery
本文针对高分辨率多光谱图像中水体分割标签难以获取的问题，提出了一种两阶段框架，通过伪标签学习和局部细化提升分割精度。
0 citationsRead paper
Multispectral airborne laser scanning dataset for tree species classification: MS-ALS-SPECIES
This study addresses the scarcity of publicly available multispectral airborne laser scanning (MS-ALS) datasets with high-quality field validation, which has hindered individual tree species classification research. We present the first open MS-ALS dataset, comprising 6,326 individually delineated trees across nine species in southern Finland, acquired using the HeliALS and Optech Titan dual-system sensors to capture three-wavelength point clouds. Ground truth was collected via an efficient and scalable field protocol. Leveraging deep learning models—including point cloud segmentation and Point Transformer architectures—we achieve high-accuracy species classification, demonstrating particularly strong performance for small-sized and rare tree species. Our results validate the efficacy of multispectral ALS data for fine-grained species discrimination and establish a benchmark platform to advance future research in this domain.
0 citationsRead paper
Riverine Land Cover Mapping through Semantic Segmentation of Multispectral Point Clouds
This study addresses the need for high-precision land cover mapping in riparian zones by introducing Point Transformer v2 to semantic segmentation of multispectral LiDAR point clouds. The proposed approach effectively integrates geometric structure with spectral features—specifically intensity and reflectance—to accurately distinguish land cover classes such as sand, gravel, low and high vegetation, forest, and water bodies. By employing a multi-dataset joint training strategy, the model achieves significantly enhanced generalization capability in scenarios with sparse annotations. Evaluated on the Oulanka River dataset, the method attains a mean Intersection over Union (mIoU) of 0.950, substantially outperforming baseline approaches that rely solely on geometric features, thereby demonstrating its high accuracy and robustness for riparian zone mapping.
0 citationsRead paper
Learning Image-based Tree Crown Segmentation from Enhanced Lidar-based Pseudo-labels
Mapping individual tree crowns is essential for tasks such as maintaining urban tree inventories and monitoring forest health, which help us understand and care for our environment. However, automatically separating the crowns from each other in aerial imagery is challenging due to factors such as the texture and partial tree crown overlaps. In this study, we present a method to train deep learning models that segment and separate individual trees from RGB and multispectral images, using pseudo-labels derived from aerial laser scanning (ALS) data. Our study shows that the ALS-derived pseudo-labels can be enhanced using a zero-shot instance segmentation model, Segment Anything Model 2 (SAM 2). Our method offers a way to obtain domain-specific training annotations for optical image-based models without any manual annotation cost, leading to segmentation models which outperform any available models which have been targeted for general domain deployment on the same task.
0 citationsRead paper
Deep Learning-based Robust Autonomous Navigation of Aerial Robots in Dense Forests
To address autonomous UAV navigation challenges in dense forests—characterized by GNSS denial, low visibility, slender and irregular obstacles, and degraded perception—this paper proposes a semantic-enhanced end-to-end navigation framework. Methodologically, it integrates stereo visual–inertial tightly coupled odometry, semantics-guided deep feature encoding, neural motion primitive evaluation, and introduces two novelties: a lateral maneuver control module and a temporal consistency-aware planning suppression mechanism; additionally, a real-time safety action filtering layer ensures flight stability. Evaluated across three real-world northern forest sites, the framework achieves 100% task completion in medium- and high-density scenes and 80% in extremely dense shrubland. Compared to baseline methods, it improves success rate by 23%, reduces trajectory jitter by 41%, and lowers collision rate by 67%, significantly enhancing robustness and safety in complex forest environments.
0 citationsRead paper