Riverine Land Cover Mapping through Semantic Segmentation of Multispectral Point Clouds

Sopitta Thurachen; Josef Taher; Matti Lehtomäki; Leena Matikainen; Linnea Blåfield; Mikel Calle Navarro; Antero Kukko; Tomi Westerlund

cs.CVMar 23, 2026

Riverine Land Cover Mapping through Semantic Segmentation of Multispectral Point Clouds

Authors: Sopitta Thurachen, Josef Taher, Matti Lehtomäki, Leena Matikainen, Linnea Blåfield, Mikel Calle Navarro, Antero Kukko, Tomi Westerlund,

Organizations: Department of Remote Sensing and Photogrammetry, Finnish Geospatial Research Institute FGI, Vuorimiehentie 5, Espoo, 02150, Finland · Faculty of Technology, University of Turku, Turku, 20014, Finland · Department of Geography and Geology, University of Turku, Turku, 20014, Finland · Faculty of Geological Sciences, Complutense University of Madrid, Avda. de Séneca, 2 Ciudad Universitaria, Madrid, 28040, Spain

Abstract

Accurate land cover mapping in riverine environments is essential for effective river management, ecological understanding, and geomorphic change monitoring. This study explores the use of Point Transformer v2 (PTv2), an advanced deep neural network architecture designed for point cloud data, for land cover mapping through semantic segmentation of multispectral LiDAR data in real-world riverine environments. We utilize the geometric and spectral information from the 3-channel LiDAR point cloud to map land cover classes, including sand, gravel, low vegetation, high vegetation, forest floor, and water. The PTv2 model was trained and evaluated on point cloud data from the Oulanka river in northern Finland using both geometry and spectral features. To improve the model's generalization in new riverine environments, we additionally investigate multi-dataset training that adds sparsely annotated data from an additional river dataset. Results demonstrated that using the full-feature configuration resulted in performance with a mean Intersection over Union (mIoU) of 0.950, significantly outperforming the geometry baseline. Other ablation studies revealed that intensity and reflectance features were the key for accurate land cover mapping. The multi-dataset training experiment showed improved generalization performance, suggesting potential for developing more robust models despite limited high-quality annotated data. Our work demonstrates the potential of applying transformer-based architectures to multispectral point clouds in riverine environments. The approach offers new capabilities for monitoring sediment transport and other river management applications.

Figures & tables

Figure 1: Oulanka river in the northeastern part of Finland

Figure 2: A map of Finland highlighting the three study areas: NS, HN, and JM.

Scanner	1	2	3
Model	VUX-1HA	miniVUX-3UAV	VQ-840-G
Wavelength (nm)	1,550	905	532
Approx. point density (points/m 2 )	1,400	500	1,600
Number of returns (max.)	12	5	5
Maximum scanning angle ( ∘ )	360	120	28 $\times$ 40
Laser beam divergence (mrad)	0.5	0.5 $\times$ 1.6	1

Table 1: The specifications of the laser scanners in the multispectral laser scanning system HeliALS for Oulanka river

Point Attribute	Description
3D coordinates	X, Y, Z positional data from LiDAR scanning
Intensity	Strength of the return signal
Reflectance	Surface reflectivity
Amplitude	Peak amplitude of the return waveform
Deviation	The deviation of the return echo waveform from the system impulse response
Return number	Sequential identifier of each return from one laser pulse

Table 2: Point cloud attributes

Class	Description
Sand	Loose, granular surface material
Gravel	Coarse rock fragments
High vegetation	Tall, dense vegetation such as trees
Low vegetation	Shorter vegetation, grasses, or ground cover
Forest floor	Moss or other low vegetation under high vegetation
Water	Rivers and water bodies

Table 3: Land cover classes and their characteristics.

Figure 4: Example images of the six land cover classes used in the dataset (a) sand, (b) gravel, (c) high vegetation, (d) low vegetation, (e) forest floor, (f) water

Panel of Table 4 — Table 4: Class distribution (%) of annotated data in the three study areas

Panel of Table 5 — Table 5: Details of the three study areas

Panel of Figure 5 — Figure 5: Oulanka point cloud data colored by 3-channel intensities as RGB and the corresponding orthophotos

Augmentation	Range	Description
Scaling	0.8–1.2	Random uniform resizing of the point cloud
Translation	$\pm$ 0.2 units	Shifts all points along x, y, and z axes

Table 6: Data augmentation techniques applied during training.

Channel Set	Feature configuration
XYZ	•3D coordinates
Channel 1	•3D coordinates • 1-channel intensity, reflectance, amplitude, deviation (all from VUX)
Channels 1–2	•3D coordinates • 2-channel intensity, reflectance, amplitude, deviation (all from VUX and miniVUX)
Channels 1–3	•3D coordinates • 3-channel intensity, reflectance, amplitude, deviation (all from VUX, miniVUX, VQ)
All features	•3D coordinates • 3-channel intensity, reflectance, amplitude, deviation (all from VUX, miniVUX, VQ) • Return number, number of returns

Table 7: Feature configurations for different channel sets in the multispectral point cloud ablation study

Channel set	mIoU	Sand	Gravel	High veg.	Low veg.	Forest floor	Water
XYZ	0.643	0.487	0.194	0.874	0.685	0.765	0.852
Channel 1	0.865	0.849	0.908	0.905	0.773	0.763	0.993
Channels 1–2	0.937	0.988	0.991	0.903	0.934	0.813	0.993
Channels 1–3	0.938	0.993	0.971	0.910	0.938	0.818	0.999
All features	0.950	0.990	0.971	0.931	0.930	0.878	0.998

Table 8: Ablation study results showing IoU for each class and mIoU across all classes on the JM test dataset

Channel set	mAcc	Sand	Gravel	High veg.	Low veg.	Forest floor	Water
XYZ	0.751	0.733	0.258	0.893	0.818	0.857	0.949
Channel 1	0.932	0.849	0.984	0.924	0.912	0.924	0.998
Channels 1–2	0.970	0.988	$\approx 1.000$	0.914	0.972	0.950	0.999
Channels 1–3	0.971	0.994	$\approx 1.000$	0.919	0.968	0.948	$\approx 1.000$
All features	0.977	0.990	0.998	0.938	0.970	0.966	0.999

Table 9: Ablation study results showing Acc for each class and mAcc across all classes on the JM test dataset

Figure 6: Confusion matrix for using all spectral features together with pulse information (row-wise normalization). Values shown are rounded to three decimal places.

Channel set	mPrecision	Sand	Gravel	High veg.	Low veg.	Forest floor	Water
XYZ	0.764	0.592	0.438	0.976	0.808	0.877	0.892
Channel 1	0.924	$\approx 1.000$	0.922	0.978	0.835	0.814	0.995
Channels 1–2	0.964	$\approx 1.000$	0.992	0.986	0.960	0.850	0.994
Channels 1–3	0.963	$\approx 1.000$	0.971	0.985	0.969	0.857	0.999
All features	0.971	$\approx 1.000$	0.973	0.992	0.958	0.906	0.999

Table 10: Ablation study results showing Precision for each class and mean Precision across all classes on the JM test dataset

Feature	mIoU	Sand	Gravel	High veg.	Low veg.	Forest floor	Water
Intensity	0.937	0.995	0.980	0.909	0.940	0.802	0.999
Reflectance	0.934	0.993	0.978	0.894	0.930	0.813	0.998
Amplitude	0.767	0.605	0.400	0.908	0.874	0.817	0.999
Deviation	0.719	0.606	0.222	0.883	0.877	0.727	0.997

Table 11: Ablation study results showing IoU and mIoU for each spectral feature. Each model variant is trained using only one input feature at a time with geometry to evaluate the individual contribution of each feature to semantic segmentation performance.

Training data	mIoU	Sand	Gravel	High veg.	Low veg.	Forest floor	Water
NS	0.773	0.857	0.836	0.895	0.571	0.486	0.994
NS + HN	0.950	0.990	0.971	0.931	0.930	0.878	0.998

Table 12: Per-class IoU and mIOU performance comparison comparing models trained on the NS dataset alone and combined NS+HN dataset

Model	mIoU	Sand	Gravel	High veg	Low veg	Forest floor	Water
PTv2	0.950	0.990	0.971	0.931	0.930	0.878	0.998
DGCNN	0.744	0.803	0.331	0.895	0.682	0.787	0.967
RandLA-Net	0.580	0.597	0.384	0.897	0.07	0.544	0.989
RF (500 trees)	0.682	0.718	0.822	0.671	0.577	0.340	0.967

Table 13: Per-class IoU and mIoU comparing PTv2 with RF, DGCNN and RandLA-Net.

Figure 7: Orthophoto (left), prediction using only geometry (middle), and prediction using full feature set (right)

Panel of Figure 8 — Figure 8: Sediment prediction results: Model prediction highlighting sediment mapping (left) and corresponding orthophoto of the study area (right)

Panel of Figure 9 — Figure 9: Outer bank mapping: prediction (left) and corresponding orthophoto (right)

Panel of Figure 10 — Figure 10: Isolated vegetation patch mapping on the upper point bar:prediction (left) and corresponding orthophoto (right)

Panel of Figure 11 — Figure 11: Forest floor mapping on the upper point bar : prediction (left) and corresponding orthophoto (right)

Point Transformer

See all 8 papers

May 21, 2026

3D LULC classification using multispectral LiDAR and deep learning: current and prospective schemes

Land Use Land Cover (LULC) classification is essential for national 3D mapping, geospatial analysis, and sustainable planning. Multispectral (MS) LiDAR provides synchronized spatial-spectral information, and deep learning (DL) enables 3D point cloud semantic segmentation; however, adoption is limited by the lack of publicly available urban and suburban MS LiDAR datasets aligned with National Mapping and Cadastral Agencies (NMCAs) classification schemes. This study addresses these gaps by introducing L1 and L2 NMCA-aligned LULC classification schemes and a new benchmark MS LiDAR dataset. We evaluate seven state-of-the-art DL models and perform spectral ablation studies at both levels of detail. Results show that Point Transformer V3 achieves the best performance, with mIoU of 79.4% (L1, 8 classes) and 58.9% (L2, 20 classes) using a dual-wavelength LiDAR system (532 nm and 1064 nm). Ablation results show that multispectral information improves performance over geometry-only inputs, with gains of 1.1 percentage points at L1 and 7.8 points at L2. These results highlight the value of LiDAR reflectance for fine-grained material discrimination and support the evolution of NMCA LULC schemes toward higher semantic detail. The Loosdorf-MSL dataset contributes a new benchmark for consistent national and international LULC mapping.

Narges Takhtkeshha, Aldino Rizaldy, Markus Hollaus +3

3D Optical Metrology (3DOM) Unit, Bruno Kessler Foundation (FBK), Trento, Italy · Department of Geodesy and Geoinformation, TU Wien, Vienna, Austria · Helmholtz-ZentrumDresden-Rossendorf (HZDR), Helmholtz Institute Freiberg for Resource Technology (HIF), Freiberg, Germany +2

Remote Sensing Image Understanding Point Cloud Classification Land Cover Classification 3D Semantic Segmentation Remote Sensing

Ask PDF

Mar 23, 2026

Benchmarking Deep Learning Models for Aerial LiDAR Point Cloud Semantic Segmentation under Real Acquisition Conditions: A Case Study in Navarre

Recent advances in deep learning have significantly improved 3D semantic segmentation, but most models focus on indoor or terrestrial datasets. Their behavior under real aerial acquisition conditions remains insufficiently explored, and although a few studies have addressed similar scenarios, they differ in dataset design, acquisition conditions, and model selection. To address this gap, we conduct an experimental benchmark evaluating several state-of-the-art architectures on a large-scale aerial LiDAR dataset acquired under operational flight conditions in Navarre, Spain, covering heterogeneous urban, rural, and industrial landscapes. This study compares four representative deep learning models, including KPConv, RandLA-Net, Superpoint Transformer, and Point Transformer V3, across five semantic classes commonly found in airborne surveys, such as ground, vegetation, buildings, and vehicles, highlighting the inherent challenges of class imbalance and geometric variability in aerial data. Results show that all tested models achieve high overall accuracy exceeding 93%, with KPConv attaining the highest mean IoU (78.51%) through consistent performance across classes, particularly on challenging and underrepresented categories. Point Transformer V3 demonstrates superior performance on the underrepresented vehicle class (75.11% IoU), while Superpoint Transformer and RandLA-Net trade off segmentation robustness for computational efficiency.

Alex Salvatierra, José Antonio Sanz, Christian Gutiérrez +1

Department of Statistics, Computer Science and Mathematics and Institute of Smart Cities (ISC), Public University of Navarre (UPNA), Campus de Arrosadía s/n, Pamplona, 31006, Navarre, Spain · Tracasa Instrumental, Cabárceno, 6, Sarriguren, 31621, Navarre, Spain

3D Point Cloud Segmentation 3D Semantic Segmentation Remote Sensing Benchmarks Airborne Laser Scanning

Ask PDF

Jun 8, 2026

An Enhanced Geometric-Spectral Feature Learning Framework for Airborne Multispectral Point Cloud Classification

Multispectral point cloud (MPC) is composed of 3D spatial-spectral information, which holds tremendous potential for accurate land-cover classification. However, the representation power of classification models is limited by inherent high-dimensional and heterogeneous spatial-spectral information, unbalanced sample distribution, and inter-class spectral similarity of airborne MPCs. We build two MPC datasets and propose an enhanced geometric-spectral feature learning framework based on attentions for airborne MPC classification. A key component in our model is a two-stream feature fusion method with attention mechanisms, which enhances the representation capability of spatial-spectral features from high-dimensional heterogeneous MPCs. The first stream aims to extract position-encoded global spectral features with fusion self-attention, and the second stream comprises a multikernel point convolution and feature aggregation attention to extract spectral-guided geometric features. We then develop a residual attention fusion block to integrate the most informative geometric-spectral features from the two parallel streams. Another important contribution of this work is a joint loss function to improve the learning ability on unbalanced and interclass similar samples. Experimental results on two airborne MPC datasets demonstrate the effectiveness of the proposed method compared with the state-of-the-art methods. Furthermore, the codes and datasets used in this paper will be made available freely at https://github.com/HITlixian/TGRS_GSFF.

Xian Li, Yanfeng Gu, Aleksandra Pižurica

School of Electronics and Information Engineering, Harbin Institute of Technology, Harbin 150001, China · Department of Telecommunications and Information Processing, UGent-GAIM, Ghent University, 9000 Ghent, Belgium

Representation Learning Point Cloud Classification Remote Sensing

Ask PDF

Abstract

Point Transformer

Explore similar work

3D LULC classification using multispectral LiDAR and deep learning: current and prospective schemes

Benchmarking Deep Learning Models for Aerial LiDAR Point Cloud Semantic Segmentation under Real Acquisition Conditions: A Case Study in Navarre

An Enhanced Geometric-Spectral Feature Learning Framework for Airborne Multispectral Point Cloud Classification