Multispectral airborne laser scanning for species classification in Espoonlahti (MS-ALS-SPECIES) dataset

General description Data for benchmarking tree species classification from multispectral airborne laser scanning point clouds. The dataset contains 6326 samples across nine species (pine, spruce, birch, maple, aspen, rowan, oak, linden and alder) that are collected from a suburban area in Espoonlahti, Finland. The point clouds have been acquired with two different actively illuminated three wavelength airborne laser scanning systems: The helicopter-borne HeliALS point clouds are close-range (altitude AGL = 100 m) and have high-density (> 1000 points/m2). The Optech Titan point clouds have been collected with a fixed-wing aircraft from medium range (altitude AGL = 700 m) and have medium density (35 points/m2). The dataset has been utilized in publication:Taher, J., Hyyppä, E., Hyyppä, M., Salolahti, K., Yu, X., Matikainen, L., ... & Hyyppä, J. (2025). Multispectral airborne laser scanning for tree species classification: a benchmark of machine learning and deep learning algorithms. arXiv preprint arXiv:2504.14337, https://doi.org/10.48550/arXiv.2504.14337 . Data description The ground-truth reference is provided in two separate .csv - files: final-segments-with-species.csv : Tree species labels for the tree segments (6238 samples). training-and-test-segments-with-species.csv : The tree species labels used in the benchmarking study Taher et al., 2025. The segments are divided into 1065 training segments and 5261 test segments. The participants of the benchmarking study were given the species labels of the 1065 training segments, and they were given the task to predict the species for the 5261 test segments. Compared to final-segments-with-species.csv, the training-and-test-segments-with-species.csv includes a few additional segments, and a small number of incorrect species labels that have been later corrected in final-segments-with-species.csv with another round of field visits. The point cloud data is saved segment-wise as las-files, and the filename of each tree segment includes the test site and the segment index. The point cloud data is contained in three separate .zip - folders: full_data_HeliALS.zip : All tree segment point clouds from the helicopter-borne three wavelength (Riegl VUX-1HA 1550 nm, Riegl miniVUX-1DL 905 nm, Riegl VQ-840-G 532 nm) laser scanning system. sample_data_HeliALS.zip : A stratified subset of the HeliALS data (10 random samples from each species category). full_data_Optech_Titan.zip : All tree segment point clouds from the fixed-wing aircraft based three wavelength (channel 1 = 1550 nm, channel 2 = 1064 nm and channel 3 = 532 nm) Optech Titan laser scanning system. Ground-truth reference The columns in the final-segments-with-species.csv ground-truth reference sheet contain the following information:test_site_name corresponds to the original spatial division of the Optech Titan LAS files (either A or B), segment_id refers to the index of the tree segment, height is the height of the segment in meters, species_code refers to the integer encoded tree species (ranges from 1 to 9, see table below), profile_category and profile_category_name describe the quality of the watershed segmentation (some segments may contain multiple trees, for example). The columns crown_class and crown_class_name describe the position of the tree in the forest canopy (dominant, co-dominant, etc.). The mapping from the species code into the species name is given by: species code species name 1 Pine (Pinus sylvestris) 2 Spruce (Picea sp.) 3 Birch (Betula sp.) 4 Maple (Acer platanoides) 5 Aspen (Populus tremula) 6 Rowan (Sorbus sp.) 7 Oak (Quercus robur) 8 Linden (Tilia sp.) 9 Alder (Alnus sp.) For further details with respect to profile categories and crown classes in the reference sheet, please refer to Section 2.6 in Taher et al., 2025. The reference sheet training-and-test-segments-with-species.csv has an additional column, test_set_flag, which indicates whether the sample was included in the test set in the benchmarking study by Taher et al. (2025). The ground-truth species labels were collected during the summer of 2024. HeliALS point cloud data The HeliALS data has been stored in LAS version 1.2 point format 3. The tree segment point cloud files have been populated with the following standard and extra bytes fields: xyz-coordinates: The x- and y-coordinates follow the ETRS-TM35FIN coordinate system. The z-coordinates have ellipsoidal height values. intensity: Intensity value of the point. The values are recorded with 16-bit resolution and are thus between 0 and 65,536. return number: Pulse return number for a given output pulse. Note, that the maximum value is seven due to LAS version 1.2 limitations, although the scanners (VUX-1HA and VQ-840-G) are capable of measuring return numbers higher than that. number of returns: Total number of returns for a given output pulse. Similarly to the return number attribute, the maximum value is seven. scan direction flag: The direction in which the scanning mirror was moving at the time of the given output pulse (left or right with respect to track). edge of flight line: The last point on a given scan line before a change in direction. classification: The standard ASPRS classification with the following classes: 0 = created, never classified, 2 = ground and 7 = low point (noise). The ground points were classified using cloth simulation filter (Zhang et al. 2016) and the noise points by statistical outlier method. scan angle rank: The cross-track component of the laser beam angle with respect to nadir. user data: Indicates the scanner from which the point originated: 1 = VUX-1HA (1550 nm), 2 = miniVUX-1DL (905 nm) and 3 = VQ-840-G (532 nm). point source id: Refers to original 1 minute data files in sequential order. gps time: GPS week time at which the point was acquired. red, green and blue: The color channels are populated with nearest neighbor interpolated (search radius = 0.25 m) intensity values (red = 1550 nm, green = 905 nm and blue = 532 nm) that are clipped, transformed, scaled and gamma-corrected. The intensity values at the red channel are clipped to reside between [30000, 50000], at the green channel between [32000, 45000] and at the blue channel between [15000, 55000]. The transformation, scaling and gamma correction has been carried out for each channel after the clipping operation followingly: I_out = (2^16)*((I_in - I_min)/(I_max - I_min))^gamma, where gamma = 2.5 . amplitude : Echo signal amplitude in decibels. reflectance : Echo signal reflectance in decibels. deviation : Deviation of the echo pulse shape from the ideal system impulse response. Notification for the user: For a more detailed definition of the reflectance, amplitude and echo deviation attributes for the Riegl laser scanners, please refer to "Pfennigbauer, M., & Ullrich, A. (2010, April). Improving quality of laser scanning data acquisition through calibrated amplitude and pulse deviation measurement. In Laser Radar Technology and Applications XV (Vol. 7684, pp. 463-472). SPIE." Additional information on HeliALS dataAcquisition date: The HeliALS dataset was acquired on July 20 and 28, 2023.Data acquisition and preprocessing of the data for tree species benchmarking: Taher et al., 2025. Multispectral airborne laser scanning for tree species classification: a benchmark of machine learning and deep learning algorithms, https://arxiv.org/abs/2504.14337 Optech Titan point cloud data The Optech Titan data has been stored in LAS version 1.2 point format 3. The tree segment point cloud files have been populated with the following standard and extra bytes fields: xyz-coordinates : The x- and y-coordinates follow the ETRS-TM35FIN coordinate system. The z-coordinate has an undefined offset (approximately 10 to 20 cm) to the N2000 height system. intensity : The range corrected intensity value of the point. The values are recorded with 12-bit resolution and are thus between 0 and 4096. The intensity range correction has been applied in the following manner: I_out = I_in*(R^2/R_ref^2) , where I_in is the original intensity, R is the measurement range and R_ref is the reference measurement range (R_ref = 700 m). return number : Pulse return number for a given output pulse (the Optech Titan system provides up to four returns per output pulse). number of returns : Total number of returns for a given output pulse. scan direction flag : The direction in which the scanning mirror was moving at the time of the given output pulse (left or right with respect to track). classification : The standard ASPRS classification with the following classes: 1 = unclassified, 2 = ground, 5 = high vegetation, 6 = building, 7 = low point (noise). scan angle rank : The cross-track component of the laser beam angle with respect to nadir. user data : Indicates the wavelength channel from which the point originated: 1 = channel 1 (1550 nm), 2 = channel 2 (1064 nm) and 3 = channel 3 (532 nm). point source id : Refers to the ordinal of the flight line during data collection. gps time : GPS week time at which the point was acquired. red, green and blue : The color channels are populated with nearest neighbor interpolated (search radius = 1.0 m) intensity values (red = 1550 nm, green = 1064 nm and blue = 532 nm). The intensity values are clipped to reside in range [0, I_median + 2*I_std] and scaled to obtain the output color value, as follows:I_out = (I_in/I_max)*(2^16 -1) , where the processing has been done independently for each color channel. height agl : Height of the point above ground level. Additional information on Optech Titan dataAcquisition date: 14 June 2016Further information on the data and its processing is available in the following publications:Data acquisition: Karila et al., 2018. The effect of seasonal variation on automated land cover mapping from multispectral airborne laser scanning data. International Journal of Remote Sensing, https://doi.org/10.1080/01431161.2018.1528023.Preprocessing of the data for tree species benchmarking: Taher et al., 2025. Multispectral airborne laser scanning for tree species classification: a benchmark of machine learning and deep learning algorithms, https://arxiv.org/abs/2504.14337 Acknowledgements We would like to acknowledge the following people for participating to the data collection, data curation, and the benchmarking study: Sopitta Thurachen, Ville Luoma, Markus Holopainen, Gefei Kong, Hongchao Fan, Petri Rönnholm, Matti Vaaja, Antti Polvivaara, Samuli Junttila, Mikko Vastaranta, Stefano Puliti, Rasmus Astrup, Joel Kostensalo, Mari Myllymäki, Maksymilian Kulicki, Krzysztof Stereńczak, Raul de Paula Pires, Ruben Valbuena, Juan Pedro Carbonell-Rivera, Jesús Torralba, Yi-Chen Chen, Lukas Winiwarter, Markus Hollaus, Gottfried Mandlburger, Narges Takhtkeshha, Fabio Remondino, Maciej Lisiewicz, Bartłomiej Kraszewski, Xinlian Liang, Jianchang Chen, Eero Ahokas, Kirsi Karila, Eugeniu Vezeteu, Petri Manninen, Roope Näsi, Heikki Hyyti, Siiri Pyykkönen and Peilun Hu. Cite Any scientific publication using this dataset should cite the following paper: "Taher, J., Hyyppä, E., Hyyppä, M., Salolahti, K., Yu, X., Matikainen, L., ... & Hyyppä, J. (2025). Multispectral airborne laser scanning for tree species classification: a benchmark of machine learning and deep learning algorithms. arXiv preprint arXiv:2504.14337", https://doi.org/10.48550/arXiv.2504.14337 . While this dataset is provided under its respective license, we specifically authorize its application in training deep learning models. We request that researchers wishing to create derivative datasets or republish these materials in a new format reach out to the corresponding authors beforehand.