Today，approximately 56% of the world's population—4.4 billion people—live in cities. Urban trees play a significant role in mitigating global climate change^［1］ and are uniquely susceptible to climate change impacts. Urban Forest Effects model^［2，3］ is widely used in urban areas globally to estimate urban forest structure，species diversity，and ecosystem functions. However，conducting urban forest inventories is labor-intensive，especially on private properties，and the results are often not spatially detailed. While remotely sensed data is commonly used in forest applications，traditional optical remote sensing methods struggle to capture three-dimensional forest structures，especially in unevenly aged，mixed-species forests with multiple canopy layers^［4］.

Airborne laser scanning（ALS）is effective for extracting biophysical variables and revising forest inventory maps. The successful use of ALS data has been demonstrated for various applications. For example，ALS has been used to estimate tree height^［5，6］，identify tree species^［7-9］，and estimate tree volume biomass^{［10，11］}，and growth^{［12，13］}. Tree species information at an individual tree level is particularly useful in growth and yield estimates and has been primarily studied for forest applications，such as updating forest inventories. Tree species classification using ALS has not been intensively studied compared with studies on the successful use of ALS for other forest attribute mapping because of the lack of spectral information.

Previous studies have also revealed that combining multispectral information with 3D ALS data can improve the accuracy of tree extraction and tree species classification，as we can take advantage of both datasets. However，challenging factors limit the effective operational use of the fused datasets^{［14，15］}. For example，geometric and radiometric registration between two datasets is demanding because data are normally acquired at different times using different sensors. The recently developed multispectral laser scanning technique is becoming an attractive option for forest mapping because it can provide not only a dense point cloud but also spectral information，which can simplify data processing and facilitate the interpretation of data.

Given the limitations of traditional optical remote sensing in capturing three-dimensional forest structures，it is essential to explore the potential of multispectral laser scanning for urban tree inventories，particularly for species classification. This study aims to assess the feasibility of using multispectral ALS data for urban tree species classification and to analyze the information content of features derived from point clouds and intensity data.

The MLS datasets used in this study were acquired in a suburban area in Espoolahti，southern Finland（60°9′18″N，24°38´24″E）in the southern Boreal Forest Zone. We choose around 822 trees in this area as our field dataset. The land area is approximately 5 km². In our research，we concentrated solely on the vegetated areas，excluding the sea using a water mask created from topographic map data. The area included a diverse range of boreal tree species.

The points were updated through visual interpretation of Titan data and open datasets from the City of Espoo，the National Land Survey of Finland，Google Maps，and Google Street View. Field checks validated the analysis and resolved uncertainties. The reference points' attributes included species，geographic location，living conditions，tree height，and planting date for each tree.

Multispectral Optech Titan data（Teledyne Optech，Toronto，ON，Canada）for the study area were collected in May and June 2016 in collaboration with TerraTec Oy（Helsinki，Finland）from a 650 m flight height. The data acquisition was carried out using a fixed-wing aircraft flying at a constant altitude. The sensor comprises three Titan channels：green（532 nm），near-infrared（1 064 nm），and shortwave infrared（1 550 nm）. Each channel provided separate point clouds. In our preprocessed dataset，the point densities over land areas were approximately 9 points/m² for Channel 1，9 points/m² for Channel 2，and 8 points/m² for Channel 3.

TerraScan（TerraSolid Oy，Helsinki，Finland）was used to preprocess the ALS data and differentiate between ground and nonground points using a standardized procedure. This procedure involved removing noise，such as points detected below the ground level or above the canopy. Subsequently，the point clouds were height-normalized. Ground elevation was subtracted from the point cloud height measurements using a digital terrain model created from the classified ground points of the three channels to eliminate potential discrepancies.

Radiometric calibration of ALS intensity is crucial to ensure successful classification. Therefore，in this study，we implemented relative radiometric calibration. We observed that the intensity values were higher in the middle of the flight path compared to other areas and decreased with scanning height. A range correction was applied to mitigate such effects.

where $I_c$ is the modified intensity，$I$ is the original intensity，$D_i$is the distance from the LiDAR to the point cloud and $D_{\mathrm{r}\mathrm{e}\mathrm{f}}$is the flying altitude（650 m）.

Individual trees were detected using a minimum curvature-based algorithm，which started with creating a canopy height model（CHM）. According to our field dataset of each tree coordination，we set the potential crown area within 5 m². A local maximum filtering algorithm was used to find the treetops in this area. Subsequently，the watershed segmentation method was used to delineate tree crown boundaries without setting a flow threshold in the CHM. Eventually，the point cloud of each tree from the multispectral ALS dataset was created. In the segmentation process，the shape and position of individual tree crowns were identified using the segment boundaries and the location of the highest point within each segment. In this study，first return points from all three channels were utilized to generate CHM.

In this experiment，the features were primarily divided into two types：intensity features and geometric features. The maximum height（H_max）of each tree was calculated from the highest point of all point cloud in each tree segment.

Simultaneously，we got 137 features in each channel from the multispectral ALS data.

In this study，we use 8 machine learning algorithms to compare the classification of tree species.：extra trees（ET），random forest（RF），K-nearest neighbour（KNN），logistic regression（LR），linear discriminant analysis（LDA），classification and regression tree（CART），naive bayes（NB），support vector machine（SVM）. Tree species were estimated based on prediction models by 8 machine learning algorithms using tree features as predictors and tree species as a response for correctly detected trees.

As presented in Fig. 2，using all the intensity and geometric features，the overall tree species classification performed best in the extra tree algorithm and reached 71.7%. When we only use channel 1 features for classification，overall values can only reach 65.7%. Only using features from channel 2 yielded overall values that can only reach 68.3%. Only using features from channel 3 yielded overall values that can only reach 64.8%. The accuracy of all the classifications for each species is shown in Fig. 3.

The confusion matrix analysis reveals a model that performs well for most classes but struggles with a few，particularly Quercus and Sorbus according to Table 2 and Fig. 4. Certain classes，such as Acer，Larix，and Thuja，exhibit high accuracy（≥93%），indicating the model’s ability to correctly classify instances for these classes. By addressing these shortcomings through feature refinement，data augmentation，and model optimization，the overall classification accuracy can be significantly improved. Future work should focus on integrating domain-specific knowledge to enhance feature representation and reduce class overlap.

We also investigated which input features and channels are most relevant for tree species classification based on the measure provided by the RF algorithm for assessing feature importance. If a feature influences the prediction，permuting its values should affect the model error. If a feature is not influential，then permuting its values should have little or no effect on the model error. Table 3 lists the top three features in the classifications based on different combinations of the features. The most important features in the classification based on point cloud features were penetration and higher-level percentiles. Two density-related features at higher and middle layers were also scored as important as higher percentiles. In the case of classification using single-channel features，the 1 064nm wavelength（Channel 2）appears to provide the most valuable information for distinguishing between pine，spruce，and birch species. This is followed by the 1 550nm wavelength（Channel 1）and then the 532nm wavelength（Channel 3）.

Multispectral LiDAR data improved the classification accuracy by approximately 5% to 10% for all channels compared to each channel. This proves our hypothesis about the ability of mALS features in classification. For example，the overall accuracy of 71.7% was obtained in multispectral LiDAR all-channel data，while accuracies of 65.7%，68.3%，and 64.8% were achieved when using only Channel 1，Channel 2，and Channel 3，respectively. Our findings demonstrated the advantage of combining multichannel features over single-channel data in classifying urban trees. However，the sample size of each tree species in this experiment was uneven，which may have affected the model's accuracy. Consequently，a larger and more representative sample will be used in future research. The imbalance in measurement samples reduced classification accuracy to some extent. Addressing this limitation will be a key focus in subsequent studies.

In this study，eight machine learning algorithms were evaluated for their classification performance，each demonstrating distinct strengths and limitations. The selection of an appropriate classification algorithm depends on the specific characteristics of the dataset，including size，dimensionality，and the underlying relationship between features and class labels. Extra trees（ET）and random forests（RF）proved effective in our study due to their ability to handle large，high-dimensional datasets and their robustness against overfitting，which suited the conditions of our dataset. Naive Bayes（NB）was efficient and scalable，especially for high-dimensional data，but its assumption of feature independence limited its applicability in cases with high feature correlation.

It is also important to note that overall accuracy（OA）is influenced by factors such as species composition，stand structure，age，and the methods used to select the best features，which vary among studies. In this research，however，the intensity of laser returns was not calibrated. This limitation can be addressed in future studies. First，we can investigate whether calibrated intensity affects classification results. Second，the use of MCI features in this study mitigated potential variations in intensity.

In conclusion，the ability of mALS compared to single-channel ALS（SCI-Ch）data to characterize tree species in urban areas was assessed in this study. Our classification results indicate that mALS data provided more accurate results than single-channel ALS data for urban tree species classification.