湖北林业科技

The combination of LiDAR features and multispectral image features with different resolutions generated under the same point cloud density is utilized to explore the effects of different feature classification schemes on the classification accuracy of urban garden tree species. In this study, the acquired 20p.m-2 airborne LiDAR data were used to generate feature rasters of 0.5, 1.0, 3.0, and 5.0m. Combined with the multispectral image features acquired during the same period, the extracted feature variables were classified into six groups of classification schemes, and a combination of object-oriented segmentation and three machine learning algorithms was used to classify sixty typical urban garden tree species and screen the optimal features for comparative analysis. The results showed that: (1) the classification scheme based on the combination of multispectral features and LiDAR features extracted at 0.5m grid scale shows the best results on all three machine learning classifiers, XGBoost, RF and KNN, with overall accuracies of 81.77%, 80.79% and 76.85%, respectively; (2) at the same point cloud density, with the features generated from the LiDAR data raster resolution decreases, the accuracy of the combined classification scheme also decreases gradually, even lower than the results of using a single data source features for classification; (3) using the RF importance measure algorithm to screen feature variables in different classification schemes, the importance of LiDAR data-derived variable features in the combined classification scheme decreases as the resolution of the generated feature raster decreases. In summary, the higher the resolution of the raster features generated under the same point cloud density, the higher the accuracy of the classification of urban garden tree species in combination with multispectral data, and the better the classification effect.