The use of mobile lidar data and Gaofen-2 image to classify roadside trees

Abstract

Roadside trees are a vital component of urban greenery and play an important role in intelligent transportation and environmental protection. Quickly and efficiently identifying the spatial distribution of roadside trees is key to providing basic data for urban management and conservation decisions. In this study, we researched the potential of data fusing the Gaofen-2 (GF-2) satellite imagery rich in spectral information and mobile light detection and ranging (lidar) system (MLS) high-precision three-dimensional data to improve roadside tree classification accuracy. Specifically, a normalized digital surface model (nDSM) was derived from the lidar point cloud. GF-2 imagery was fused with an nDSM at the pixel level using the Gram–Schmidt algorithm. Then, samples were set including roadside tree samples from lidar data extracted by random sample consensus and other objects samples from field observation using the Global Positioning System. Finally, we conducted a segmentation process to generate an object-based image and completed the roadside tree classification at object level based on a support vector machine classifier using spectral features and completed local binary pattern (CLBP) texture features. Results show that classification using GF-2 alone and using nDSM alone result in 67.34% and 69.39% overall accuracy respectively with serious misclassification. The fusion image based on GF-2 and nDSM yields 77.55% overall accuracy. This means that the fusion of multi-source data is a great improvement over individual data. After adding the CLBP texture feature to the classification procedure, the classification accuracy of the fusion image is increased to 87.76%. The addition of CLBP texture features can clearly reduce the noise . Our results indicate that the classification of urban roadside trees can be realized by the fusion of satellite data and mobile lidar data with CLBP texture feature using the target-based classification method. Results also suggest that MLS data and CLBP texture features have the potential to effectively and efficiently improve the accuracy of satellite remote sensing classification.