The refined target classification has always been a research hotspot in remote sensing and is also a prerequisite for studies on biomass calculation, global carbon cycle, and energy flow. With the continuous expansion and refinement in remote sensing detection, more effective and accurate target classification is becoming more complex and difficult. 3D spatial information and rich spectral information are typical attributes of a target, which is significant data support for target classification. Hyperspectral lidars have been successfully designed and structured for target classification to achieve the integrated acquisition of 3D spatial information and spectral information. With an aim at this new type of remote sensing data, how to develop and exploit its potential in target classification is of research significance. Therefore, to realize high-precision recognition and classification under complex scenes, we propose a target classification process of spatial-spectral feature optimization selection dependent on the hyperspectral lidar. This method can not only reduce feature redundancy and select the optimal feature combination for target classification but also reduce computational efficiency and save costs, thereby providing new research ideas for refined target classification with hyperspectral lidar.

With the continuous expansion of remote sensing detection, detection targets become more diversified and complicated. Constructing various spatial-spectral features based on spectral information and spatial information is a mainstream method to improve the accuracy of target classification. Based on the technological advantages of the integrated imaging detection of high spatial resolution and hyperspectral resolution, we construct spectral index features of the vegetable index and color index, and geometric features for target classification. Extracting lots of spatial-spectral classification features can enhance the classification accuracy, yet it may produce feature redundancy, increase the calculation cost, affect the classification efficiency, and even lead to declining classification accuracy. Therefore, we put forward a target classification process of spatial-spectral features optimization selection dependent on the hyperspectral lidar. In the feature space built by the hyperspectral lidar, these spatial-spectral features with the best classification significance are determined based on the marine predator algorithm by iterative search and selection to minimize the classification error. Finally, considering the feature heterogeneity of the selected feature combination, the feature correlation is calculated to eliminate feature redundancy and determine the optimal feature combination, thereby improving classification accuracy.

To further explore the technological advantages of hyperspectral lidar for target classification under complex scenes, and to compare and verify the feasibility and universality of the proposed method, we design six different classification strategies with different feature combinations. Classification results of these feature combinations are determined by a random forest algorithm. Total accuracy, average accuracy, Kappa coefficient, accuracy rate, and recall rate are adopted to evaluate the classification results of each category. Table 4 shows that the six different classification strategies yield sound classification results with the total accuracy higher than 89%, the average accuracy of more than 68%, and Kappa coefficient greater than 0.85. Compared with the results of the first three classification strategies, the classification results of the fourth strategy which integrates original spectral information, elevation value, index features, and geometrical features, have been greatly improved. Additionally, the overall accuracy can reach 95.57% with the average accuracy of 84.37% and the Kappa coefficient of 0.9380, whereas the elapsed time is the longest at 5.16 s. The predicted result of target labels is shown in Fig. 6(d). Based on the spatial-spectral feature optimization selection method, the optimal feature combination could be determined to eliminate feature redundancy and enhance classification accuracy. The overall accuracy and average accuracy are increased by 1.56% and 4.36%, respectively, and the elapsed time is reduced by 1.55 s. The predicted results of target labels are shown in Fig. 8(f). The classification results demonstrate that this method can determine the optimal spatial-spectral features for target classification, and provide a new research idea for refined target classification with hyperspectral lidar.

As a new active remote sensing technology, the hyperspectral lidar can combine the technology advantages of passive hyperspectral imaging and lidar scanning imaging and has great application potential in refined target classification under complex scenes. Therefore, we propose a target classification process of spatial-spectral feature optimization selection dependent on the hyperspectral lidar. The index features constructed by the spectral band optimization and geometric features constructed by the local neighborhood surface fitting are extracted and employed to target classification. Finally, the optimal feature combination is determined by the proposed method to achieve high-precision target classification under complex scenes with the scanning scene of 14 different targets. Based on the spatial-spectral feature optimization selection to determine the optimal spatial-spectral feature combination, it can effectively eliminate the characteristic redundancy. This increases the overall classification accuracy by 1.56% and the average classification accuracy by 4.36%, and the elapsed time is reduced by 1.55 s. However, there is a certain degree of misclassification because the spatial structures of some targets are so complex that the laser irradiates to the edge of targets or only part of the laser irradiates to the surface of a target, thus leading to a large deviation in spectrum acquirement. The classification results could be smoothed by the boundary algorithm or conditional random field algorithm to eliminate the salt and pepper noise and improve the classification accuracy.