Objective Lidar scanning obtain point cloud date can not only directly measure the three-dimensional (3D) model of the object, but also reveal the intensity of the object. The laser intensity data reflects a variety of the characteristics of the target surface, which can be applied for registration of different measuring stations and filtering of point cloud data. It can also be used to extract and classify the target object by using the intensity data or combining the intensity data with the point cloud RGB data, so as to provide a basis information for feature extraction and leaf area calculation and biomass estimation. However, the influence factors such as the angle and distance will impact the laser intensity data of the same features. These deviations reduce the accuracy of point cloud registration, classification, and extraction, which is not conducive to the full use of point cloud information. Therefore, it is needed to establish a correction model to make the intensity data accurately reflects the feature information for rapid extraction.

Methods Based on lidar ranging equation, first, a pre experiment processing is designed to predict the influence of scanning background, illumination change, and leaf inclination on the intensity data of the research objects. Second, the polynomial models are analyzed to fit the intensity data corrected equation, and the standard value of each material is defined through indoor experiment. Through the model, the angle, which is not easy to measure, is converted into height. After that, the intensity correction models of seven different materials (ginkgo leaves front and back, ginkgo branch, white paper, soapberry leaves front and back, and soapberry branch), with six different distances (2 m, 3 m, 4 m, 5 m, 6 m, and 7 m) and six different heights (0 m, 0.5 m, 1 m, 1.5 m, 2 m, and 2.5 m) is going to be established. Third, the coordinate transformation method of outdoor standing trees (two species: ginkgo and soapberry) is designed, and the correction model is used to obtain the corrected data of the reflectance of branches and leaves for each species. Finally, a threshold method and a random forest method are selected for intensity classification, and the classification result is analyzed to achieve the purpose of using point cloud intensity data for ground feature classification.

Results and discussions 1) In the comprehensive analysis of various correction models, quadratic polynomial usually has the characteristics of simple calculation way and good simulation effect. For the seven kinds of materials, the data range after correction is less than 0.1, and the reflection intensity of each material is more stable than raw data, and is not affected by the distance and height. 2) Before correction, the peak value of reflected intensity of point cloud with leaves is smaller than that of yellow leaves and leaves off. The reflected intensity of leaves is less than that of branches. The reflected intensity of ginkgo ranges from 0.163 to 0.506 and that of soapberry ranges from 0.182 to 0.505. After correction, the range of intensity data for all materials decreases by an order of magnitude and fluctuates within a very small range of value. Among the correction results, the minimum range appears only 0.005 in the data of front of gingko leaves. While in the data of back of gingko leaves the range is also smaller, only 0.007. 3) For standing tree experiment, reflectance for single target and all return point cloud are all drawn. Experimental results show that the curves of gingko reflected intensity are smoother. The reflected intensity interval of ginkgo point clouds is 35, that is, the reflected intensity is 0.175 with a largest number of clouds. After the correction of the leaf model and the branch model, the reflection intensity of the most standing point clouds is 0.035 and 0.065. For soapberry, the intensity interval is 18, and the reflected intensity is 0.09 with a largest number of clouds. After the correction of the leaf model and the branch model, the reflection intensity of the most standing point clouds are 0.05 and 0.065. 4) Different classification methods are used to separate leaves and branches from the corrected intensity of the two tree species, the threshold classification is more applicable. The highest accuracy of the threshold classification with row data is 37.406%. After correction, the classification accuracy of ginkgo leaf model could reach 75.780% which increases by 83%. With the addition of RGB information, the classification accuracy of the random forest model with corrected data is improved from 85.645% to 91.504% which increases by 6.8%. The accuracy of leaf model correction for both species are 91.504% and 84.323%, respectively.

Conclusions The calibration model method established in this paper can accurately correct the laser intensity data of the natural diffuse reflection target object. The selected natural standing-tree verifies the experiment successfully. Branches and leaves can be distinguished by the reflection intensity after calibration, which provides the possibility to further use point cloud data for branch and leaf separation and tree species identification.