Generalizing Predictive Models of Sub-Tropical Forest Inventory Attributes Using an Area-Based Approach with Airborne LiDAR Data

【Objective】Airborne LiDAR is an advanced technology used for the inventory and monitoring of forest resources and ecology, many site-specific and species-specific models had been developed for the estimation of forest inventory attributes. However, the estimations between forest types derived from these models were poorly comparable, thus it would be necessary to develop a model or formula with a stable structure and being suitable for various forest types. 【Method】In this paper, a south subtropical hilly region with an area of 22 100 km2 was taken as the study site to estimate three forest attributes such as the stand volume(VOL), basal area(BA)and mean diameter at breast(DBH)of four forest types(Chinese fir, masson pine, eucalyptus and broadleaf forest)using the airborne discrete return LiDAR and field plot data. Seven LiDAR-derived metrics which describing the complementary 3D structural aspects of the stand canopy were selected to construct five multivariate power models. We tested the performances of these models with 383 field plot measurement data. 【Result】The results indicated that the model consisting of the LiDAR-derived mean point cloud height, canopy coverage, variation coefficient of leaf area density, variation coefficient of point cloud height distribution and 50% height quantile density had the best performance. The R2 of VOL prediction models of four forest type were 0.765, 0.711, 0.748 and 0.683, respectively, the relative root mean square error(rRMSE)ranged from 18.53% to 36.32%, and the mean prediction error(MPE)ranged from 3.37% to 6.95%. The R2 of BA estimation models were 0.572, 0.582, 0.706, and 0.568, respectively, the rRMSE ranged from 16.11% to 30.82%, and the MPE ranged from 3.27% to 5.89%. The R2 of DBH estimation models were 0.574, 0.501, 0.709 and 0.240, respectively, the rRMSE ranged from 1.09% to 28.27%, and the MPE ranged from 1.83% to 5.55%. The relative differences of R2 between the optimal generalizing formula and the optimal model of three attributes of four forest types were less than 5%, and those between rRMSE and MPE were less than 7%. 【Conclusion】The metrics of our model offer clear insights on forestry biophysics, have greats in forestry analytics by accurately depicting the three-dimensional structure of the stand canopy, and perform well in the estimation of various forest types and different forest parameters. The model provides accurate generalization for adaptation, which is beneficial to the operational application of the airborne LiDAR technology on the dynamic monitoring of forest resources. © 2021, Editorial Department of Scientia Silvae Sinicae. All right reserved.