Comparison of machine learning algorithms for forest parameter estimations and application for forest quality assessments

Forest parameters have been estimated using various regression methods based on satellite data. However, there are a few concerns regarding further application of these predicted parameters. Current forest quality (FQ) assessments are mainly based on forest parameters collected in the field. To evaluate FQ quickly and comprehensively, forest parameters estimated from satellite images can be used. Black locust (Robinia pseudoacacia) plantations have experienced increases in tree mortality and reductions in tree growth caused by human destruction and poor natural conditions on the Loess Plateau. Assessing the FQ of the black locust plantations in these areas is critical. In this study, four machine learning algorithms (MLAs) classification and regression tree (CART), support vector machine (SVM), artificial neural network (ANN) and random forest (RF) were first implemented and compared regarding their ability to estimate forest parameters of black locust plantations on the Loess Plateau. The results indicated that among the four MLAs, the CART method achieved the lowest accuracy, the SVM and ANN methods had moderate performances, and the RF obtained the highest accuracy and lowest error for all forest parameter estimates. As a result, the RF algorithm was chosen to predict the forest parameters. The highest R-2 values among the forest parameters were predicted for the diameter at breast height (DBH, R-2 = 0.85 and relative root mean square error (rRMSE) = 0.18), and the lowest R-2 values were predicted for the forest aboveground biomass (AGB, R-2 = 0.66 and rRMSE = 0.25). These predicted forest parameters as well as the topographic factors derived from a digital elevation model (DEM) were used to assess the FQ of the study area. Among the four stand age classes, the middle-aged forest had the lowest proportion of poor quality forests (7%) and the highest proportion of good quality forests (52%), whereas overmature forests presented a moderate proportion of poor quality forests (23%) and good quality forests (21%). These results provide the necessary foundation for forest management and guidance for regional reforestation on the Loess Plateau.