Investigation on the potential of regression kriging for mapping oriental beech forest site productivity in research forest of Tarbiat Modares University

• Published in: مجله جنگل ایران
• Main Authors: Zahra Ahadi, S.J. Alavi, Seyed Mohsen Hosseini
• Format: Article
• Language: Persian
• Published: Iranian Society of Forestry 2018-01-01
• Subjects: geostatistics; interpolation; kriging; machine learning; random forest; regression
• Online Access: http://www.ijf-isaforestry.ir/article_57291_2c3593a2223854f92088d73362996b9e.pdf

Forest resources mapping is a prerequisite for sustainable forest management. Site productivity is a key indicator of forest ecosystem services like wood production, carbon sequestration, etc. Due to the extent of Hyrcanian forests and mountainous areas in these forests that are sometimes difficult to access, it seems necessary to find suitable methods for mapping the quantitative parameters in these forests. In this study, site form index which is the most reliable criterion for evaluating site productivity of mixed and uneven stands was used. This study aims at mapping beech forest site productivity by using regression kriging in research forest of Tarbiat Modares University. For this purpose, 123 0.1 ha circular sample plots were laid out in beech dominated stands. The height and diameter of beech trees with DBH ≥ 7.5 cm within each plot was recorded. Some primary and secondary variables were also extracted from DEM in sample plots to be used in regression kriging. We investigated the differences between two predictive approaches: random forests and linear regression as the base model for regression kriging technique. Results of 10-fold cross-validation demonstrate that by using criteria such as mean error, mean absolute error, root mean square error, relative mean error, relative root mean square error, the random forests algorithm outperforms the linear regression and kriging techniques, with average decreases of ca. 70% in Root Mean Squared Error (RMSE). Hence, the regression kriging technique with random forest as the base model is recommended in order to better understand the more complex environment-beech forest site productivity relationships.