LiDAR-derived methods for volume estimation and individual tree detection in Eucalyptus spp. plantations

Abstract

Accurate and spatially explicit measurements of forest attributes are considered of great importance for sustainable forest management and environmental protection. Improvements in the management of eucalyptus plantations result in multiple industrial and environmental benefits. Remote sensing techniques can increase planting management efficiency by reducing or replacing field sampling that requires a longer time and therefore higher costs. Airborne Light Detection and Ranging (LiDAR) systems have become an important remote sensing technique for forest inventory, mainly because this technology can provide high accuracy and spatially detailed information on forest attributes across entire landscapes. Remote sensing data from LiDAR combined with machine learning techniques and automated individual tree detection algorithms present great potential for modeling forest attributes. This dissertation is focused on the comparison of predictive methods of total stem volume and number of individual trees in plantations of Eucalyptus spp. from LiDAR-derived data. More specifically evaluating: 1- the combined impact of sample size and parametric and non-parametric modeling techniques; 2- the accuracy of algorithms for automatic individual trees detection. The modeling technique that presented the best performance was verified for the OLS method, which was able to provide results comparable to the traditional approaches of forest inventory using only 40% of the total field plots, followed by the Random Forest (RF) algorithm for the same sample size. The Dalponte e Silva automatic detection algorithms presented more accurate results with the lowest commission and omission errors, and consequently better F-scores in most of the sampled plots, obtaining comparable results.