Comparison between machine learning classification and trajectory-based change detection for identifying eucalyptus areas in Landsat time series

Abstract

In forestry, where species management has inter-annual durations, time series longer than one year have been used to map planted areas, estimate biophysical parameters, and determine planting age. Eucalyptus plantations have characteristics that enhance the use of time series in their classification, such as clear-cutting before planting or previous rotation and the rapid growth of large vegetation cover, which remains stable throughout the rotation period. This article compares the performance of two classification methods for mapping Eucalyptus areas: an object-oriented classification (GEOBIA) using products generated by the LandTrendr change detection algorithm based on trajectories and a classification using machine learning (random forest). Using a Landsat time series from 1985 to 2020, tests were conducted in a pilot area in Rio de Janeiro, Brazil. Both methods showed high accuracy in detecting Eucalyptus areas. However, the trajectory-based classification proved slightly superior, achieving a global accuracy of 0.988 and an F-Score of 0.975, while the classification using the random forest algorithm achieved a global accuracy of 0.954 and an F-score of 0.849. Regarding identifying the initial year of planting, both methods proved effective without showing significant differences (p-value = 0.1003). However, detecting the initial year using the LandTrendr algorithm proved more assertive. Both methods revealed periods of increase and stabilization in Eucalyptus planting throughout the time series, proving promising for determining the location and age of each stand and, thus, obtaining information about the time of use of that area for Eucalyptus cultivation.