Automated Estimation of Forest Stand Age Using Vegetation Change Tracker and Machine Learning

Abstract

The ability to automatically delineate forest stands and determine their age is useful for natural resources professionals. Two common approaches to estimating forest area and age-class distributions are inventory-based methods, such as Forest Inventory and Analysis (FIA), and remote sensing based methods. Vegetation Change Tracker (VCT) is an algorithm that uses time series stacks of Landsat images to identify forest disturbances. However, additional computation is required to identify type of disturbance. This paper evaluates the usefulness of machine learning tools, such as support vector machine (SVM), for reclassifying VCT disturbances as stand-clearing disturbances or partial disturbances. Overall accuracy for a 2010 VCT disturbance map of the entire state of Virginia was determined to be 87 percent. 100 percent of 2010 Virginia clearcut harvests recorded in a reference dataset were classified as disturbances by VCT. Neighboring disturbed pixels, as classified by VCT, were clumped together and reclassified as stand-clearing disturbances or partial disturbances using SVM and variables for average disturbance magnitude and shape and size metrics of the clumped pixels, with an overall accuracy rate of 86 percent. The users and producers accuracy rates for stand-clearing disturbances were 88 percent and 95 percent respectively. In addition, an algorithm was developed in R for determining years since last stand-clearing disturbance for each pixel in a time series stack of reclassified VCT disturbance maps from 1984 to 2011. Neighboring pixels of the same age, in number of years since last stand-clearing disturbance, were clumped together and correspond, in general, to clearcut harvest boundaries.