Predicting post-fire vegetation recovery patterns in three different forest types

<p>Wildfire disturbances severely modifies the ecosystem structure and natural regeneration processes. Predicting mid- to long-term post-fire vegetation recovery patterns, is pivotal to improve post-fire management and restoration of burned areas forest ecosystems management. Currently, many research efforts have been conducted, in order to monitor and predict wildfires, using Machine Learning and Remote sensing techniques. Instead, the method proposed in this study combines Satellite images and Data Mining algorithm to process data collected by time series and regional forest dataset to predict post-fire vegetation recovery patterns. For this reason, we analysed Normalized Burn Ratio (NBR) patterns from Landsat Time series (LTS), to assess post-fire vegetation recovery for several wildfires that occurred in three different forest Corine Land Cover classes (311, 312, 313) in the Basilicata region during the period 2005-2012. Random Forest model, was used to classify the observed recovery patterns and investigate the influence of burn severity, topographic variables, climate and spectral vegetation indices on post-fire recovery. Image acquisition and Random Forest classifier was undertaken in Google Earth Engine (GEE). Results of bootstrapping classification, across forest type, show high percentage for high recovered (HR) classes and moderate recovered (MR) classes and moderate-low percentage for low (LR) and unrecovered (UR) classes. Specifically, in the holm- and cork-oak and oak forests show medium to high recovery rates, while Mediterranean pine and conifer-oak forests show the slowest recovery rates. Different post-fire recovery patterns are related to fire severity, vegetation type and post-fire environmental conditions. Our methodology shows that post-fire recovery classification, using RF classifier provides a robust method for both local and broad scale monitoring for mid- to long-term recovery response.</p><p>Keywords: Wildfires, Post-fire recovery, Landsat Time Series (LTS), Google Earth Engine, Wildfires, Machine Learning, Random Forest.</p>