Mapping forest natural disturbances dynamics in the Aosta Valley (Italy) through long-term trends derived from Landsat time series and innovative statistical approaches

Forest dynamics triggered by natural disturbances occurred in the Aosta Valley region were spatially mapped over time using long-term trends derived from Landsat time series spanning over 35 years, from 1985 to 2019. Among biotic and abiotic disturbance agents, the following were selected: wildfires, windthrows, snow avalanches, landslides and insect outbreaks. Landsat TM, ETM+ and OLI images acquired during the vegetative season (from June to September) with less than 80% cloud cover were employed to create synthetic images at one-year interval using the geometric median approach at the pixel-level. Forest dynamics due to disturbance occurrence and the following vegetation recovery were explored through inter-annual time series of different spectral indices such as normalized vegetation indices (Normalized Burn Ratio, Normalized Moisture Index) and the tasseled cap band transformations (wetness, angle). Changes in the linear trends of the spectral indices time series caused by disturbance occurrence were detected using a novel bottom-up approach in which a wavelet basis is adaptively constructed by merging neighbouring segments of the data. This method doesn’t require a priori knowledge of the time series parameters making it fully automated. Prior to perform the trend analysis, vegetation indices time series were filtered to remove residual invalid pixels and fill gaps of one-year length. Considering abrupt disturbances, this method highlighted sensitivity toward both high and low magnitude events and was able to accurately detect different severity degrees within the perimeter of the affected forest area. Historical wildfire perimeters and crown fires patches provided by the forest fire fighting corps of the Aosta Valley were used to perform preliminary severity maps validation. Considering two severity classes, ‘low to moderate’ and ‘moderate to high’, maps produced using the Normalized Burn Ratio achieved an overall accuracy of 83%. Future work is aimed to validate all the selected natural disturbance agents using historical field data available at the regional scale. Moreover, a rigorous and wide scale-based assessment of the capabilities of the algorithm in tracking post-fire forest recovery will be performed by integrating forest structure data from filed surveys and airborne LiDAR measurements.