Optimizing remote sensing workflows using machine learning techniques to quantify windthrow severity across Australian temperate forest.

Windthrow events, characterized by tree uprooting or breakage by strong winds, can result in substantial structural change and biomass loss, altering biodiversity and carbon dynamics. They have led to severe disturbance of native forests in temperate Australia in recent years, and yet their extent and impacts remain largely unquantified. This study aims to advance the monitoring and understanding of windthrow dynamics by integrating remote sensing technologies, machine learning, and field-based data to quantify windthrow severity and associated impacts on the structure of native eucalypt-dominated forests. As a case-study example, it focuses on a major storm event on 9 June 2021 that affected an estimated 40,000 ha of the Wombat State Forest in Victoria, southeastern Australia. The study employs high to very-high resolution satellite and aerial imagery [PlanetScope (3m), NearMap (7.5cm)] and derived indices (Normalized Difference Vegetation Index, NDVI; Blue Normalized Difference Vegetation Index, BNDVI) to nominally map None, Low, Medium, and High severity windthrow zones. These zones were used in stratified random sampling to select 650 (30m×30m) plots in the NearMap imagery, which were analyzed for change in canopy cover using a machine learning workflow involving a Random Forest model. The workflow provided canopy cover reduction estimates from pre- to post-event scenario with high accuracy (96.9%), precision (92.5%), recall (92.8%,) and F1-score (92.68%) across plots in high windthrow severity locations (260) initially and significantly reduced the amount of time and labour for this task. Building on these canopy-level estimates, the final stage will upscale damage quantification across the entire Wombat State Forest by training PlanetScope imagery with very-high resolution canopy cover estimates data from NearMap while employing machine learning models integrating spectral predictors (including dNDVI, dBNDVI, and key multispectral bands). This will produce a high-resolution windthrow severity map, enabling an accurate assessment of windthrow severity across the large and heterogeneous landscape. These outputs will enable biomass-loss estimation from canopy and tree-fall metrics, and will support risk models that integrate remote sensing, biophysical variables, and climate data to predict windthrow susceptibility across the landscape of Australian temperate forests.