Forest cover and canopy health mapping in Australian subalpine landscape: supervised machine learning models for Sentinel-2 and Landsat images

Forests and woodlands worldwide appear to be increasingly vulnerable to decline and mortality, endangering forest-reliant biodiversity and ecosystem services. Consistent with this vulnerability, increased fire frequency and widespread insect-induced dieback has been reported across the subalpine woodlands of southeastern Australia. High-quality spatiotemporal information on the occurrence of the dieback is, however, scarce. This study aimed to develop a method to map distribution of subalpine woodlands and assess their canopy health using Sentinel-2 and Landsat series multispectral satellite imagery. Three-month growing-season geometric median surface reflectance was used for pixel-based supervised classification and regression. Forest masking models were trained on aerial LiDAR-derived forest cover. Canopy condition regression models were trained on ground-based assessment data surveyed at 85 vegetation plots. We tested random-forest (RF), support vector machine (SVM), and multiple linear regression (MLR) to find the algorithm that provides the best accuracy. Cross-validation experiments were undertaken to optimize the model configurations. Using Sentinel-2 surface reflectance, the best forest masking model was SVM (OA = 0.849 ± 0.003), while MLR performed best for canopy condition (<i>r</i>^<i>2</i> = 0.73 ± 0.07). The models built on Landsat surface reflectance performed similarly. The accuracy within our study area is greater than those of existing global/continental satellite-derived forest cover maps. In conclusion, constraining the region of interest and optimizing model configurations can improve forest cover mapping. Integrated with canopy condition regression, our method can identify forest cover and assess canopy condition of Australian subalpine woodlands with reasonable accuracies.