ES-01. Conifer seedling detection in a postfire conifer forest using drone data and machine learning

Abstract
Quantifying vegetation recovery after disturbance informs important metrics such as carbon, water, and nutrient budget needed for ecosystem resilience and forest management assessment and implementations. However, large uncertainties are associated in quantifying recovery due to; 1) lack of in-situ samples that represent large scale disturbances and 2) the coarse resolution of remotely sensed data in comparison to the size of vegetation in early phases of recovery. Very high-resolution photography (~ 3 cm) in uncrewed aerial systems (UAS) supplies unprecedented opportunities to overcome these challenges and to fill the gaps in understanding of cross scale ecosystem processes. In this study we applied a novel machine learning algorithm (one-class SVM) to a set of UAS-derived predictor variables (texture, vegetation indices, and point cloud distribution within individual vegetation polygons) to detect conifer seedlings (0.5 m – 6 m) in a mountainous postfire conifer forest in the Southern Rockies. First, we implemented a workflow to delineate individual vegetation (can include seedlings, trees, and shrubs) from a larger burned area (~40 ha) using UAS RGB images and structure from motion technology. Then, the above variables were extracted per vegetation polygon and input to the trained one-class SVM to detect conifer seedlings. Our results show that we can detect conifer seedlings with an overall accuracy of 86%. We observe that the accuracy of seedling detection varies by seedling size. We show that low cost UAS images and machine learning can help to advance forest recovery quantification across larger regions than traditional field plots and to enhance ecosystem resilience assessments and forest management strategies. In addition, our study demonstrates that one-class classification can be successfully used for remote sensing applications where in-situ samples of other classes except the major class are extremely limited and unavailable for training.