Scale-space image analysis for mapping spotted knapweed in grassland ecosystems.

Abstract

Premise: This study investigates the value of scale-space representations of spatial features for the identification and mapping of invasive plants using high-resolution multispectral imagery. Scale-space representations combine the two-dimensional spatial domain with a third dimension representing spatial scale, such that features can be represented across both space and image spatial resolution.

Methods: Gaussian pyramids were constructed to create representations of the features across five image resolution levels (between 3 and 46 cm) within the scale space. A case study was employed to evaluate the importance of features across the scale space for the identification and mapping of spotted knapweed in a grassland ecosystem using multispectral imagery acquired using a remotely piloted aircraft system (RPAS). Classification was performed using a random forest classifier that was constructed using a process that included feature and hyperparameter optimization.

Results: The results of this case study show that classifiers constructed with only high spatial resolution features were associated with better classification performance than those constructed with only lower spatial resolution features; however, the best classification is achieved by leveraging features from throughout the feature scale spaces.

Discussion: These findings suggest that optimization of the spatial scale at which features are computed can substantially improve remote sensing image analysis and potentially inform RPAS flight planning to improve mission efficiency.