PS 33-213 - Deep learning and high-resolution drone imagery accurately maps shrub species in diverse California coastal communities

Abstract: Accurate maps of wildland plant species distributions across large scales are needed for effective conservation efforts. Such mapping is challenging in diverse, dense plant communities because labor-intensive ground-based surveys can cause damage to sensitive flora. Very high-resolution satellite imagery (~30 cm/pixel) is widely available on a spatial and temporal scale needed for plant community conservation across large areas, but imagery can be expensive and lack sufficient resolution to identify plants to species. In contrast, inexpensive drone-based imagery combined with deep-learning analysis can aid in the accurate classification of plant species, in these communities across large regions. We sought methods to map species distribution in large areas of complex chaparral, coastal sage scrub, and oak woodland habitat using deep learning methods and drone imagery. Our Convolutional Neural Network (CNN) coupled with object-based image analysis (OBIA) classification approach outperformed random forest and support vector machine methods. It could even distinguish two congeneric shrub species with similar morphological characteristics. Similar accuracies were attained when applied to neighboring sites. The CNN+OBIA method had the greatest overall classification accuracy (OA) and reliability (κ) across all application sites (Mean OACNN+OBIA = 0.85, Mean κCNN+OBIA = 0.81). CNN+OBIA was the most effective model for classifying several target species and cover types including: Q. agrifolia (Mean κCNN+OBIA = 0.97), bareground (Mean κCNN+OBIA = 0.96), standing deadwood (Mean κ CNN+OBIA = 0.89), A. fasciculatum (Mean κCNN+OBIA = 0.80), A. pumila (Mean κCNN+OBIA = 0.86), A. tomentosa (Mean κCNN+OBIA = 0.72), and coastal sage scrub community (Mean κCNN+OBIA = 0.69). This work is key to the accurate species mapping at the scale and resolution needed for conservation research and endemic species monitoring in chaparral and other wildland plant communities. Locally rare species and intermixed canopies present the greatest challenges to applying these methods.