Many drylands of the world have experienced dramatic changes in vegetation from perennial grasslands to woody plant dominance over the past several centuries. These states are thought to be very stable under current climate as a result of feedbacks between woody plants and soil properties such that grass recovery rarely occurs, and restoration is difficult. When grass recovery occurs, spatiotemporal heterogeneity in recovery patterns at multiple scales makes it challenging to identify the drivers and processes governing those dynamics. New technologies that link big data analytics with computer vision and image processing are promising approaches that can move restoration ecology forward.
A 15,708 m2 research site in the northern Chihuahuan Desert was cleared in 1996 and kept barren until 2007. Following treatment and a concurrent multi-year (2004-08) period of above-average rainfall, grasses appeared to recolonize the site. Our objectives were to: 1) analyze the patterns of vegetation change through time from over-head photos and 2) use deep learning and UAV fine-scale imagery to estimate above ground biomass and determine the drivers of grass biomass.
Results/Conclusions
We found that the site transited from bare-soil to increased litter and short-lived forbs followed by the establishment of grasses. The perennial grasses Aristida sp. and Dasyochloa pulchella representing 26% and 4% of the foliar cover. Mesquite were also identified across the site and represented much less of the foliar cover (<2%). In addition, 15 additional species of annual and perennial grasses and forbs, each representing <1% of the foliar cover, were also identified.
Automated classification of the grass recovery site with human guided deep learning was able to rapidly classify and provide volumetric estimates of the study area vegetation. Estimates of above-ground biomass were a function of soil moisture, soil water redistribution, and erosion processes.
This approach, offered continuous classification of the composition, and once trained, requires much less time and effort to complete than traditional field methodologies (e.g. quadrat, line point intercept). Due to the continuous surface inputs, this approach may also increase the likelihood of identifying relatively rare, cryptic, or newly occurring species that may go undetected.