Remote Sensing Time-Series Analysis, Machine Learning, and K-Means Clustering Improves Dryland Vegetation and Biological Soil Crust Classification

Abstract

Dryland and semi-arid vegetation communities, although appearing to the casual observer as relatively simplistic and homogeneous, are in fact the opposite. Upon further inspection, semi-arid vegetation is highly complex and heterogeneous at almost any scale. The same holds true for biological soil crust. Growing concern about global changes in climate, nutrient cycles, and land use have required increasing scrutiny of our understanding of these communities and all of their constituents, as we seek to improve forecasting models and inform land management decisions. This thesis aims to provide insight to the paradigm of how we create and interpret vegetation classifications in a semi-arid ecosystem. In the first chapter, I examine the potential of new remote sensing imaging platforms in combination with machine learning algorithms and cloud computing as they apply to time-series analyses for vegetation classification. The results of this indicate that sinusoidal approximations ("Harmonic Models") of vegetation indices are able to predict vegetation cover with nearly the same accuracy as monthly composites, and that a combination of both perform no better than either. Additionally, I examine how assigning classes to training data (e.g. species-level, plant functional type) influence the classification accuracy, interpretability, and potential uses. Stricter class membership requirements at increasingly aggregated scales (e.g. PFT) lead to greater accuracies. Finding the implication of this conclusion unsatisfactoryat the extreme, everything was either cheatgrass or bare, the shrub class succumbing almost entirely to vii errors of omission-I investigated approaching other methods to assign classes to field data that captured more of the realities of semi-arid vegetation within our study area. To this extent, k-means clustering was used to determine what community classes were present in the field data. The outcome of this approach is a class where each potential constituent cover has a known distribution. Overall accuracies were found to be lower for this approach. However, the classification outcomes quantify overlapping distributions of cover types (e.g. 'sagebrush' or 'shrub') between classes. These accuracies are assessed using 'fuzzy' confusion matrices. This enables more information to be preserved through the remote sensing classification process, and reserves more interpretation for the map user than a typical 'hard' classification. Importantly the distributions of cover types are likely most representative of field conditions and thus more useful to land managers making holistic decisions about restoration or fuel management. For the second chapter, I delved deeper into the potentials of new remote sensing and computing platforms to predict biological soil crust cover. The growing field of research on biological soil crust points to potentially significant implications for nutrient and water cycling, in addition to positive effects on native vascular vegetation. However, spatial data are lacking due to remot...

Coverage and Rainfall Response of Biological Soil Crusts Using Multi-Temporal Sentinel-2 Data in a Central European Temperate Dry Acid Grassland

“…With the launch of the multispectral optical satellite mission Sentinel-2 (S2) within the ESA Copernicus earth observation program, global data of high spatial and temporal resolution is freely available, allowing for continuous global monitoring. The mission is designed to be comparable to other predecessor missions such as Landsat 8 [69], and was used to estimate BSC coverage and characteristics within large-scale areas in the northern and southern hemispheres [28,30,53].…”