Patterning emergent marsh vegetation assemblages in coastal Louisiana,USA, with unsupervised artificial neural networks

Abstract

Questions: Are self-organizing maps (SOMs) useful for patterning coastal wetland vegetation communities? Do SOMs provide robust alternatives to traditional classification methods, particularly when underlying species response functions are unknown or difficult to approximate, or when a need exists to continuously classify new samples obtained under ongoing long-term ecosystem monitoring programs as they become available? Location: Coastal Louisiana, USA. Methods: A SOM was trained from in-situ observations of 559 vegetation species relative cover data from 2,526 samples collected over 8 years at 343 locations across coastal Louisiana. Hierarchical cluster analysis was applied to the SOM output to delineate vegetation community types, and indicator species analysis was conducted.Salinity and flood duration were compared across the delineated community types. Results:The SOM patterned the 2,526 training samples into 260 output neurons, which were further clustered into eleven community types. Clear gradients in salinity and flood duration existed among the community types, and geographic zonation of the communities was evident across the landscape. At some locations assemblages were temporally stable; at other locations, they varied considerably. Samples not used in training the network were effectively projected onto the SOM and assigned to one of the delineated community types. Conclusions:The SOM was effective in delineating plant communities in the region that were qualitatively similar to those obtained in previous investigations. Being robust to skewed distributions and the presence of outliers, SOMs provide an alternative to traditional distribution-based statistical approaches. Their ability to efficiently classify new data into existing community types makes their use an ideal approach to classifying samples obtained from ongoing, long-term ecological monitoring programs.

Wetland Biomass and Productivity in Coastal Louisiana: Base Line Data (1976–2015) and Knowledge Gaps for the Development of Spatially Explicit Models for Ecosystem Restoration and Rehabilitation Initiatives

“…Despite the spatial extension and dominance of Sagittaria lancifolia [69], the total number of studies is small (n = 21) and lacking during some periods (e.g., 1990-1994; 2002-2004; 2010-2013) in the last two decades. This lack of consistent field or greenhouse AGB studies is surprising because S. lancifolia is a dominant species generally included as the most representative of the freshwater category in modeling studies [32,70,71].…”

“…This statistical modeling approach (e.g., clustering, detrended correspondence analysis, multinomial logistic regression, neural network) shows how robust field-based measurements of plant structural attributes (diversity, coverage) and hydroperiod (e.g., tidal amplitude, flood duration) can provide good estimates to characterize plant species assemblages along complex hydrological and stressor (e.g., salinity) gradients. Furthermore, this study shows the tremendous utility of long term acquired in situ environmental data, as presently collected by the Coastwide Reference Monitoring System (CRMS) program since 2006 (https://www.lacoast.gov/crms2/home.aspx), and with more consistent data acquisition since 2008 in 390 sites across Louisiana; this environmental network [29] is advancing the predictive capability of models to assess vegetation communities' structural properties at a large scale in the near future (e.g., [71]).…”

“…The current probabilistic models proposed by Snedden and Steyer [70] and the LAVeg Model [32] represent examples of this incremental development in vegetation models that explicitly incorporate in situ measurements that reflect local and regional ecological processes and environmental variable interactions (e.g., salinity, tidal amplitude, percentage of time flooded) that are relevant to management decisions in coastal Louisiana. More recently, Snedden [71] used artificial neural networks to evaluate marsh vegetation assemblages across coastal Louisiana and found that this approach was comparable to statistics-based classification schemes. Given the dominance of spatially skewed vegetation and outliers, this approach could potentially be linked to high-resolution spatial information to delineate zonation and functional attributes, such as biomass and productivity.…”

Redefining floristic zones in the Korean Peninsula using high‐resolution georeferenced specimen data and self‐organizing maps

“…Therefore, approaches to delineate floristic zones using these data can be developed reliably and accurately. In To analyze accumulated species distribution data, artificial neural networks (ANNs) are increasingly used as an alternative to traditional statistics to analyze multidimensional data (Chon, 2011;Cottrell, Olteanu, Rossi, & Villa-Vialaneix, 2018;Snedden, 2019).…”

Wetland Classification Based on a New Efficient Generative Adversarial Network and Jilin-1 Satellite Image

“…Notably no labeled samples are required, the unsupervised methods can be easily used for wetland classification. Typical unsupervised classification methods include k-means cluster [14] and self-organizing maps (SOM) [15]. Unfortunately, it is hard to explore the relationship between clusters and class labels with too little a priori knowledge.…”