Classifying Coastal Wetland Vegetation Communities with Unsupervised Artificial Neural Networks Active
Wetlands are often classified by their vegetation, which can help scientists track how these landscapes change over time. USGS turns to unsupervised artificial neural networks to help guide this classification process.
The Science Issue and Relevance: Vegetation community composition is often chosen as the basis for the classification of wetland ecosystems because it can reflect the integration of multiple ecological processes acting on site or landscape scale more effectively than any other factor or set of factors. Patterns of co-occurring plant species can be used to infer spatial and temporal gradients in salinity, inundation, and disturbance regimes. Thus, classifying communities in coastal wetlands based on vegetation can serve to describe many aspects of ecology across coastal landscapes, and to track the trajectory of those landscapes through time. Existing vegetation community classes obtained with multivariate statistical approaches are problematic in that when classification is repeated after the addition of new samples, cluster memberships of previously classified samples can change. This limitation precludes well-established existing classification systems from classifying current and new samples, and impedes our ability to analyze long-term trends in vegetation community structure at local and global scales.
Methodology for Addressing the Issue: The self-organizing map (SOM), which provides a topology-preserving non-linear projection of multivariate data onto a 2-dimensional space (map), is a relatively new artificial neural network approach that is very robust to these limitations. Each neuron (cell) in the map represents a unique “virtual” community assemblage, and neighboring neurons share similar assemblages. In addition to providing a stable classification system which can be used to classify new data as they become available, the nonlinear, nonparametric nature of SOMs makes them appropriate for analysis of nearly any kind of species composition data, irrespective of their statistical properties.
In this project, an SOM is trained from vegetation species cover data obtained at nearly 4000 marsh sites across coastal Louisiana in late summer 2007. Subdividing the trained SOM into regions that signify distinct community types is accomplished by submitting the species weight vectors of each map unit of the SOM to cluster analysis.
Future Steps: Once a suitable network is trained from the 2007 survey data, species cover samples obtained from the Coastwide Reference Monitoring System (CRMS) and Coastal Wetlands Planning, Protection and Restoration Act (CWPPRA) monitoring will be projected onto the SOM to examine how vegetation community composition responds to restoration activities, climate variability, and disturbance. Eventually, predictive models linking vegetation community types to hydrologic variables such as inundation, salinity, and tidal amplitude will be developed.
- Overview
Wetlands are often classified by their vegetation, which can help scientists track how these landscapes change over time. USGS turns to unsupervised artificial neural networks to help guide this classification process.
Sources/Usage: Some content may have restrictions. View Media DetailsNeuron (cell) from a self-organizing map (SOM) The Science Issue and Relevance: Vegetation community composition is often chosen as the basis for the classification of wetland ecosystems because it can reflect the integration of multiple ecological processes acting on site or landscape scale more effectively than any other factor or set of factors. Patterns of co-occurring plant species can be used to infer spatial and temporal gradients in salinity, inundation, and disturbance regimes. Thus, classifying communities in coastal wetlands based on vegetation can serve to describe many aspects of ecology across coastal landscapes, and to track the trajectory of those landscapes through time. Existing vegetation community classes obtained with multivariate statistical approaches are problematic in that when classification is repeated after the addition of new samples, cluster memberships of previously classified samples can change. This limitation precludes well-established existing classification systems from classifying current and new samples, and impedes our ability to analyze long-term trends in vegetation community structure at local and global scales.
Methodology for Addressing the Issue: The self-organizing map (SOM), which provides a topology-preserving non-linear projection of multivariate data onto a 2-dimensional space (map), is a relatively new artificial neural network approach that is very robust to these limitations. Each neuron (cell) in the map represents a unique “virtual” community assemblage, and neighboring neurons share similar assemblages. In addition to providing a stable classification system which can be used to classify new data as they become available, the nonlinear, nonparametric nature of SOMs makes them appropriate for analysis of nearly any kind of species composition data, irrespective of their statistical properties.
In this project, an SOM is trained from vegetation species cover data obtained at nearly 4000 marsh sites across coastal Louisiana in late summer 2007. Subdividing the trained SOM into regions that signify distinct community types is accomplished by submitting the species weight vectors of each map unit of the SOM to cluster analysis.
Sources/Usage: Some content may have restrictions. View Media DetailsSelf-organizing map (SOM), which provides a topology-preserving non-linear projection of multivariate data onto a 2-dimensional space Future Steps: Once a suitable network is trained from the 2007 survey data, species cover samples obtained from the Coastwide Reference Monitoring System (CRMS) and Coastal Wetlands Planning, Protection and Restoration Act (CWPPRA) monitoring will be projected onto the SOM to examine how vegetation community composition responds to restoration activities, climate variability, and disturbance. Eventually, predictive models linking vegetation community types to hydrologic variables such as inundation, salinity, and tidal amplitude will be developed.