Record Details
Spatial and Temporal Dynamics of Broad-scale Predictive Models : Influences of Scale
ScholarsArchive at Oregon State University
| Field | Value |
|---|---|
| Title | Spatial and Temporal Dynamics of Broad-scale Predictive Models : Influences of Scale |
| Names |
Hooper, Samuel
(creator) Kennedy, Robert E. (advisor) |
| Date Issued | 2017-01-05 (iso8601) |
| Note | Graduation date: 2017 |
| Abstract | Developing accurate predictive distribution models requires adequately representing relevant spatial and temporal scales, as these scales are ultimately reflective of the relationships between distributions and influential environmental conditions. In this research, we considered both spatial and temporal scale and the influence each has on predicting broad-scale distributions of two disparate but related phenomena: land cover and bird distributions. Employing machine-learning algorithms, we first developed land cover time series datasets covering all of California, Oregon, and Washington with a model that simultaneously reflects local-scale heterogeneity and broad-scale homogeneity. We then used these and other land cover time series datasets to investigate the effects of temporal resolution on species distribution models. In the second chapter, we focused on the importance of accurately representing the spatial scale of relationships between predictors and a response variable for broad-scale predictive models. Using both a novel machine-learning algorithm and a novel predictor dataset, we developed dense time series forest canopy cover (FCC) and impervious surface cover (ISC) datasets at a 30-meter spatial resolution for all of California, Oregon, and Washington. To develop both datasets, we employed a spatial ensemble modeling method using a population of locally defined and spatially overlapping decision trees, making it both appropriate at continental-scales and sensitive to local variation in predictor-response relationships. Our predictor variables were products of LandTrendr, a tool for developing time series images and derivatives from the Landsat archive. To develop the most accurate time series of FCC and ISC, we first tested two model parameters, sample size and estimator size. Using the best-performing configuration of each, we then compared our models with locally defined estimators to bagged decision trees, the most comparable model with globally defined estimators. Using the best-performing models and LandTrendr imagery, we developed yearly FCC and ISC maps, spanning 1990-2012. To test the temporal extensibility of our models, we compared our predicted 2011 maps to 2011 maps from the National Land Cover Database. We found that model performance for both FCC and ISC decreased with increasing estimator size and that models with locally defined estimators outperformed bagged decision trees. We also found that our models performed well when extending learned predictor-response relationships to predict 2011 FCC and ISC distributions. These results, in concert with several novel byproducts of the models that we developed, demonstrate that representing local-scale spatial relationships is critical to producing accurate broad-scale distribution models. In the third chapter, we investigated the influence of temporal scale on an avian species distribution model (SDM) by comparing models developed with different temporal resolutions of land cover predictor data. We expressed temporal resolution as the time interval between representative dates of predictor data. By adjusting the temporal resolution of land cover predictors but still using yearly response data, models at coarser resolutions were potentially trained with bird observations that did not match the land cover contemporaneous with observations. We tested different temporal resolutions for five bird species to capture how habitat preference may affect SDM response to temporal resolution. Our results confirmed that temporal resolution has a noticeable effect on SDMs, but in unexpected ways. For three of the five species, the best-performing models were produced at the coarsest temporal resolution, with other species showing better performance at moderate resolutions. We also found only subtle evidence supporting the idea that habitat preference influenced SDM response to temporal resolution. These results indicate that 1) temporal dynamics of avian species-environment relationships are highly complex and specific to individual species and 2) that representation of temporal scale has a prominent effect on model outcome. |
| Genre | Thesis/Dissertation |
| Access Condition | http://creativecommons.org/licenses/by-nc-nd/3.0/us/ |
| Topic | machine learning |
| Identifier | http://hdl.handle.net/1957/60148 |
