| Abstract |
- The use of Landsat data has historically been constrained to spectral and spatial information derived from a carefully selected image or set of images. However, free and open access to Landsat imagery combined with advances in data storage and computing are revolutionizing how the Landsat temporal domain is used to map and monitor land surface properties and land cover change. Since the opening of the USGS archives in 2008, many different time series analysis approaches have been developed without a unified framework for characterizing information extracted from dense time series of Landsat imagery.
In Chapter 1, we define Spectral-Temporal Features (STFs) as discrete or continuous features derived from time series of remotely sensed observations. Like spectral indices, STFs represent a transformation of the original image data and can provide new information about land surface properties and other biophysical parameters. STFs offer a number of improvements over conventional spectral or spatial inputs, including seamless coverage over large extents, more consistent and stable feature sets for classification through time, and new information on both spectral and temporal variability in reflectance that can be related to biophysical parameters.
To demonstrate how STFs can be applied in practice, we present a series of case studies spanning a range of geographic locations within different ecosystem types, and study objectives. These case studies illustrate relationships between different STFs and various biophysical parameters and yield insight into the specific ecological metrics that can be discovered and characterized with the spectral-temporal domain.
With the release of collection-style Landsat products and continued advances in pre-processing algorithms, as well as availability of tiled Analysis Ready Data and improved access to cloud- and cluster-based computing resources such as Google Earth Engine, the Australian Data Cube, and Sentinel Hub, time series approaches are becoming increasingly prevalent. We argue that STFs provide new information on both spectral and temporal variability in reflectance in different ecosystems that can be related to biophysical parameters. Thus, there is a critical need to continue to review and standardize the discussion and application of STFs for locally-accurate mapping and monitoring of forested ecosystem dynamics.
In Chapter 2, we test the utility of primary STFs derived from time series of all available Landsat TM/ETM+ observations, including Global Surface Water (GSW) features, for discriminating among wetlands at different categorical resolutions within the National Wetlands Inventory (NWI) classification taxonomy. We examine two key types of primary STFs, 1) reflectance STFs, which characterize reflectance values of spectral indices used, and 2) day of year (DOY) STFs, which quantify the timing of their associated reflectance STFs. As an exploratory measure, we also abstract and evaluate spatial-temporal climate features, such as the per-pixel annual maximum of daily maximum temperature, to yield insight into potential drivers of wetland characterization. Using an NWI reference dataset from two Oregon ecoregions in distinctly different eco-hydrological climate zones, we test classification agreement and examine relative performance of different classification inputs across ecoregions and wetland categories. We test an array of classifications that use consistent training and testing datasets, but vary the features and feature-sets used as model inputs. Beyond classification performance, we also explore categorical-level agreement and the importances of different features for differentiation within different wetland categories. Our aim is not to estimate the accuracy of the reference NWI map or monitor wetland change over time and space, but rather to build a framework for multi-level wetland classification across climate gradients.
We found that STFs feature-sets consistently produced high overall accuracies and were able to accurately delineate wetland habitats across climate gradients and wetland categorical resolutions even further when combined with other features. Additionally, accuracies decreased with increasing categorical resolution in both energy-and water-limited ecosystems. Evaluation of individual feature importance for distinguishing between different wetland habitats showed that different features are more important for different climate gradients and categorical resolutions. However, GSW Occurrence was consistently valuable for both ecoregions across all categorical resolutions, exemplifying the value of utilizing the GSW dataset for wetland classification. Further, although not all types of features were found to be important in overall classification, in quantifying correlations between individual features and individual wetland habitat classification probabilities, we found that all feature types were had and strong positive and negative correlations with individual habitats. This indicates the importance of using the various features as inputs for wetland classification.
In Chapter 3, we use all available Landsat imagery from 1985 - 2017 to explore how Pacific Northwest wetland ecosystems are changing over time in different climate zones and at varying categorical resolutions. Additionally, we investigate the long term changes in abstracted Landsat spectral-temporal features that are closely associated with different aspects of wetland hydro-ecological processes. We found that our annual classification model built from Landsat spectral-temporal features, climate-temporal features, and ancillary datasets performs well in showing change in wetland habitat. Individual STFs also display distinct changes in intra-annual wetland dynamics in the context of wetland land use change. In terms of long term wetland change, Willamette Valley wetlands are trending toward more non-vegetated wetlands, fewer vegetated wetlands, and extreme annual-conditions with the lower extrema occurring earlier in the year. In addition to other drivers, this change may be attributed to increased precipitation and increased temperature. In contrast North Basin wetlands are trending towards more vegetated wetlands, fewer non-vegetated wetlands, and extreme annual-conditions with the lower extrema occurring earlier in the year, except for max TCW which is trending towards later annual occurrence. Timing and persistence are key for wetland habitats and this study begins the work to examine change in both occurrence of wetland habitat type and timing of key hydrologic and phenological features and ecosystem drivers.
|
|---|