Author(s): Sarina Adeli | Bahram Salehi | Masoud Mahidanpari | Lindi J. Quackenbush
Author Affiliation(s): Department of Environmental Resources Engineering, College of Environmental Science and Forestry, State University of New York, Syracuse, New York, USA | Department of Environmental Resources Engineering, College of Environmental Science and Forestry, State University of New York, Syracuse, New York, USA | C-CORE and Department of Electrical Engineering, Memorial University of Newfoundland, St. John’s, NL A1B 3X5, Canada | Department of Environmental Resources Engineering, College of Environmental Science and Forestry, State University of New York, Syracuse, New York, USA
Session: WETLANDS AND WATER BODIES MAPPING AND MONITORING I
Frequent challenges in the preservation of wetlands, a land cover type that is crucial for climate regulation, biodiversity preservation, and carbon sequestration arise from the lack of reliable wetlands distribution maps on a national scale. The rate of loss in wetlands due to natural and man-made stressors is reported about 1.4 percent from 2004 to 2009 in the USA. As such, the management of wetlands mandates a cost-effective, robust, and semi-automated technique that can ensure their efficient conservation. However, there are some challenges in monitoring wetlands due to their vast extent and inaccessibility. The availability of free-of-charge and high-resolution earth observation data in along with the emerging of cloud computing platforms’ ability in handling big data has made up-to-date wetland mapping on a large scale a real possibility. In this presentation, we will present our state-wide wetland method and results for Minnesota using multi-source remote sensing data utilizing Google Earth Engine cloud computing. In this study, we adopted an object-oriented Random Forest (RF) classification scheme by leveraging dual-polarimetry C-band Sentinel-1, multi-spectral Sentinel-2, and dual polarimetry L-band ALOS-PALSAR. To assure wall-to-wall coverage over Minnesota state we generated yearly and seasonal composites. Once the unwanted artifacts (cloud-cover) were removed, we left with about 5,000 imageries. The spatial clusters for segmentation were extracted using the Simple Non-Iterative Clustering (SNIC) algorithm. We also included textural Gray-Level Co-occurrence Matrix (GLCM) to obtain textural features. Spectral information of vegetation, water, and soil indices was extracted along with ratio and span for the radar imagery. To be consistent with the National Wetland Inventory of Minnesota, we used the Cowardin classification scheme for defining the reference data. Our initial results demonstrate that, depending on the availability of field data for each ecozone, overall accuracies changed from more than 80.3% to 88.1%. The variable importance analysis suggests that Sentinel-2 spectral features are dominant in terms of their capability for wetland delineation. Overall, although there are some limitations in the current computational ability of GEE on a large scale, the initial results of this study confirmed the unique ability of available EO data along with the computational ability of GEE for producing detailed wetlands extents on a state-wide scale.
Sarina Adeli