The Sacramento-San Joaquin Delta community level classification maps derived from WorldView-2 multispectral data

Since 2004, airborne hyperspectral imagery has been acquired annually over the Sacramento - San Joaquin Delta (Delta) in northern California to map submerged and floating invasive species. However, between 2009 and 2013, there was a gap in data collection due to lack of funding. The objective of this project was to fill that gap to the extent possible using WorldView-2 images acquired over the Delta during this gap period. There were no images available in 2009, however, more than 70% of the Legal Delta was mapped each year from 2010 to 2014 to fill the gap in the time series of existing data. Digital Globe provided the University of California, Davis, with all images acquired over the Legal Delta in this time period. These top-of-atmosphere (TOA) reflectance images were then further georegistered and resampled to 2x2m pixels and prepared for analysis. Further, each flightline was processed using multiple spectral mapping methods such as spectral angle mapper, spectral mixture analysis, spectral indexes, and continuum removal over water and cellulose absorption bands. The outputs of these transformations were used as inputs to a Random Forests classifier. Concurrent with image acquisition, field data were user-interpreted across the study area for training and validation of the classification products. The data were divided into test and training polygons. These polygons were overlaid on the transformed files and pixel data were extracted corresponding to the polygons. The training data were used to train the Random Forests classifier to identify 7 classes (water, submerged aquatic vegetation (SAV), emergent marsh, soil, non-photosynthetic vegetation (NPV), floating aquatic vegetation (FAV), and riparian vegetation). The classifier was validated quantitatively using the test data at both pixel and polygon level using overall accuracy and kappa metrics. A unique classifier was created for each individual date of imagery. The corresponding classifier was then applied to all images from that date and class maps were produced. Mosaics of these class maps are published in this dataset.