Three decades of Landsat-derived spring surface water dynamics in an agricultural wetland mosaic; Implications for migratory shorebirds.

Satellite measurements of surface water offer promise for understanding wetland habitat availability at broad spatial and temporal scales; reliable habitat is crucial for the persistence of migratory shorebirds that depend on wetland networks. We analyzed water extent dynamics within wetland habitats at a globally important shorebird stopover site for a 1983-2015 Landsat time series, and evaluated the effect of climate on water extent. A range of methods can detect open water from imagery, including supervised classification approaches and thresholds for spectral bands and indices. Thresholds provide a time advantage; however, there is no universally superior index, nor single best threshold for all instances. We used random forest to model the presence or absence of water from >6200 reference pixels, and derived an optimal water probability threshold for our study area using receiver operating characteristic curves. An optimized mid-infrared (1.5-1.7 μm) threshold identified open water in the Sacramento Valley of California at 30-m resolution with an average of 90% producer's accuracy, comparable to approaches that require more intensive user input. SLC-off Landsat 7 imagery was integrated by applying a customized interpolation that mapped water in missing data gaps with 99% user's accuracy. On average we detected open water on 26000 ha (3% of the study area) in early April at the peak of shorebird migration, while water extent increased five-fold after the migration rush. Over the last three decades, late March water extent declined by ~1300 ha per year, primarily due to changes in the extent and timing of agricultural flood-irrigation. Water within shorebird habitats was significantly associated with an index of water availability at the peak of migration. Our approach can be used to optimize thresholds for time series analysis and near-real-time mapping in other regions, and requires only marginally more time than generating a confusion matrix.