Accelerating declines and geographic shifts of salt marshes and species of conservation concern in the northeastern United States have prompted land managers to revise management objectives and monitoring methods to inform decision making. The FWS National Wildlife Refuge System (NWRS) Interior Region 1 (IR1) manages salt marshes across 26 NWRs from Virginia to Maine, and recently updated salt marsh management objectives focusing on four primary vegetation-based classes and five non-vegetated classes. Previous field methods offered insights into vegetation community dynamics and responses to management actions. This project builds upon these insights and uses satellite imagery and machine learning to provide a fine-scale (2-meter resolution) and cost-effective method for monitoring the extent and distribution of key salt marsh vegetation communities across NWRs.
In September 2020, 2,676 vegetation field training data samples were collected on seven NWRs in IR1 using survey-grade equipment. Training data for non-vegetation classes were digitized visually using satellite imagery from DigitalGlobe’s WorldView-2 and -3 satellite images. Images were primarily collected during the growing season (June–September) from 2018 through 2020. Imagery was preprocessed using ENVI’s radiometric and atmospheric corrections to achieve surface reflectance values. Thirty-nine input variables at 2-meter resolution were identified: 30 vegetation indices, five soil indices, one water index, a digital elevation model (DEM), and a composite measure of all spectral bands (2 Principle Component Analysis (PCA) axes). Preliminary results identify several important variables: elevation, normalized difference mud index, PCA axis, and structure insensitive pigment index. Overall classification error of the best model was 8.4% (91.6% model accuracy). Water, development, bare sediment, other vegetation, and mudflat classes had the lowest error (0.3% to 3.3%), while mid-marsh low marsh classes had moderate error (11% to 23%), and brackish marsh and upper marsh classes had the highest error (50% to 67%), which were commonly misclassified as mid-marsh (27% to 38%). Exploration of class probabilities shows many pixels are a mixture of vegetation classes (not homogeneous) to aid in interpreting results.
This approach provides the foundation for a more accurate, fine-scale, cost-effective monitoring for salt marsh vegetation on NWRs. Predictive maps can help assess management effectiveness of restoration projects when field methods are expensive, infrequent, and have high error. Incorporating these methods into a monitoring framework could provide timely spatial salt marsh information to managers to assess management progress, help understand changes in migratory bird habitat, and provide more evidence of salt marsh instability (unvegetated to vegetated marsh ratio) to guide management interventions.
Maps derived using 2-meter imagery from WorldView-2 and -3 providing high spatial information of salt marshes for the US Fish and Wildlife Survey (FWS) National Wildlife Refuge System (NWRS) Interior Region 1 (IR1). Panel A: Twenty-six NWRs with salt marsh ecosystems. Panel B: WorldView-2 imagery of Prime Hook NWR with training data (circles) of each class (colors). Panel C: Area of Prime Hook NWR with example vegetation training data locations (circles). Panel D: Example prediction map (2-meter) of two correctly predicted classes (mid-marsh and brackish marsh) with training data locations (circles). Panel E: Example probability of each pixel (2-meter) classified as mid-marsh class (yellow = higher probability, purple = lower probability).