Neutral models as a way to evaluate the Sea Level Affecting Marshes Model (SLAMM)

A commonly used landscape model to simulate wetland change – the Sea Level Affecting Marshes Model (SLAMM) – has rarely been explicitly assessed for its prediction accuracy. Here, we evaluated this model using recently proposed neutral models – including the random constraint match model (RCM) and growing cluster model (GrC), which consider the initial landscape conditions instead of starting with a blank or randomized initial map as traditional neutral models do. Thus, the SLAMM's performance, due to processes accounted for in the model, could be more accurately assessed. RCM allocates change randomly in space, while in the GrC, change allocation is prioritized at the locations with pairs of to-be-increased land type and to-be-reduced land type adjacent to each other. The metrics we applied to evaluate the SLAMM vs. the neutral models accounted for five main components in map comparison: (1) reference change simulated correctly as change (hits), (2) reference persistence simulated correctly as persistence (correct rejections), (3) reference change simulated incorrectly as change to the wrong category (wrong hits), (4) reference change simulated incorrectly as persistence (misses), and (5) reference persistence simulated incorrectly as change (false alarms). These methods improved the way that we currently evaluate land change models, where we either do not compare to a neutral model, or the neutral model does not have the same boundary conditions and constraints as the assessed dynamics models. The results showed that the SLAMM could simulate wetland change more accurately compared to the GrC and RCM at a 10-year time step for the lower Pascagoula River basin, Mississippi, with higher hits and correct rejections, and lower misses and false alarms. The magnitude of simulated changes using the SLAMM was 46% of reference changes. The number of wrong hits for the SLAMM was also lower than those for the neutral models after combining some land or water types into broader categories. After the aggregation, the SLAMM performance improved substantially. How the errors of this relatively short-term simulation propagate into longer-term predictions requires further investigation. This study also showed the importance of implementing elevation data with high vertical accuracy, and conducting local calibration when we apply the SLAMM.