Landcover

Overview

Several landcover models exist that span the Cascadia region and are widely used. However, there is no single landcover product with a rich taxonomy of landcover classes that is seamless across the entire Cascadia region, high resolution, available annually over the historical period from 1984 to present, and dynamically updated each year. Though some existing models share some of these qualities, the lack of a model that has them all hinders progress towards a deep understanding of the current status and recent trends in landcover across our region.

To develop a new landcover model for the region that meets all the above criteria is an enormous challenge. Perhaps the most difficult task is to gather the required field observations of landcover needed to train a new model across this large region. Our approach to this problem was to train a new model through transfer learning (reviewed in Wang et al. 2023), where the information contained in existing landcover models from across the region is used as an input into a machine learning classifier that generates a new prediction based on a potentially different set of predictors. In essence, this approach identifies areas of model agreement across the various existing landcover products and attempts to extend that consensus understanding across the region and across years of input data. The result is a seamless model of annual landcover predictions that spans the period from 1984-present, updated each year automatically in the TerrAdapt cloud-based workflow. We used the limited amount of field observations of landcover available in public repositories and government databases to validate the new landcover model.

In the following sections, we describe the landcover taxonomy, training data, model training methods, and validation in more detail.