Estimating changes and trends in ecosystem extent with dense time-series satellite remote sensing

Quantifying trends in ecosystem extent is essential to understanding the status of ecosystems. Estimates of ecosystem loss are widely used to track progress toward conservation targets, monitor deforestation, and identify ecosystems undergoing rapid change. Satellite remote sensing has become an important source of information for estimating these variables. Despite regular acquisition of satellite data, many studies of change in ecosystem extent use only static snapshots, which ignores considerable amounts of data. This approach limits the ability to explicitly estimate trend uncertainty and significance. Assessing the accuracy of multiple snapshots also requires time-series reference data which is often very costly and sometimes impossible to obtain. We devised a method of estimating trends in ecosystem extent that uses all available Landsat satellite imagery. We used a dense time series of classified maps that explicitly accounted for covariates that affect extent estimates (e.g., time, cloud cover, and seasonality). We applied this approach to the Hukaung Valley Wildlife Sanctuary, Myanmar, where rapid deforestation is greatly affecting the lowland rainforest. We applied a generalized additive mixed model to estimate forest extent from more than 650 Landsat image classifications (1999-2018). Forest extent declined significantly at a rate of 0.274%/year (SE = 0.078). Forest extent declined from 91.70% (SE = 0.02) of the study area in 1999 to 86.52% (SE = 0.02) in 2018. Compared with the snapshot method, our approach improved estimated trends of ecosystem loss by allowing significance testing with confidence intervals and incorporation of nonlinear relationships. Our method can be used to identify significant trends over time, reduces the need for extensive reference data through time, and provides quantitative estimates of uncertainty.