A Framework for Integrating Dense Landsat Time Series and Community-based Monitoring Data to Characterize Forest Changes in the Tropics

Abstract:
We present an integrated monitoring framework that uses all available Landsat data and a continuous stream of community-based monitoring data to characterize forest changes in an Afromontane forest system in southern Ethiopia. To demonstrate this framework, we trained random forest models using a suite a spectral-temporal variables derived from the Landsat time series data as covariates and forest disturbance reports from local forest rangers as training data. We classified deforestation and degradation with out-of-bag (OOB) class accuracies of 74% and 69%, respectively, with an overall OOB accuracy rate of 71%. While these models did not show improvements for the deforestation class compared to previous studies in our site, our approach succeeded in mapping diffuse forest degradation with greater certainty than previously achieved. Forest change classification accuracies improved as more ground-based observations from local rangers became available, demonstrating the utility of such a continuous data stream. The random forest models also revealed that short-wave infrared bands, or indices based on these bands, consistently provided the most important information for distinguishing deforestation, degradation and no-change classes. Given the continuous acquisition of satellite-based and in situ observations, this framework provides a flexible approach to monitoring forest changes which could be used to ingest observations from newly launched sensors such as Sentinel-1 and 2 combined with continued Landsat-7 and 8 acquisitions. Furthermore, with modifications to the parameters measured by local forest rangers, this framework could be scaled up to monitor a broader range of relevant forest variables.
Abstract type: Posters