A LandTrendr Optimization Approach for Improved Mapping of Annual Land Cover
Abstract
Earth-observation (EO) data provide the foundation for much mapping and monitoring of environmental conditions at broad scales, but the increasing richness and depth of EO data archives require corresponding improvements in algorithms to tap those observations. Because they distill and stabilize long-term spectral records from EO data (primarily the Landsat family of sensors), the LandTrendr (LT) algorithms have been used both for tracking disturbance and land cover in the United States and elsewhere. Additionally, since being added to the Google Earth Engine (GEE) platform, the LT algorithms are increasingly being applied for forest disturbance mapping around the world. Like many change detection approaches, the LT algorithms require that users specify the spectral index and algorithm fitting parameters to ensure good separation of change from no change. This level of control is attractive to many users, but also requires significant expertise. Additionally, the choice of ideal spectral index and parameters may vary within ecosystem types within the same region. Here, we report on new approaches to optimize the LT algorithms for eventual landcover mapping in Cambodia as part of a NASA-SERVIR Science Team project in the Lower Mekong Basin. The optimization workflow leverages GEEs computational capacities to run LT across spectral and parameter domains at a sample of points distributed across the study domain, and uses a Pareto-like optimization strategy to balance capture of temporal features with avoidance of over-fitting. We tested the approach with 5000 points and several hundred parameter runs, selecting the best LT drivers to apply to the entire country. Using a web-based interpretation tool for validation, we found that the optimization strategy successfully identifies appropriate index and parameter values. When carried through to mapping to the country of Cambodia, the approach shows promise as a means of flexibly mapping change without the need for expert intervention.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMGC54B..03K