REMOTE SENSING OF LANDSLIDES IN THE MONTEVERDE, COSTA RICA CLOUD FOREST: A MACHINE LEARNING APPROACH

Landslides, which are often associated with major rain or seismic events, are an important contributor to the biodiversity of montane rainforest ecosystems. By opening gaps in the tree canopy and allowing sunlight to reach the forest floor, they change the temperature and light conditions that provide germination cues to pioneer plants capable of growth only under the high light in canopy gaps. Large landslides are often visible in high resolution satellite images, and the availability of time series of such imagery makes it increasingly feasible to associate landslides with specific weather or seismic events. The goal of our research was to develop a machine learning method for the automatic identification of landslides and other erosional features in high resolution satellite imagery. Using imagery of the cloud forest near the Monteverde Cloud Forest Reserve (MCFR) in Costa Rica from the Planetscope (4-band) and the RapidEye (5-band) satellite constellations, we first generated training and validation sets through the visual identification of surface features. Paired with a supervised machine learning algorithm (random forest), we then combined pixel-based spectral information with texture measures and topographic variables (computed from the imagery and remotely sensed digital elevation models) to classify pixels in the satellite imagery into distinct classes, separating landslides as well as other features. Based on the validation data, the overall accuracy of the classifier was estimated to be 99%. The approach is further validated by ground-truth data collected in July 2019 in the MCFR. The classifier was applied to satellite imagery collected before and after recent major weather events (Hurricanes Nate and Otto) that impacted the MCFR area, and the resulting maps demonstrate the impacts of these storms.