Forest disturbance detection using deep learning approaches

Fire, wind, drought, and insect outbreaks are causing rapid forest decline worldwide. In recent years, the number of forest losses due to such disturbance events has reportedly increased in the temperate and southern boreal regions. However, there is a lack of large-scale and long-term observational records of different disruption types that would link forest stress and mortality to a specific cause of disturbance. Therefore, it is difficult to understand the temporal development of the various forest disturbances and to attribute these signs with certainty to climate change. The increase in quantity and quality of remote sensing data at high spatial and temporal resolutions, along with advances in machine learning for environmental applications, hold great promise for distinguishing between these different disturbance types.

Our conceptual plan for implementing the research is centered on using Sentinel 2 reflectance data and constructing different deep-learning models to identify disturbance types over a case-study region in North America. We focus on large wind throw and bark beetle outbreaks in the construction of a comprehensive data set. The relevance of different features for distinguishing between these two disturbance types is evaluated by comparing spatial and temporal patterns, as well as the relative importance of different reflectance bands.