Due to climate change, the number of storms that cause damage to forests has been increasing over the past years. Trees knocked over by wind throw are a loss of timber if not harvested in time, therefore the damaged area has to be identified as quickly as possible. Current state of the art forest damage detection methods are mainly based on change detection using a before/after comparison as well as manual damage assessment using satellite and aerial images, which takes a great amount of time and effort. In this thesis two custom deep learning models based on the U‐Net architecture are trained on satellite and aerial images to predict storm damage in forests. The study area is located in the state of Bavaria in Germany and covers 160 km^2 of forest, fields and villages. PlanetScope satellite images of PlanetLabs Inc. with a resolution of 3 × 3 m are rapidly available after storm events and are used for first damage assessment. Aerial ortho images with a resolution of 0.2 × 0.2 m are used to get a more accurate prediction if necessary. The achieved accuracy of the models are comparable to current state of the art detection methods when used on similar test data, but is considerably lower when used on a different data set. Methods to improve the transferability are discussed and implemented using transfer learning. The models are integrated into ArcGIS Pro for damage prediction on large scale and subsequent processing steps of the results.
«
Due to climate change, the number of storms that cause damage to forests has been increasing over the past years. Trees knocked over by wind throw are a loss of timber if not harvested in time, therefore the damaged area has to be identified as quickly as possible. Current state of the art forest damage detection methods are mainly based on change detection using a before/after comparison as well as manual damage assessment using satellite and aerial images, which takes a great amount of time an...
»