This master’s thesis aims to train machine learning models that detect wind-felled forests using low-resolution satellite images. Today, drones and helicopters are often used to localize and estimate damaged forests, a time-consuming and expensive process. Using low-resolution satellite images to find wind-felled forests would be more cost-effective and less time-consuming. During this thesis, the main focus is to investigate how training data from different time periods affects models' performance.
This thesis aims to create models that segment the wind-felled forest to make it easier for forest owners to find it. U-net is used for this thesis because it is one of the most prominent models for image segmentation. The models are trained on data from around the village Malå and from June to September 2023. This area was chosen due to the large damage caused by the storm Hans at the beginning of August 2023. This thesis uses a comparison-based approach, which means that the model analyses images from before and after the storm and compares them. To train the models labeled data is needed. Gathering and labeling data is expensive, so investigating how to most effectively train models is important. This thesis specifically explores how to create training sets that result in robust models on data from different dates. Models are trained on data with different constructions to investigate how performance differs between models trained on one pair of days compared to multiple pairs of days and some tests are also done on normalization and infrared satellite images. The models are evaluated on satellite images from new days and areas to assess their ability to generalize to new scenarios.
The results show that the models trained on multiple pairs of days (one image before and one after the storm) perform better on new days than models trained with data from one pair of days. The result also shows that a training set constructed from multiple days performs differently depending on how it was constructed. The results also showed that using infrared satellite images may increase the performance and that normalization of satellite images seems promising, but more investigation is needed.