DescriptionThe WMO report on The Global Climate in 2015-2019 apprises of an acceleration in climate change, which could have a colossal impact over environmental systems (marine environment, freshwater systems, terrestrial ecosystems, soil, forests, etc.). This phenomenon can lead to extensive environmental damage, affecting ecosystems, human health, economies, well-being etc. Diminishing/eliminating (some of) these dire consequences requires social action, which can only be prompted by raising awareness,
locally, in our communities, and globally. To this end, we aim to present a series of machine learning methods for environmental damage and post-disaster assessment from spaceborne imagery. We primarily focus on deforestation and assess damage of the forest area in the Kvarken Region. We select a collection of images of the Kvarken Area from Google Earth Engine, comprising both damaged and intact forest areas. Initially, we perform unsupervised learning methods to automatically delimit forest areas from coastal and urban areas.
Consequently, we train models to classify the damaged forest areas. In the past decades, Convolutional Neural Networks (CNNs) brought about remarkable results for object detection and image segmentation. Several well-known datasets (Imagenet, VOC, COCO) were used to train different CNN architectures. However, the specificity of the satellite images hinders the results obtained using these pre-trained architectures. For this reason, we train and test three CNN architectures on the collected satellite imagery of the Kvarken region to classify the damaged forest areas at pixel level. Moreover, we use QGIS to map the damaged areas. We compare these three architectures and their accuracies. This work illustrates how machine learning methods can be used to support damage assessment in the aftermath of possible disturbances or disaster at regional scale.
|Period||20 Jan 2020 → 22 Jan 2020|
|Event title||2020 Finnish Satellite Workshop and Remote Sensing Days|
|Degree of Recognition||International|