Document Type
Paper
Start Date
6-10-2023 9:50 AM
End Date
6-10-2023 10:20 AM
Abstract
The monitoring of open water and early identification of oil spills in the Alaska Arctic has become increasingly critical due to the rise in oil and gas exploration and shipping activities, facilitated by the increasing number of ice-free days resulting from global warming. This escalating risk of oil spills is further compounded by potential accidents in offshore operations, illicit oil discharges, and knowledge gaps in Arctic coastlines, rapidly changing due to rising seas, permafrost melting, and coastal erosion. To address these pressing challenges, we propose a deep learning model based on MobileNet neural networks to detect oil spills in remotely sensed images. Compared to traditional pattern recognition methods, the proposed model can learn from examples to map input new data into the design and automatically optimize the training objective without designing rules and specifying critical parameters to solve the inference task. The experiments demonstrated we were able to obtain an overall accuracy of 0.93 with our proposed methods.
DOI
10.5703/1288284317671
Fine Tuning MobileNet Neural Networks for Oil Spill Detection
The monitoring of open water and early identification of oil spills in the Alaska Arctic has become increasingly critical due to the rise in oil and gas exploration and shipping activities, facilitated by the increasing number of ice-free days resulting from global warming. This escalating risk of oil spills is further compounded by potential accidents in offshore operations, illicit oil discharges, and knowledge gaps in Arctic coastlines, rapidly changing due to rising seas, permafrost melting, and coastal erosion. To address these pressing challenges, we propose a deep learning model based on MobileNet neural networks to detect oil spills in remotely sensed images. Compared to traditional pattern recognition methods, the proposed model can learn from examples to map input new data into the design and automatically optimize the training objective without designing rules and specifying critical parameters to solve the inference task. The experiments demonstrated we were able to obtain an overall accuracy of 0.93 with our proposed methods.