Machine Learning Analysis for Radiation Monitoring Data in the Fukushima Region

Abstract:

Nine years have passed since the radionuclide release occurred at the Fukushima Dai-ichi Nuclear Power Plant (FDNPP). During the accident, radionuclides were deposited on soil and plants through wet and dry deposition. Radiocesium (134Cs and 137Cs) is currently the main contaminant in the environment. Over the past nine years, the region around FDNPP has experienced remarkable recovery. In addition, many studies have reported that the decrease in the air dose rates – including the reduction associated with radiocesium transport in the environment – has been accelerated compared to the physical decay. An extensive monitoring program has been established after the accident and still continues to this day. Quantifying the changes in air dose rates is important for planning the residents’ return and future decontamination efforts. Recently, machine learning (ML) techniques represent one of the most rapidly expanding areas of science and technology in the past ten years. ML enables us to discover hidden patterns and identify key factors in large datasets, and also to estimate unobservable parameters or project them into the future. The objective of this study is to explore the ML application for the radiation monitoring data in the Fukushima region. In particular, we aim to characterize the spatial variability of environmental decay of dose rates across the region. We test a variety of ML methods, including supervised and unsupervised learning. We use different types of datasets, including car survey data, fixed-point observations, and monitoring posts.