Breakpoint detection through neural nets: A feasibility study

A variety of statistical methods are available to detect sudden changes, or breakpoints, in time series when used as multi-temporal change detection technique. However, these methods are unreliable in the presence of noise. Neural nets might detect breakpoints better. These deep learning models are able to generalize and optimize well, even in the presence of noise. This research tests the feasibility of different neural net architectures to detect breakpoints in generic linear time series. Two relatively simple neural nets are proposed, combined with four different descriptions of breakpoint, and trained on synthetic data. The neural nets are tested on two datasets: On a separate synthetic dataset and on Australian rainuse-efficieny (RUE) time series, a surrogate for dryland ecosystem functioning. Some of the neural nets built performed exceptionally well on synthetic data, outperforming a benchmark statistical method with margin. The direct translation to RUE time series was less successful. The results shows great promise for the use of neural nets in change detection. A generalist change detection approach by use of neural nets is likely not optimal. Current developments in deep learning, as well as choosing the right user-case, show great promise to unlock the full potential of neural nets in time series analysis.