Sinkholes exhibit precursory deformation patterns. Such deformation patterns can be studied using InSAR time-series analysis over constantly coherent scatterrers (CCS). In the past we identified Heaviside and Breakpoint changes as two important forms of anomalous behaviour. It is challenging to efficiently detect and classify these sudden step and sudden velocity changes in deformation time series, especially in the presence of tens of thousands CCS. To address this challenge, we propose to classify these forms of anomalous behaviour with a deep learning-based supervised time series classification. In this study, we used a two-layered Bidirectional Long Short Term Memory (LSTM) classification model for this purpose. The classified deformation classes were analyzed as well in the context of scattering mechanisms. We implemented this model on a sinkhole affected region spanning <inline-formula><tex-math notation="LaTeX"><?TeX ${\sim }63\times 44$?></tex-math></inline-formula> km<inline-formula><tex-math notation="LaTeX"><?TeX $^{2}$?></tex-math></inline-formula> in Ireland, using 104 Sentinel-1 A SAR images acquired between 2015 and 2018. Our results show that the CCS with a linear trend can be correctly classified with a maximum accuracy of <inline-formula><tex-math notation="LaTeX"><?TeX ${\sim }99 \%$?></tex-math></inline-formula> whereas for the CCS categorized as anomalous Heaviside and Breakpoint changes the accuracy drops to a maximum of 62%. Multi-threshold-based filtering of samples increased the classification accuracy by as much as 50%. We conclude that the method that we propose is effective in detecting anomalous deformation changes. Future research should investigate how it can be applied to other hazard-related detection and classification problems.