There are now more interferograms being generated from global satellite radar datasets than can be assessed by hand. The reliable, automatic detection of true displacement from these data is therefore critical, both for monitoring deformation related to geohazards and understanding solid earth processes. We discuss improvements to an unsupervised, event agnostic method for automatically detecting deformation in unwrapped interferograms. We use an anomaly detection framework that recognises any deformation as “anomalies” by learning the ‘typical’ spatio-temporal pattern of atmospheric and other noise in sequences of interferograms. Here, we present developments to our prototype model, ALADDIn (Autoencoder-LSTM based Anomaly Detector of Deformation in InSAR) using (1) a self-attention training technique to exploit redundancy in interferogram networks to capture the temporal structure of signals and (2) the addition of synthetic data for training. We evaluate the impact of these developments using two geophysical scenarios: (1) the detection of the same M_w 5.7 earthquake used to test our original model (20.03.2019, south-west Turkey), (2) the persistent uplift of Domoyu volcano (17.05.2017 to 14.12.2018, Argentina). We make a quantitative evaluation of the performance of our method using synthetic test data and find that for peak displacements exceeding a few cm and of length-scale greater than a few hundred metres, overall detection accuracy is 80 to 90%.