*= Equal Contributors
We suggest a Self-supervised Anomaly Detection method, known as SeMAnD, to detect geometric anomalies in Multimodal geospatial datasets. Geospatial information includes acquired and derived heterogeneous information modalities that we rework to semantically significant, image-like tensors to deal with the challenges of illustration, alignment, and fusion of multimodal information. SeMAnD is comprised of (i) a easy information augmentation technique, known as RandPolyAugment, able to producing numerous augmentations of vector geometries, and (ii) a self-supervised coaching goal with three parts that incentivize studying representations of multimodal information which are discriminative to native modifications in a single modality which aren’t corroborated by the opposite modalities. Detecting native defects is essential for geospatial anomaly detection the place even small anomalies (for instance, shifted, incorrectly linked, malformed, or lacking polygonal vector geometries like roads, buildings, landcover, and many others.) are detrimental to the expertise and security of customers of geospatial functions like mapping, routing, search, and advice programs. Our empirical research on check units of several types of real-world geometric geospatial anomalies throughout 3 numerous geographical areas demonstrates that SeMAnD is ready to detect real-world defects and outperforms domain-agnostic anomaly detection methods by 4.8-19.7% as measured utilizing anomaly classification AUC. We additionally present that mannequin efficiency will increase (i) as much as 20.4% because the variety of enter modalities will increase and (ii) as much as 22.9% as the range and energy of coaching information augmentations will increase.