*= Equal taxpayers
We propose a self-supervised anomaly detection technique, called SeMAnD, to detect geometric anomalies in multimodal geospatial datasets. Geospatial data comprises acquired and derived heterogeneous data modalities that we transform into semantically meaningful image-like tensors to address multimodal data representation, alignment, and fusion challenges. SeMAnD is composed of (i) a simple data augmentation strategy, called RandPolyAugment, capable of generating various augmentations of vector geometries, and (ii) a self-supervised training objective with three components that encourage the learning of multimodal data representations that discriminate to the locals. changes in one modality that are not corroborated by the other modalities. Detecting local defects is crucial for geospatial anomaly detection where even small anomalies (e.g., displaced, incorrectly connected, poorly formed, or missing polygonal vector geometries such as roads, buildings, land cover, etc.) are detrimental to the experience. and the security of users of geospatial services. applications such as mapping, routing, search and recommendation systems. Our empirical study on test suites of different types of real-world geometric geospatial anomalies in 3 diverse geographic regions demonstrates that SeMAnD is capable of detecting real-world defects and outperforms domain-independent anomaly detection strategies by 4.8%. and 19.7%, as measured by anomaly classification. AUC. We also show that model performance increases (i) up to 20.4% as the number of input modalities increases and (ii) up to 22.9% as diversity increases and robustness increases. training data.