Spatial Data Mining in Aerial Object Detection Datasets for Finding Co-Locations and Anomalies

The remote sensing community has developed algorithms to extract diverse geographic information from electro-optical and radar imagery, and study them along with data from other weather sensors such as wind, temperature and humidity. Simultaneously, the computer vision community has adopted machine learning algorithms for detection problems with promising success. Many such algorithms utilize labeled datasets to learn object signatures for remote sensing. Meanwhile, research in the data mining domain also has evolved spatial algorithms for discovering association and outliers in the geographic data. In a first, the current work introduces techniques developed in data mining domain to the computer vision community utilizing object detection datasets for applications in remote sensing. Specifically, this work applies the concepts of rule mining to geo-registered aerial object detection datasets.Aerial object detection datasets were originally intended for training a classifier. However, this work innovates the use of ground-truth labels with geo-location to identify co-located classes. The paper establishes a method to adapt the aerial object detection dataset to mining friendly data. A theoretical analysis of the co-location frequency and size of search area is presented. The search area size for a set of objects is heuristically estimated. The work utilizes state-of-the-art data mining methods and uses the Support and Lift metrics to study the effects of varying count of object instances per class. In addition to finding prominent co-locations and anomalies, this work demonstrates how composite objects can be detected, which have varying spatial configurations of their constituents. The developed method is tested on publicly available XView (60-classes) and FAIR1M (37-classes) datasets.