Hierarchical Machine Learning Approach for Real-Time Imaging: Classification and Object Detection

Current All-Sky imagers measure airglow, providing a valuable tool for studying atmospheric gravity waves due to its relationship with airglow intensity. Targeted airglow emissions include atomic oxygen, sodium, and hydroxyl compounds. Gravity waves cause these emissions to fluctuate in intensity, allowing an indirect way to study these waves to further understand and improve atmospheric dynamics and modeling. However, issues such as vignetting occurring at the outer edges of a traditional imager can render captured images unusable for accurate observation. The All-Sky Imaging science team from the University of Houston’s Undergraduate Student Instrumentation Project (USIP) aimed to collect airglow data by real-time processing for a dual optical array. This proof of concept allows for improved imaging by a primary camera, with a near 180-degree field of view utilizing a layered convoluted neural network trained on decades of all-sky image data. The system begins with a ResNet50 architecture to give binary classification of airglow presence, optimizing object detection by YOLO with creating bounding boxes for prominent lines of airglow activity. A secondary camera with a narrower 80 degree field of view then automatically reorienting for direct target of detected regions for high-resolution imagery. This development allows for a new approach to automatic image evaluation for optical and atmospheric applications.