Revisiting Image Quality Effects on Encoding Robustness for Remote Sensing Models

Image quality factors associated with training and test datasets significantly impact the performance of a computer vision algorithm. This is particularly important for object detection and classification tasks with remotely sensed imagery. Optimizations in training methods as well as network architecture can sometimes overcome challenges in scale invariance, though these methods aren’t robust to all image variances. Standoff factors typical of remote-sensing data are significantly different from hand-held imagery, by which most computer vision models are trained. Hence, robust algorithms for detection and classification tasks using satellite imagery continue to challenge the state of the art. Arguably, typical feature extraction mechanisms in traditional deep learning algorithms break down for this domain. New or additional training sets from remote sensing sources are often required to properly train classifiers from overhead imagery, with transfer learning as a secondary alternative for enhancing models at the scale in which satellite imagery is typically taken. We revisit the challenges in object classification and detection using remote sensing in this study, and account for a variety of image quality factors that may impact detector performance. New baselines may need to be considered to enhance the performance of models in the remote sensing domain. Finally, an emphasis on further characterization and understanding of image quality factors from overhead imagery will result in proper optimizations in state-of-the-art algorithms suited for remote sensing. The resulting research should help bridge the performance gap in detection and classification tasks between hand-held and remotely sensed imagery.