The increasing availability of inexpensive high-resolution satellite imagery data has made remote sensing an ideal frontier for applied deep learning research. However, such data is highly unstructured, with factors such as time, location, weather conditions, and sensor specifications creating discrepancies across sources, making it difficult to extract meaningful insights with conventional computer vision methods.
In order to take advantage of such satellite data, which is available in large quantities but highly unstructured, we propose the use of domain adaptation techniques such as domain-adversarial training and cycle consistency to generate analogous data across different domains and conditions, consequently learning universally generalizable representations of satellite image data.
General representations would allow for more effective use of unstructured data, making possible large-scale employment of deep learning methods for downstream tasks such as monitoring environmental conditions, detecting natural disasters, and estimating agricultural output.
Computer Vision, Remote Sensing, Domain Adaptation, GANs
Basic understanding of Computer Vision, GANs, and Domain Adaptation
- ragaDani (Participant)