The realm of space technology is in a perpetual state of expansion. Our collective endeavour to build Machine Learning (ML) and Artificial Intelligence (AI) models suitable for space applications is steadily gaining momentum, driven by the potential such models have to transform space research. Key among their capabilities is the efficient processing of vast volumes of data collected by satellites, covering topics such as aerial mapping, weather prediction, and deforestation. Rapid events like natural disasters, however, have remained a challenge due to the limitations of current satellite data processing techniques.
A recent breakthrough has seen researchers successfully train ML models in space for the first time, an approach that bypasses the need for Earth-based training. These models, trained using a method called few-shot or active learning, focus on identifying the most critical features necessary for training. This process effectively reduces data dimensions, leading to a significant increase in the speed and effectiveness of the model. As a component of the broader Computer Vision model category, this innovative model zeroes in on detecting the presence of cloud cover, thereby forming a vital part of a more complex classification model.
This model operates in two primary stages. The initial stage involves capturing images and training on Earth, while the secondary stage, which takes place on the satellite, applies a binary classification technique to determine cloud cover presence. Although such training typically requires numerous iterations, or epochs, our compact model completed the process in an astounding one and a half seconds. The model’s flexibility, reflected in its ability to automatically adapt to various data forms, is another noteworthy feature.
Despite these promising advancements, the pursuit of further development continues, as researchers aim to create models capable of managing complex datasets, including hyperspectral satellite images. The model’s performance metrics, such as recall, precision, and F1 score, are impressively high, underlining its potential for broader application.
As the potential of AI in space research becomes increasingly evident, opportunities are opening up not just for near-Earth studies, but also for deep space exploration. Researchers are ready to explore the unknown, armed with the power of AI, signifying a pivotal step towards our understanding of the universe.
Check out the Paper. All Credit For This Research Goes To the Researchers on This Project.