Weakly-Supervised Learning for Earth Observation

Kurzfassung

While satellites stream petabytes of remote sensing data from multiple sensors each year, human annotations of such modalities are unable to keep that pace. My presentation highlights the success of recent deep learning methodologies to mitigate the data labeling challenge for Earth observation. Datasets derived from radar, multi-spectral, and hyper-spectral missions enable the training of geospatial foundation models where an encoder compresses semantic information into feature vectors. Neural compression opens an opportunity to benchmark the state-of-the-art for data representations useful for a wide variety of remote sensing applications. Aside from the current excitement surrounding foundation models, we explore robust training of artificial neural networks with noisy labels. We introduce the concept of weak labels auto-generated from sparse, but high-quality remote sensing data such as airborne LiDAR. We demonstrate use cases for urban climate resilience by quantifying local climate zones in metropolitan areas such as New York City, and we tap into how contrastive learning assists the classification of Arctic Sea Ice from radar altimetry measurements.