Weakly-Supervised Learning for Earth Observation

Abstract

While satellites stream petabytes of remote sensing data from multiple sensors each year [1], human annotation of such modalities are unable to keep that pace. My presentation highlights the success of recent deep learning methodologies such as self-supervised learning to mitigate the data labeling challenge for Earth observation [2]. Benchmark datasets derived from the Sentinel-1/2 radar/multi-spectral missions enable the training of geospatial foundation models [3,4]. For hyperspectral sensors such as DLR's EnMAP satellite, semi-supervised approaches have been proven successful, too [5]. Aside from the current hype surrounding foundation models, we explore robust training of artificial neural networks exploiting noisy labels. We introduce the concept of weak labels auto-generated from sparse, but high-quality remote sensing data such as airborne LiDAR [7]. We demonstrate use cases for urban climate resilience by quantifying local climate zones in metropolitan areas such as New York City [8,9] [1] https://doi.org/10.1016/j.scitotenv.2023.168584 [2] SSL review: https://doi.org/10.1109/MGRS.2022.3198244 [3] SSL4EO-S12: https://doi.org/10.1109/MGRS.2023.3281651 [4] SoftCon: https://doi.org/10.48550/arXiv.2405.20462 [5] HyperPAWS: https://doi.org/10.1109/IGARSS52108.2023.10282971 [6] AIO2: https://doi.org/10.1109/TGRS.2024.3373908 [7] AutoGeoLabel: https://doi.org/10.1109/BigData52589.2021.9672060 [8] AutoLCZ: https://doi.org/10.48550/arXiv.2405.13993 [9] DeepLCZChange: https://doi.org/10.1109/IGARSS52108.2023.10281573