Meta-Learning For Few-Shot Time Series Classification

State-of-the-art deep learning methods require large quantities of labeled data pairs for high performance. While satellite data is now available in abundance, ground truth labels remain scarce. Moreover, ground truth labels are distributed unevenly around the globe; high-resource regions (e.g. US, Europe) have many more labels than low-resource regions (e.g. Africa, parts of Asia). We hypothesize that a great deal of information can be shared across tasks involving satellite imagery, due to shared image acquisition and similarities in landscapes and objects worldwide. Meta-learning is a machine learning discipline in which models are explicitly trained to adapt to new tasks with few labeled data pairs per task. In this work, we use model agnostic meta-learning (MAML) to train a neural network on tasks from high-resource regions to improve land cover classification accuracies in low-resource regions. We observe that a meta-learned neural network outperforms both the same network trained from scratch and the network pretrained on high-resource data and finetuned on low-resource data.