Spatially transferable dwelling extraction from Multi-Sensor imagery in IDP/Refugee Settlements: A meta-Learning approach

Dwelling information is very important for various applications in humanitarian emergency response. For this, Earth observation is crucial to have spatially explicit and temporally frequent observations. Coupled with advances in computer vision, especially with the proliferation of state-of-the-art deep learning models, are providing a new opportunity for automatic information retrieval from remotely sensed imagery. Despite their proven performance, they have two known limitations, viz, the requirement of intensive data for training and lack of universal generalization under changing scene characteristics and respective data distributions. To tackle this problem, the current study has investigated the relevance of a meta-learning approach for the creation of a spatially transferable optimal model for dwelling extraction in IDP/refugee settlements. The study followed a Model Agnostic meta-Learning (MAML) strategy with newly designed and tested variates with weighted loss gradient update plus self-supervision in the adaptation phase to the target locations using a few samples. The approach is tested using multi-sensor, multi-temporal satellite imagery from eight IDP settlements. Furthermore, a thorough investigation is undertaken on task-specific transfers and their association with deep-embedded feature space and image structural similarity. Results indicate that for some target sites, task-specific transfers perform better than MAML approaches. When MAML is trained with a weighted loss gradient update, it yielded better performance. The best performance (MIoU 0.623 and an F-1 score of 76.7 %) was achieved when MAML is aided with self-supervision using pseudo-labels from unlabelled target data. In all experimental setups, though increasing adaptation samples contribute to positive transfer, the marginal contribution from additional samples is decreasing and stagnates when the adaptation sample size reaches similar to 35 % of the target dataset.