Search

Dwelling information is essential for humanitarian emergency response during or in the aftermath of disasters, especially in temporary settlement areas hosting forcibly displaced people. To map dwellings, the integration of very high-resolution remotely sensed imagery in computer vision models plays a key role. However, state-of-the-art deep learning models have two known downsides: 1) lack of generalization across space and time under changing scenes and object characteristics, and 2) extensive demand for annotated samples for training and validation. Both could pose a critical challenge during an emergency. To bypass this problem, this study deals with unsupervised domain adaptation for instance segmentation using a single-stage instance segmentation model, namely segmenting objects by location (SOLO). The goal is to adapt a SOLO model trained on a labeled source domain to detect dwellings in an unlabeled target domain. In this context, we study three domain adaptation techniques based on adversarial learning, domain discrepancy, and domain alignment mapping. We also propose domain similarity at different levels to understand its implication on domain adaptation. Experiments are conducted on very high-resolution satellite images obtained from four temporary settlement areas located in different countries and exhibiting various spatial characteristics. Analysis results show that in most source–target combinations unsupervised domain adaptation improves the performance by a large margin even surpassing a model trained with supervised learning. There is also an observed performance deviation among implemented strategies and different source–target dataset combinations. From the in-depth analysis of domain similarity at the image, object, and deep feature space levels, the former is more correlated with unsupervised domain adaptation performance.

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 ∼ 35 % of the target dataset.

Earth-observation-based mapping plays a critical role in humanitarian responses by providing timely and accurate information in inaccessible areas, or in situations where frequent updates and monitoring are required, such as in internally displaced population (IDP)/refugee settlements. Manual information extraction pipelines are slow and resource inefficient. Advances in deep learning, especially convolutional neural networks (CNNs), are providing state-of-the-art possibilities for automation in information extraction. This study investigates a deep convolutional neural network-based Mask R-CNN model for dwelling extractions in IDP/refugee settlements. The study uses a time series of very high-resolution satellite images from WorldView-2 and WorldView-3. The model was trained with transfer learning through domain adaptation from nonremote sensing tasks. The capability of a model trained on historical images to detect dwelling features on completely unseen newly obtained images through temporal transfer was investigated. The results show that transfer learning provides better performance than training the model from scratch, with an MIoU range of 4.5 to 15.3%, and a range of 18.6 to 25.6% for the overall quality of the extracted dwellings, which varied on the bases of the source of the pretrained weight and the input image. Once it was trained on historical images, the model achieved 62.9, 89.3, and 77% for the object-based mean intersection over union (MIoU), completeness, and quality metrics, respectively, on completely unseen images.