A Framework for Rapid Active Learning in Resource-Constrained Environmental Sensing Domains

Abstract

Recent advancements in deep convolutional neural networks have enabled accurate, efficient, and intelligent feature learning for a wide variety of classification tasks. However, there remains a research to practice gap that has limited the deployment of these vision systems in socially relevant domains. In particular, the prohibitive cost of manual data annotation for training these classification models, and the unique qualities of naturally observed datasets, such as imbalanced, multispectral input features, can limit the ability to develop robust, stable, and generalizable classification models in resource-constrained settings. In this work, we proposed a novel active learning methodology to enable efficient data labeling in these challenging classification settings, accounting for common dataset constraints of socially focused domains such as small sample sizes, rare positive class samples, and multispectral/thermal input features. We additionally test the efficacy and generalizability of our proposed framework in two distinct domains of ecology and disaster management. We find that the proposed approach significantly reduces the number of manual annotations required to develop stable and accurate classification models, as well as rapidly identifies important, rare positive samples through the developed automatic querying procedure. In turn, we find that this reduction in the manual annotation requirements for classification training translates to significant time-saving measures for human annotators, which can enable more ubiquitous, equitable, and accessible deployment of deep learning-based classification models in important social domains.