Detecting Aquaculture with Deep Learning in a Low-Data Setting
Keywords: remote sensing, image segmentation, image classification, contrastive learning, attention pooling, representation learning, convolutional neural networks
TL;DR: Deep learning models are proposed to detect aquaculture from satellite imagery in the Amazon basin with as few as 300 labelled examples using percentile data, attention pooling, and contrastive pretraining.
Abstract: Aquaculture is growing rapidly in the Amazon basin and detailed spatial information is needed to understand the trade-offs between food production, economic development, and environmental impacts. Large open-source datasets of medium resolution satellite imagery offer the potential for mapping a variety of infrastructure, including aquaculture ponds. However, there are many challenges utilizing this data, including few labelled examples, class imbalance, and spatial bias. We find previous rule-based methods for mapping aquaculture perform poorly in the Amazon. By incorporating temporal information through percentile data, we show deep learning models can outperform previous methods by as much as 15% with as few as 300 labelled examples. Further, generalization to unseen regions can be improved by incorporating segmentation information through masked pooling and using contrastive pretraining to harness large quantities of unlabelled data.
Submission Number: 6
Loading