Accelerating ecosystem monitoring through computer vision with deep metric learning

Abstract

Significant progress has been made in the application of deep learning models to ecosystem monitoring. Deep learning has opened up new opportunities in the interpretation of ecological data, such as detecting and identifying objects in images and acoustic monitoring analysis. However, these have created new challenges, such as the need to further improve the efficiency of data processing due to the increasing volume of data, the need for more advanced feature extraction methods due to the complexity of data characteristics, and limitations of available annotated data. In this study, we focused on deep metric learning as a new application for environmental observation data to overcome these challenges. The extraction of features such as patterns and changes from large and complex environmental observation data using a deep metric learning approach may provide new opportunities for monitoring ecosystems experiencing unprecedented loads from climate change and human activities. While these methods demonstrate the potential of deep metric learning for flora and fauna and various datasets, they also suggest challenges to overcome, such as the need for more valid training datasets, diverse data collection, training time proportional to the data volume, and the identification of unknown classes. We expect that deep metric learning will be a powerful tool for various ecosystem monitoring systems, from remote sensing of wide-area data to ecological data obtained through field surveys.