Universal Robust Domain Adaptation for Remote Sensing Image Classification using Deep Optimized Transfer Learning

Computer Science and Engineering

The detection of algal blooms in lakes and reservoirs via remote sensing presents a significant environmental challenge, impacting aquatic ecosystems and human health. While conventional algorithms relying on remote sensing reflectance have shown effectiveness in certain contexts, achieving high accuracy across multiple lakes remains a challenge, particularly with single- threshold-based approaches. This study investigates the performance of various machine learning (ML) algorithms for pinpointing algal bloom locations using Sentinel-2 images in Chinese eutrophic inland lakes. Through comprehensive testing of four ML models - random forest (RF), extreme gradient boosting, artificial neural network, and support vector machine - in lakes Taihu, Chaohu, and Dianchi, alongside index-based methods such as the floating algae index, this research provides insights into their accuracy, stability, and robustness. Results indicate that the RF model exhibits superior performance compared to other ML models, maintaining an overall accuracy above 0.90 across various lakes. Notably, even when trained on data from a single lake, the RF model achieves a commendable accuracy of 0.88 for other lakes. In summary, this comparative analysis underscores the promising potential of ML techniques in enhancing the detection of algal blooms in diverse remote sensing scenarios.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.