Dissolved Oxygen Estimation in Freshwater Aquaculture Ponds from UAV Multispectral Imagery via Coupled Transfer Learning Strategies Wenxu Lv (1, 2), Peng Cheng (1, 2), Guohong Li (1, 2*), Ruiyin Tang (1, 2*), Yancang Wang (1, 2*), Xuqing Li (1, 2)
1North China Institute of Aerospace Engineering, Langfang 065000, China.
2Aerospace Remote Sensing Information Processing and Application Collaborative Innovation Center of Hebei Province, Langfang 065000, China
Abstract
Dissolved oxygen (DO), as a key indicator of aquatic environmental health, is a critical parameter for maintaining the stability of aquatic ecosystems. Its dynamic variations directly reflect water quality conditions and significantly affect the growth and development of aquatic organisms. In recent years, the integration of UAV-based multispectral imagery with machine learning algorithms has become a research hotspot for accurate estimation of water quality parameters. However, the adaptability and stability of models across different periods remain key challenges limiting their widespread application. This study proposes a transfer learning-based approach for dynamic DO monitoring to enhance the temporal generalization ability of remote sensing estimation models. Multispectral UAV data and in-situ DO measurements acquired in 2023 were used as the training and validation sets, while the 2024 dataset served as an independent test set. Three machine learning algorithms-RF, XGBoost, and LightGBM-were employed to construct baseline estimation models, combined with three transfer learning strategies, namely IR, MMD, and their hybrid integration, to investigate model generalization across time. The results show that: (1) compared with baseline models, IR improved the average coefficient of determination by 12%, while reducing RMSE and MAE by 4% and 4.5%, respectively- MMD improved the average coefficient of determination by 32%, while reducing RMSE and MAE by 11% and 15%, respectively- (2) the hybrid IR-MMD strategy achieved the most significant improvement, with average coefficient of determination increased by 38% and RMSE and MAE reduced by 18% and 21%, respectively. The hybrid transfer learning model based on RF achieved the best performance on the test set, with coefficient of determination of 0.58, RMSE of 3.01 mg/L, and MAE of 2.77 mg/L, substantially outperforming traditional models.
