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)

1) North China Institute of Aerospace Engineering, Langfang 065000, China.
2) Aerospace Remote Sensing Information Processing and Application Collaborative Innovation Center of Hebei Province, Langfang 065000, China

Abstract

Dissolved oxygen is a key indicator of aquatic environmental health. Its dynamic changes directly reflect water quality conditions. In recent years, the integration of UAV multispectral imagery and machine learning algorithms has emerged as a research hotspot for accurately estimating water quality parameters. However, the temporal adaptability and stability of such models remain major challenges limiting their broader application. This study proposes a transfer learning based approach for dynamic monitoring of dissolved oxygen, aiming to enhance the temporal generalization ability of remote sensing estimation models. Data in 2023 were used as the training and validation sets, while data from 2024 were employed as the test set. Three machine learning algorithms, RF, XGBoost, and LightGBM, were used to construct estimation models. Three transfer learning strategies, including instance reweighting (IR), maximum mean discrepancy (MMD), and a hybrid method combining both, were applied to improve the models^ temporal generalization performance. The results show that first, compared with baseline models, instance reweighting increased the coefficient of determination by an average of 12 percent, and reduced RMSE and MAE by 4 percent and 4.5 percent, respectively. MMD improved the coefficient of determination by 32 percent, with average reductions of 11 percent in RMSE and 15 percent in MAE. Second, the hybrid transfer learning strategy combining IR and MMD yielded the most significant improvements, increasing the coefficient of determination by 38 percent, and decreasing RMSE and MAE by 18 percent and 21 percent, respectively. The hybrid RF based transfer learning model achieved the best performance on the test set, with a coefficient of determination of 0.58, RMSE of 3.01 mg/L, and MAE of 2.77 mg/L. These findings demonstrate that transfer learning strategies can effectively mitigate the impact of temporal variability on remote sensing based water quality estimation.

Keywords: Transfer learning- temporal generalization- UAV multispectral imagery- freshwater aquaculture-dissolved oxygen

Topic: Topic B: Applications of Remote Sensing