Answers are grounded in this paper and cite their sources.
Traditional machine learning models used for coastal water quality remote sensing inversion require a substantial amount of identically distributed data, however, the scarcity of measured data often leads to significant small-sample problems. While transfer learning is an effective method to address these issues, traditional transfer learning methods face limitations in areas such as source domain sample weight initialization, negative transfer suppression, and weight decay control. This paper proposes an enhanced Two-stage TrAdaBoost.R2 algorithm, which improves algorithm performance through three key improvements: 1) The source domain sample weights are initialized based on the similarity between source and target domain samples using a logistic regression (LR) model; 2) A removal strategy based on the initial source domain sample weight percentiles is introduced to suppress negative transfer by removing source domain data with low similarity to the target domain samples; 3) A source domain sample weight update rule with error threshold control is designed to ensure that the weights of source domain samples are not overly reduced when the error is too small, thus balancing knowledge transfer effectiveness with model stability. This study uses the inversion of dissolved oxygen (DO) concentration in the coastal waters of Tangshan as an example, with the Yellow River Estuary serving as the source domain. The effectiveness of the method is confirmed by integrating Sentinel-2 multispectral data with a random forest model as the base learner. The results demonstrate that the improved transfer learning algorithm achieves an R2 of 0.81 and an RMSE of 0.54 mg/L on the target domain test set, representing improvements of 35% and 17.4% compared to the traditional random forest model and the baseline transfer model, respectively. Application to a second dataset from the Yellow Sea coast (Qingdao–Nantong) resulted in R2 = 0.80 and RMSE = 0.43 mg/L on the Nantong test set, representing 27.0% and 12.7% gains over RF and baseline transfer learning, respectively. The improved algorithm effectively alleviates negative transfer and enhances the robustness of the model.
No takes yet. Share an insight, caveat, or question.
Li et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: