A few-shot ResNet transfer learning method for robust chlorophyll-a retrieval in optically-variable waters under distribution shift and data scarcity

Existing algorithms for chlorophyll-a (Chla) retrieval encounter two key challenges: poor cross-domain generalization, and the scarcity of paired field Chla measurements with corresponding spectrum. Water optical properties vary strongly with environmental conditions, causing pronounced distribution shift across different spatial and temporal domains. However, most models assume that spectra in new domains follow the same distribution as the training data, which leads to sharp performance degradation under distribution shifts. Furthermore, the absence of effective sample selection strategy increases data demand for model development, exacerbating conflicts with limited field observations. To address these issues, we proposed FSResTL-Chla, a few-shot ResNet transfer learning method for Chla retrieval. We constructed a universal Chla-spectrum model with a pre-trained ResNet under different optical water types. To ensure the representativeness and informativeness of few-shot samples, a self-adaptive representative sample selection strategy was proposed, coupled with fine-tuning to adjust the universal model to distribution shift. FSResTL-Chla showed favorable performance across multiple optical water types using 7–12 samples, with an average coefficient of determination ( R 2 ) of 0.68. Despite using only 18% of all observations, FSResTL-Chla achieved performance ( R 2 = 0.52–0.79) comparable to or exceeding that of the best traditional models ( R 2 = 0.49–0.80), which commonly required 70% − 80% of the observations for model development. This demonstrates that FSResTL-Chla maintains strong spatiotempral generalization, whereas traditional models often fail under changing conditions. By improving cross-domain generalization while reducing dependence on sample size, FSResTL-Chla provides a valuable tool for environmental monitoring and management.