StratTune: a spectral adaptive hierarchical fine-tuning transfer approach for red tide detection

Spectral detection technology, which allows for the rapid and non-invasive acquisition of spectral characteristics, is widely employed in red tide monitoring. Due to the weak spectral signals of low-concentration red tides and the susceptibility of the detection process to environmental noise, most existing spectral detection methods operate within concentration ranges higher than the baseline concentration during red tide events. In addition, the limited sample size often fails to capture the spectral response variations across different stages of red tide growth. Regarding the above issues, this paper proposes StratTune, a spectral adaptive hierarchical fine-tuning transfer method for red tide spectral detection. By establishing a novel two-dimensional domain difference evaluation system, combining gradient sensitivity and spectral activation strength, to address small-sample spectral analysis. Our Domain-Aware Fine-Tuning (DAF) leverages feature activation and gradient metrics as dual-domain indicators, enabling synchronized fine-tuning in both spatial and spectral dimensions. The Hierarchical Policy Network (HPN) customizes fine-tuning routes and filtering strategies based on importance scores for each instance. Additionally, we introduce a Multi-Head Feature Adapter (MHFA), utilizing spatial sparsity and high channel similarity, to minimize the domain difference in red tide spectral images for heterogeneous transfer learning. We have developed a portable spectral detection system and constructed a database of seven common red tide species in the southeastern coastal region of China. Experimental results show that StratTune achieves high performance in small-sample (<18) and low-concentration (<1 × 106 cells/L) red tide classification and concentration detection, attaining 98.32% accuracy in species identification and an R2 of 0.9900 in concentration prediction, while reducing trainable parameters by 43.8%.