A Chlorophyll-a Concentration Inversion Model Based on Adaptive Fine-Tuning Transfer Learning Method and Self-Attention Mechanism

Chlorophyll-a (Chl-a) is essential for assessing water quality and aquatic ecosystems. Traditional in situ monitoring suffers from sparse stations and limited samples, and machine learning often overfits under such conditions. Transfer learning offers a solution by leveraging pretrained knowledge, but domain discrepancies remain a major challenge. To improve cross-domain Chl-a inversion under few-shot settings, this study proposes GSA-ICPO-Former, which integrates a Transformer architecture, an improved crested porcupine optimizer (ICPO), and a gradient sensitivity-based adaptive (GSA) fine-tuning method. The Transformer captures global spectral dependencies through self-attention; ICPO enhances hyperparameter optimization and prevents overfitting; GSA adjusts model parameters to adapt source-domain knowledge to target-domain features. After ICPO optimization, the model’s $R^{2}$ in the source domain improved from 0.73 to 0.84, and RMSE decreased from 2.39 to 1.86 $\mu$g/L, representing improvements of 15.07% and 22.18%, respectively. In the target domain, compared with ICPO-Former, GSA-ICPO-Former raised $R^{2}$ from 0.54 to 0.79 and reduced RMSE by 32.97%, demonstrating enhanced cross-domain adaptability. Ablation studies confirmed the effectiveness of each optimization component. The model also showed robust performance across different coastal environments, including both nearshore (Tangdao Bay) and open-sea (Rongcheng) regions. It accurately tracked Chl-a distribution with limited samples and showed promise in dynamic algal bloom monitoring. Overall, the proposed model effectively bridges domain gaps and maintains high prediction accuracy in low-sample, cross-regional scenarios, providing a practical tool for satellite-based coastal water quality monitoring.