Many-to-many: Domain adaptation for water quality prediction

Predicting water quality is crucial for sustainable water management. To mitigate data scarcity for specific water quality targets, domain adaptation methods have been employed, adjusting a model to perform in a related domain and leveraging learned knowledge to bridge domain differences. However, these methods often fall short by overfitting certain domain-specific patterns, overlooking consistent water quality patterns in multi-water domains. Despite regional variations, these Consistent patterns show fundamental commonalities and can be observed across monitoring sites, stemming from their widespread and interconnected nature. Addressing these limitations, we introduce the Many-to-Many Domain Adaptation framework (M2M) for prediction to bridge the gap between multi-source domains and multi-target domains, aligning shared insights with the distinct profiles of individual monitoring sites while considering their geographical interconnections. M2M adeptly addresses the formidable challenge of concurrently deciphering and integrating multifaceted patterns across an array of source and target domains, while also navigating the intricate regional heterogeneity intrinsic to the water quality of different sites. The M2M includes a domain pattern fusion module for consistent pattern extraction and numerical scale maintenance from source domains, a domain pattern sharing module for sharing pattern extraction from target domains, and an M2M learning method to ensure the training of these modules. Extensive experiments conducted on 120 diverse monitoring stations demonstrate that M2M markedly enhances the accuracy of water quality predictions using various time series encoders. Code available at https://github.com/biya0105/M2M.