XFMNet: Decoding Cross-Site and Nonstationary Water Patterns via Stepwise Multimodal Fusion for Long-Term Water Quality Forecasting

TL;DR

XFMNet is a novel deep learning framework that combines multimodal data and adaptive decomposition to improve long-term water quality forecasting across multiple sites, effectively handling nonstationarity and spatial-temporal complexities.

Contribution

The paper introduces XFMNet, a stepwise multimodal fusion network that integrates remote sensing imagery with water quality data using adaptive techniques for enhanced long-term prediction.

Findings

XFMNet outperforms existing methods on real-world datasets.

The adaptive decomposition improves handling of nonstationary signals.

Cross-attention fusion enhances integration of spatial and temporal features.

Abstract

Long-term time-series forecasting is critical for environmental monitoring, yet water quality prediction remains challenging due to complex periodicity, nonstationarity, and abrupt fluctuations induced by ecological factors. These challenges are further amplified in multi-site scenarios that require simultaneous modeling of temporal and spatial dynamics. To tackle this, we introduce XFMNet, a stepwise multimodal fusion network that integrates remote sensing precipitation imagery to provide spatial and environmental context in river networks. XFMNet first aligns temporal resolutions between water quality series and remote sensing inputs via adaptive downsampling, followed by locally adaptive decomposition to disentangle trend and cycle components. A cross-attention gated fusion module dynamically integrates temporal patterns with spatial and ecological cues, enhancing robustness to nonstationarity and site-specific anomalies. Through progressive and recursive fusion, XFMNet captures both long-term trends and short-term fluctuations. Extensive experiments on real-world datasets demonstrate substantial improvements over state-of-the-art baselines, highlighting the effectiveness of XFMNet for spatially distributed time series prediction.