Environment-Aware Deep Learning for Adaptive Dissolved Oxygen Sensing: Resolving the Information Efficiency Paradox in Aquatic Monitoring

Abstract

Continuous high-frequency monitoring of dissolved oxygen (DO) faces a fundamental conflict between excessive energy consumption and substantial data redundancy, whereas conventional fixed-frequency sampling risks frequency aliasing and the omission of acute transient ecological events. To resolve this dilemma, a knowledge-guided, seasonally adaptive dynamic sampling strategy coupled with bidirectional long short-term memory (Bi-LSTM) reconstruction is proposed. Validated via long-term datasets from two hydrodynamically distinct Yangtze River stations, this framework yields three principal findings. First, summer precipitation acts as the primary catalyst for acute hypoxic risks (DO < 5.0 mg/L), whereas winter rainfall predominantly drives reaeration, dictating the nonstationary triggers for dynamic sampling. Second, optimal data decimation exhibits pronounced spatial heterogeneity governed by site-specific environmental buffering capacities; high-fidelity reconstruction requires merely 22.99% data retention at the hydrodynamically stable Nanjinguan mainstream. Third, a counterintuitive information efficiency paradox is mathematically proven at the highly volatile Chenglingji lake outlet: the dynamic strategy (utilizing 50.23% of data) achieves a Kling-Gupta efficiency (KGE) of 0.969, vastly outperforming the uncompressed 100% full-data baseline (KGE = 0.797). This confirms that environmentally triggered sparse sampling intrinsically functions as an adaptive nonlinear filter, actively stripping stochastic environmental noise to prevent algorithmic overfitting. Ultimately, integrating physical priors into data acquisition resolves the energy-fidelity trade-off, establishing a highly resilient, low-power architectural foundation for autonomous basin-scale aquatic monitoring networks under escalating climatic extremes.