[Objective] Hyperspectral remote sensing technology offers a new approach for non-contact, real-time sensing of water quality parameters. Existing research has shortcomings in developing dedicated sensing equipment and constructing multi-parameter collaborative inversion algorithms. This study aims to systematically explore the spectral response characteristics of five key water quality parameters—chemical oxygen demand (COD), total suspended solids (TSS), total phosphorus (TP), total nitrogen (TN), and ammonia nitrogen (NH₃-N)—based on a self-developed hyperspectral water sensing instrument, and to establish an adaptive hyperspectral sensing algorithm optimization framework for multiple water quality parameters. This framework aims to overcome the adaptability limitations of traditional single algorithms in complex water environments and provide technical support for achieving 24/7 continuous online water quality monitoring. [Methods] First, continuous spectral data and corresponding measured values of water quality parameters from a large number of water samples were collected using the self-developed hyperspectral water sensing instrument. Through spectral analysis, the absorption, reflectance characteristics, and differential patterns of COD, TSS, TP, TN, and NH₃-N within the visible to near-infrared spectral range were revealed, and the sensitive characteristic bands for each parameter were selected accordingly. On this basis, a multi-level, multi-type inversion model system was constructed. At the empirical model level, single-band regression, band ratio, and normalized difference index were employed to establish statistical relationships between spectral features and water quality parameters. At the machine learning level, BP neural network, random forest (RF), and XGBoost were introduced to fully exploit the nonlinear mapping relationships within the hyperspectral data. Furthermore, an adaptive algorithm optimization framework was proposed. By comparing the inversion accuracy and stability of various models across different water quality parameters and concentration ranges, this framework automatically matched the optimal inversion algorithm for each parameter, thereby achieving the optimal configuration for multi-parameter collaborative sensing. An independent dataset partitioning strategy was adopted for model training and validation, with the coefficient of determination (R²) and mean absolute percentage error (MAPE) as the core evaluation indicators to ensure the objectivity and reliability of the assessment results. [Results] The constructed adaptive algorithm optimization framework achieved excellent performance in the inversion of all five water quality parameters. In the training and validation datasets for the optimal algorithms, the R² values of the corresponding optimal models for each parameter ranged from 0.88 to 0.99, exhibiting extremely high fitting accuracy and generalization capability. From a practical accuracy perspective, 100% (COD), 91% (TSS), 90% (TP), 90% (TN), and 95% (NH₃-N) of total samples had a MAPE below 30%, fully meeting the practical requirements for water quality monitoring. Specifically, the inversion accuracy for the COD parameter was the highest, with all samples meeting the accuracy threshold, indicating that hyperspectral data possessed exceptional characterization capability for organic pollutant concentrations. NH₃-N followed closely, with 95% of samples meeting the accuracy requirement, reflecting significant spectral response characteristics of ammonia nitrogen in specific bands. The compliance rates for TSS, TP, and TN all exceeded 90%, verifying the universality and robustness of the framework in the inversion of different types of water quality parameters. Compared with traditional single-model methods, the adaptive optimization strategy significantly improved the overall accuracy and stability of multi-parameter collaborative inversion, effectively overcoming the adaptability deficiencies of single algorithms under complex water conditions. [Conclusion] This study proposes and validates an adaptive hyperspectral sensing algorithm optimization framework for multi-parameter water quality monitoring based on a self-developed hyperspectral water sensing instrument. The core innovations of this optimization framework lies in the following: first, it breaks through the conventional paradigm of “one algorithm fits all parameters” in traditional water quality hyperspectral inversion by implementing a multi-model competitive optimization mechanism, achieving automatic matching of the optimal algorithm for each parameter. Second, it integrates empirical models and machine learning models into a unified optimization system, balancing the model interpretability with nonlinear fitting capability. Third, the research findings demonstrate strong engineering application prospects and can directly support the construction of 24/7 continuous online water quality monitoring systems.