Low-Cost IoT Water Quality Classification Using Three Deployable Sensors and Edge-Deployed Machine Learning

We quantified the accuracy cost of restricting aquaculture water quality classification to the three parameters a microcontroller can measure. Gradient Boosting achieves 87.21% accuracy and 0.8664 macro F1 from temperature, pH and turbidity alone, sacrificing 0.1198 F1 — 12.1% relative — against a full 14-parameter laboratory panel, with stability confirmed across ten random splits. A quantized inference artifact occupies 4.37 KB, well within ESP32 constraints. We also report and diagnose a negative soft-sensing result, attributing it to the synthetic con-struction of the dataset rather than to sensor limitations. Three directions follow. First, and most importantly, the model should be validated against field measurements: even fifty readings from a deployed node, labelled against a refer-ence laboratory assay or a published potability standard [20], [21], would convert this from a modelling study into a system evaluation. Second, the same ablation should be repeated on field-collected datasets such as [1], [7] and [8], where phys-ical relationships between parameters are preserved and soft-sensing may well succeed. Third, on-target latency and energy per inference should be measured on ESP32 hardware, which would allow the duty-cycling and battery-life analysis that a solar-powered deployment requires.

1.    M. M. Islam and M. A. Kashem, “Real-time dataset of pond water for fish farming using IoT devices,” Data in Brief, vol. 51, art. 109761, Dec. 2023, doi: 10.1016/j.dib.2023.109761. 2.    B. Siswanto, Y. Dani, D. Morika and B. Mardiyana, “A simple dataset of water quality on aquaponic fish ponds based on an internet of things mea-surement device,” Data in Brief, vol. 48, art. 109248, Jun. 2023, doi: 10.1016/j.dib.2023.109248. 3.    C. N. Udanor, N. I. Ossai, E. O. Nweke, B. O. Ogbuokiri, A. H. Eneh, C. H. Ugwuishiwu, S. O. Aneke, A. O. Ezuwgu, P. O. Ugwoke and A. Christiana, “An internet of things labelled dataset for aquaponics fish pond water quality monitoring system,” Data in Brief, vol. 43, art. 108400, Aug. 2022, doi: 10.1016/j.dib.2022.108400. 4.    V. Veeramsetty, R. Arabelli and T. Bernatin, “Aquaculture — Water Quality Dataset,” Mendeley Data, V1, 2024, doi: 10.17632/y78ty2g293.1. 5.    R. David, J. Duke, A. Jain, V. Janapa Reddi, N. Jeffries, J. Li, N. Kreeger, I. Nappier, M. Natraj, S. Regev, R. Rhodes, T. Wang and P. Warden, “Ten-sorFlow Lite Micro: Embedded machine learning for TinyML systems,” in Proc. Machine Learning and Systems (MLSys), vol. 3, 2021, pp. 800–811. 6.    L. Grinsztajn, E. Oyallon and G. Varoquaux, “Why do tree-based models still outperform deep learning on typical tabular data?,” in Advances in Neural Information Processing Systems 35 (NeurIPS 2022), Datasets and Benchmarks Track, New Orleans, LA, USA, 2022. 7.    R. Baena-Navarro, Y. Carriazo-Regino and F. Torres-Hoyos, “IoT monitor-ing dataset of water quality and tilapia (Oreochromis niloticus) health in aquaculture ponds in Montería, Colombia,” Mendeley Data, V1, 2024, doi: 10.17632/3g2b4sh65m.1. 8.    R. Baena-Navarro, Y. Carriazo-Regino, F. Torres-Hoyos and J. Pinedo-López, “Water quality monitoring dataset for tilapia aquaculture in Mon-tería, Colombia,” Mendeley Data, V1, 2024, doi: 10.17632/dgdr2kfbyt.1. 9.    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), Salt Lake City, UT, USA, 2018, pp. 2704–2713. 10.     J. H. Friedman, “Greedy function approximation: A gradient boosting ma-chine,” The Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, Oct. 2001.