A Lightweight Embedded Neural Network for Real‐Time Water Quality Classification in IoT ‐Based Monitoring Systems

ABSTRACT A lightweight neural‐network classifier was developed to assign surface‐water samples to the five quality classes established by Brazil's environmental standard CONAMA 357/2005. The study used a 2018 dataset from SEMASA‐Itajaí comprising 8928 hourly observations of pH, turbidity, apparent color, and local rainfall. After removing outliers, applying weighted k‐nearest‐neighbor imputation and performing min–max normalization, the data were evaluated using a 5‐fold Stratified Cross‐Validation approach to ensure robust generalization. The network, implemented in TensorFlow, consists of two fully connected layers with 32 and 16 rectified linear units, followed by a softmax output layer. Training for 100 epochs with the Adam optimizer (learning rate 0.001) achieved an average validation accuracy of 97.63% (±0.86%), with misclassifications restricted to adjacent regulatory classes. Adding rainfall as an auxiliary feature and fine‐tuning the learning rate further increased accuracy by 0.7 percentage points and reduced training variance.