Portable Multispectral Fluorometer with Embedded Machine Learning for Chlorophyll-a Estimation in Acetone-Extracted Samples

Chlorophyll-a (Chl-a) is commonly used as an indicator of phytoplankton biomass and water quality, but its determination usually depends on laboratory-based methods and specialized equipment. In this work, a portable multispectral fluorometer was developed and evaluated for Chl-a estimation in 90% acetone extracts prepared from water samples. The system integrates a nominal 420–425 nm fluorescence excitation source, 90° detection geometry, two multispectral sensors (AS7341 and AS7343), an ESP32-S3 microcontroller, and an Android application for device control and result visualization. A dataset of 76 samples was used for model development, including regression-model benchmarking, model selection, and internal testing. A Random Forest model selected for embedded implementation was exported to C++ for on-device inference. Validation using 15 independent samples yielded R2 = 0.979, root mean square error (RMSE) = 35.41 µg/L, mean absolute error (MAE) = 28.91 µg/L, and mean absolute percentage error (MAPE) = 6.14%. Agreement analysis showed a positive bias of 22.00 µg/L, while repeatability tests showed a maximum coefficient of variation (CV) of 3.29%. Embedded inference was highly consistent with Python predictions (R2 = 0.997). These results suggest that the system is a promising portable alternative for Chl-a estimation in prepared acetone extracts under the evaluated conditions.