<![CDATA[This study proposes a transfer learning framework for water quality classification in tilapia (Oreochromis niloticus) biofloc systems, leveraging a fuzzy inference system (FIS) previously developed and validated for catfish (Clarias gariepinus) biofloc monitoring as the source-domain model. Transfer is performed at the parameter level by shifting the trapezoidal membership-function boundaries to tilapia physiological thresholds, leaving the rule base and inference structure unchanged. Because a Mamdani FIS is knowledge-driven (its parameters are derived from the literature rather than learned from data), evaluation is conducted on a 200-sample literature-anchored synthetic tilapia dataset generated from Scopus-indexed biofloc parameter ranges and anchored to the noise characteristics of real IoT sensors from the same device; the synthetic set acts as a test bench that deliberately covers boundary conditions that are rare and unethical to induce in a live pond. The Transfer-FIS is compared with an Unadapted Source-FIS and a Scratch-FIS optimized entirely on tilapia data. The Transfer-FIS achieves 98.00% accuracy and a macro F1-score of 0.9733; a Wilcoxon signed-rank test over 10-fold cross-validation detects no statistically significant difference from the Scratch-FIS (98.50%, F1 = 0.9790; p = 1.000; Cohen’s d ≈ 0.21). Taking Non-Ideal as the positive class, the Transfer-FIS produces zero false negatives (missed alarms) whereas the Scratch-FIS produces three, and the Transfer-FIS reaches ~97% accuracy with only 20 target samples. These findings indicate that parameter-level fuzzy transfer is a viable, data-efficient, and operationally conservative strategy; validation on field-collected tilapia data is identified as the next step.]]>
Copyrights © 2026
