A Mamdani Fuzzy Inference System for Intelligent Greenhouse Climate Control: Integrating Temperature, Humidity, and Soil Moisture for Adaptive Precision Agricultural Decision-Making

Abstract

Climate management in the greenhouses is a non-linear, multi-variable control problem in which air temperature, relative humidity, and soil moisture interact in ways that conventional threshold-based systems cannot effectively resolve. This study designs, implements, and rigorously evaluates a Mamdani-type Fuzzy Inference System (FIS) for coordinated greenhouse climate regulation using three environmental sensor inputs mapped to a seven-class actuator output covering heating, cooling, ventilation, humidification, misting, irrigation, and steady-state maintenance. The system employs triangular and trapezoidal membership functions partitioned across the full sensor range and a 27-rule expert knowledge base validated against established agronomic guidelines. Performance was assessed on a publicly available dataset of 17,164 IoT greenhouse sensor records spanning four seasons. Results show an overall accuracy of 72.1%; because ground-truth labels were generated by a crisp threshold classifier, this figure reflects agreement with a labelling scheme that structurally favours the crisp system rather than a measure of control quality. The evaluation therefore adopts a comparative behavioural framing centred on macro-averaged F1-score, on which the FIS attains 0.55 versus 0.50 for the crisp baseline, with the largest relative gain on the minority Humidify class (F1 = 0.09 vs. 0.00). Confusion matrix analysis confirms that all misclassifications fall within agronomically proximal action pairs, with zero catastrophic control reversals recorded. Three-dimensional control surface analysis validates the monotonicity and continuity of the decision landscape across all 27 rules. The study contributes a semantic proximity evaluation framework for assessing the operational safety of agricultural control AI, a validated reusable rule base and membership function architecture suitable for microcontroller-class embedded deployment, and quantitative evidence that macro F1-score is a more informative metric than overall accuracy for multi-class agricultural control classification.