Aquaponics Automation Using Computer Vision

Abstract

Aquaponics systems are heavily plagued by the issue of maintaining better health of fish and higher growth of crops, and achieving a perfect balance between both is always a concern and requires constant attention in the form of water quality, nutrient levels, and vitality of the plant plantations. In this paper, a way of managing an aquaponics facility with real-time computer vision tracking of fish behaviour, plant progress, and water parameters is proposed. We combine multi-spectral imaging (400-1000nm) with deep learning-based analysis to identify early onset of stress in fish (98.3 % accuracy) and nutrient deficiencies in plants (95.1 % accuracy), 22% better than a conventional sensor-only-based system. An effective closed-loop control system actively controls feeding schemes, circulation, and LED grow lights by the visual feedback that decreases manual interventions by 70%. The critical innovations are: (1) a 3D CNN-LSTM combined model that would analyze the temporal-spatial characteristics of fish movement, (2) a model of leaf segmentation, which is not sensitive to water reflections, and (3) an implementation of edge computing that continuously provides service with a latency of 8ms on Raspberry Pi 4. A 25 percent improvement in crop yields (6-month trials with tilapia and lettuce) and a 25 percent decrease in feed loss as compared to the manual systems were evidenced. This is an adventure of building a vision-first sustainable model of sustainable automation within the field of aquaponics that can be applied in any small-scale or commercial situation.