Indirect Prediction of Aquaponic Water Nitrate Concentration Using Hybrid Genetic Algorithm and Recurrent Neural Network

Nitrate concentration contained in the aquaponic water has a crucial effect that influences the growth of lettuce crops. However, most of the existing methods of testing its presence remain mostly destructive, expensive, utilize multiple sensors. To address this issue, a hybrid method of indirectly identifying the nitrate concentration present in aquaponic water of lettuce growth chamber was developed, bridging the integration of computer vision and computational intelligence. The dataset is composed of 720 images of loose-leaf lettuce taken from an aquaponic vertical farm located at Morong, Rizal, Philippines. Graph-cut segmentation was used to segment vegetative green pixels from the background. A regression tree (RTree) multi-feature dimensionality reduction technique was employed to reduce the number of significant features to 15 (a⋆, correlation, B, R, S, Cb, contrast, H, Cr, energy, $\mathrm{b}^{\star}$, entropy, homogeneity, G, and L). To determine the number of optimal hidden layers for the recurrent neural network model, a multigene symbolic regression (MGSR) tool called GPTIPSv2 was employed to derive a fitness function. This function was deployed for optimization using Genetic Algorithm which led to a 107-85-69 hidden-neuron-combination pick. The GA-RNN
15
network surpassed its unoptimized version with 10.25% higher $\mathrm{R}^{2}$ for training, 12.28% higher $\mathrm{R}^{2}$ for validation and 7.28% higher $\mathrm{R}^{2}$ for testing. The findings proved that a nondestructive, cost-efficient, and accurate way of indirectly identifying nitrate concentration levels in aquaponic farm chambers is feasible.