Neural-Intelligent Water Drops algorithm to select relevant textural features for developing precision irrigation system using machine vision

Abstract

Sunagoke moss Rachomitrium japonicum is a good potential for greening material. One of the primary determinants of Sunagoke moss growth is water availability. Too much or too little water can cause water stress in plants. Water stress in plants can be detected by imaging. This study is part of on-going research aimed at developing machine vision-based precision irrigation system in a closed bio-production system for cultured Sunagoke moss. The objective of this study is to propose nature-inspired feature selection techniques to find the most significant set of Textural Features (TFs) suitable for predicting water content of cultured Sunagoke moss. The proposed Feature Selection (FS) methods include Neural-Intelligent Water Drops (N-IWD), Neural-Simulated Annealing (N-SA), Neural-Genetic Algorithms (N-GAs) and Neural-Discrete Particle Swarm Optimization (N-DPSO). TFs consist of 120 features extracted from grey, RGB, HSV, HSL and L ∗a ∗b ∗ colour spaces using ten Haralick’s textural equations. Back-Propagation Neural Network (BPNN) model performance was tested successfully to describe the relationship between water content of Sunagoke moss and TFs. Red Colour Co-occurrence Matrix (CCM) TFs, L ∗ CCM TFs, grey CCM TFs, value (HSV) CCM TFs, green CCM TFs and lightness (HSL) CCM TFs are recommended as individual feature-subset to be used for predicting water content of Sunagoke moss using Artificial Neural Networks. However, FS methods improve the prediction performance. The results show a significant difference between model using FS and models using individual feature-subsets or without FS. Comparative analysis shows the superiority of Neural-Intelligent Water Drops (N-IWD) compared to the other FS methods, since it achieve better prediction performance. The best N-IWD’s fitness function converged with the lowest validation-set Root Mean Square Error (RMSE) of 1.07 × 10 −2 when using 36 TFs.