Image processing and area estimation of chia (Salvia hispanica L.), quinoa (Chenopodium quinoa Willd.), and bitter melon (Momordica charantia L.) leaves based on statistical and intelligent methods

Abstract

The leaf area (LA) is essential for interception of solar radiation and is determinant factor in plant photosynthesis and penetration of light in the canopy and the plant growth rate. In this research, we assessed the robustness of models on an independent set of data in chia (Salvia hispanica L.), quinoa (Chenopodium quinoa Willd), and bitter melon (Momordica charantia L.). The number of 1000 leaves from each plants were selected from different levels of the canopy during the growth period and immediately transferred to the laboratory and prepared for imaging. After the LA was calculated using image processing system, regression and artificial intelligence methods were considered to LA estimation. Out of 47 evaluated regression models, a+b((L+W) 2 model explained more than 95% of phenotypic variation of LA for three plants (L and W represented length and width of leaf respectively). Among the various training algorithms studied, the trained neural network with Levenberg-Marquardt backpropagation training algorithm had the lowest MSE error and R2 value more than 0.98 in estimating the LA of quinoa, chia and bitter melon plants. In this study, prediction of LA of chia and quinoa with ANFIS (Adaptive Neuro Fuzzy Inference System) and Support Vector Regression (SVR) system indicated that using this system LA is estimated with high accuracy (R2 > 98%). Generally, results showed that methods based on artificial intelligence can estimate the LA in an acceptable manner. This research application codes can be used to generate an application on a smartphone to calculate the LA of chia, quinoa, and bitter melon.