2023 Volume 92 Issue 2 Pages 129-139
We developed a simple method for estimating nutrient index values of ‘Kinumusume’, a recommended rice variety of Okayama Prefecture, from RGB images. ‘Kinumusume’ was grown under different transplanting dates and fertilizer application conditions. Plant length, number of tillers, and SPAD were measured; and, their product value, the nutrient index value, was calculated at around 30, 20, and 10 days before heading. RGB images of the rice plants were taken just before each day’s growth measurement. Sixteen color indices were calculated from the original RGB images (whole area images) and RGB images from which only the rice area was extracted (rice area images), and models were developed to estimate nutrient index values using single regression analysis and 16 machine learning algorithms. Data from different fields were used for model development and accuracy validation. In both single regression and machine learning, rice area images tended to predict nutrient index values with higher accuracy than whole area images. This was considered to be due to the removal of the effect of the image background. The single regression model showed less decrease in accuracy from training data to test data than the machine learning models, which suggested that the single regression model would be superior to machine learning algorithms in that it can capture more general features even from a small set of data.
