Estimation of nitrate content in tomato using image features

The improper use of chemical fertilizers in crop production can result in unsafe food sources for consumers. This research focuses on estimating the accumulation of nitrate in tomatoes by analyzing images of tomato tissues. The experiments were conducted using a completely randomized design with four nitrogen levels: 400, 800, 1200, and 1600 kg.ha-1. Fifty samples were randomly selected from each treatment to create images for feature processing and develop a prediction model. The samples were sliced to a consistent thickness, and their images were prepared. The nitrate contents of the same samples were then measured in the laboratory. Color features, including R, G, and B color components, as well as non-color features such as white pixel area (WPA), total slice surface area (TSA), and the ratio of white pixel area to total slice surface area (WPA/TSA), were extracted from the images. The results showed that the nitrate content of the samples increased significantly (P<0.05) in response to the applied nitrogen fertilizer, with measurements of 1.6%, 2.7%, 2.8%, and 3.3%, respectively. Moreover, a strong correlation was found between the color components, WPA, TSA, WPA/TSA, and nitrate accumulation in the samples. Multiple regression and multilayer perceptron neural network (MLP) models were employed to predict the nitrate content. The best subset method was used to build an appropriate regression model. Various topologies and transform functions were applied to identify the best MLP model. The results indicated that an MLP model with a 3-15-1 topology and the lowest mean relative percentage error (MRPE) was the most accurate neural network model. The final regression and neural network models were validated using 60 intact samples. The neural network model achieved a MRPE of approximately 3.5%, demonstrating its precise estimation of nitrate contents compared to the regression model with an MRPE of around 5.2%.