Robustness of the RGB image-based estimation for rice above-ground biomass by utilizing the dataset collected across multiple locations

Above-ground biomass (AGB) is a critical phenotype representing crop growth. Non-invasive evaluations of AGB, including deep-learning-based red-green-blue (RGB) image analyses, are often specific to the training data. The robustness of the estimation model across untrained conditions is essential to monitor crop productivity globally, but it has yet to be fully assessed. This study aims to assess the robustness of a convolutional neural network (CNN) model for rice AGB estimation across five locations in three countries, and to demonstrate the feasibility of robust model via a practical approach. From transplanting to heading, 1957 RGB images were captured vertically downward over the rice canopy, covering approximately 1 m2. First, a base model was established using data collected from a single location. Then, its robustness was assessed using test datasets taken from the other four locations. The CNN model showed a significant variation in estimation accuracy across the untrained four locations, indicating insufficient robustness of the base model. Subsequently, we quantitatively tested the impact of improving training data diversity on model robustness by adding data from each of the four locations to the base model's training data. Adding at most 48 data points from a location achieved practical accuracy for the added location, with R2Ad above 0.8. Interestingly, adding data from one location sometimes improved the accuracy for other untrained locations as well. These findings suggest that collecting diverse training data for RGB-based estimation, combined with evaluation of robustness paves the way for on-site and instant AGB monitoring of rice.