A machine learning-based lettuce fresh weight estimation framework incorporating agronomic traits and image features

• Constructing a machine learning-based lettuce fresh weight estimation framework. • Agronomic traits and image features jointly serve as inputs to the framework. • 7 machine learning models are compared, and DNN exhibits the highest performance. • The number of leaves is the most dominant factor affecting fresh weight estimation. Fresh weight is an important indicator for quantifying the growth state of lettuce. The utilization of raw RGB images and multidimensional features, in conjunction with advanced techniques such as machine learning, has become a popular trend for estimating lettuce fresh weight. However, there is currently a lack of comprehensive studies comparing the performance of various machine learning models in the task of lettuce fresh weight estimation. Moreover, the fact that machine learning models are data-driven means that the mechanisms behind the decisions these models make based on multi-source inputs are not easy to explain. In this study, a machine learning-based framework of lettuce fresh weight estimation taking agronomic traits and image features as input was developed using two self-made datasets. Our comparative study reveals that, out of the seven machine learning models evaluated, the deep neural network (DNN) exhibits the highest accuracy in estimating fresh weight, coupled with the strongest model stability and cross-variety applicability. For the DNN and extra trees (ET) models with relatively superior performance, we implement the Shapley additive explanations (SHAP) algorithms to quantitatively analyze the impact of input features on the models. The results indicate that agronomic traits contribute more to the estimation results compared to image features, and the number of leaves is the most dominant factor. The framework is capable of identifying the dominant factors influencing the model’s output, providing not only highly accurate estimations but also enhancing the understanding of the model’s decision making.