Abstract
Moisture content in wood chips is a critical parameter for industries such as pelleting mills, bio-refineries, paper mills, and renewable energy production. The moisture level significantly influences both the quality of the final product and the efficiency of the production process. Consequently, accurate knowledge of moisture content is of substantial importance to wood chip-reliant industries. However, current methods for determining moisture content are either time-consuming or require costly equipment and specialized setups. Therefore, developing a quick and reliable method for assessing wood chip moisture content is imperative. To address this need, we evaluate fourteen Vision Transformer (ViT) architectures and introduce an optimized model, MoistViT, developed using Bayesian Optimization Hyperband (BOHB) for efficient hyperparameter tuning. Experiments on two wood chip image datasets (1600 total images) show that MoistViT achieves 91% accuracy and 92% F1-score on Source 1 and 93% accuracy and 93% F1-score on Source 2, outperforming all baseline models. Subsequently, a thorough analysis of failure cases has been carried out, including the identification of the most challenging groups of moisture levels. These analyses provide valuable insights into the complex task of determining moisture content from inherently heterogeneous wood chips. The proposed MoistViT demonstrates significant potential for real-time applications in relevant industries, which could ultimately lead to a streamlined production process.