Abstract
The moisture content (MC) of wood chips impacts their energy output and biomass loss. Traditional MC measurement through oven-drying is extremely time-intensive, while alternative techniques such as near-infrared spectroscopy are expensive. Therefore, an accurate and rapid technique for estimating the MC of wood chips is required to optimize energy production from wood chips while minimizing biomass loss. This study developed regression-based transfer learning models for rapidly predicting the MC of the loblolly pine (Pinus taeda) wood chips. MC of wood chips was adjusted across a wide range (0% to 68%) by adding calculated amounts of water. A total of 2030 RGB images were taken under controlled lighting conditions, and five-fold cross-validation was performed with 11 replications. The pre-trained networks, initialized with ImageNet weights, were fine-tuned to adapt to the wood chip MC task by freezing the first five layers, retraining the remaining layers, and adding new dense layers. All seven fine-tuned pre-trained networks outperformed the Custom CNN model developed from scratch, highlighting the advantage of transfer learning. MobileNetV2_Reg was the best model, achieving a test R2 of 0.90 and a test RMSE of 5.92%. Statistical analysis further showed that MobileNetV2_Reg differed significantly from the Custom CNN. Grad-CAM visualization confirmed that MobileNetV2_Reg focused on the wood chip region when making predictions. These results demonstrate a fast, efficient, scalable, and explainable transfer learning model for MC prediction, which can optimize profitability in industries.