Speaker
Description
Fungal sapwood discolouration is a major cause of value loss in Norway spruce sawlogs and directly affects grading outcomes according to EN 1927-1. Spores of fungi responsible for sap-stain are naturally widespread in forest ecosystems and spread via insects, wind, rain splash, and contaminated harvesting equipment, increasing the risk of quality loss during log storage.
This study evaluates a supervised deep learning approach for automated detection and quantification of sapwood discolouration in roundwood cross-sections. A dataset comprising 928 images of Norway spruce (Picea abies (L.) Karst.) was acquired. Stem discs were placed in a measuring frame and photographed using a digital camera (APS-C format) with a fixed 30 mm focal length lens, mounted on a tripod under varying natural lighting conditions typical of wood yard environments. Discoloured regions were delineated, and the masks distinguished heartwood, non discoloured sapwood, and discoloured sapwood. These masks served as ground truth labels for model training. Images and masks were converted to image resolutions of 128×128, 256×256, and 512×512 pixels, split into training (70%), test (20%), and validation (10%) subsets and augmented by rotations and flipping.
A U-Net was trained for supervised pixel-wise semantic segmentation (discoloured vs. non-discoloured wood). The encoder of the U-Net used convolutional blocks with 16-64 initial filters depending on the image resolution, and downsampling via max pooling. The decoder mirrored the encoder with upsampling and skip connections and a final 1×1 convolution with sigmoid activation produced binary probability masks. Training used the Adam optimizer, and the loss combined binary cross entropy and Dice. Evaluation metrics were the Dice coefficient and the Jaccard index (IoU). Models were trained using a batch size of 30.
Across model configurations Dice coefficients ranged from 0.871 to 0.908 and Jaccard indices from 0.783 to 0.838 on the independent test dataset, demonstrating robust segmentation performance despite heterogeneous illumination and non-standardized camera-to-object distances.
The results indicate that supervised CNN-based image segmentation enables objective, reproducible, and spatially explicit quantification of sapwood discolouration in roundwood. From a forest supply chain management perspective, such methods represent a step toward sensor-assisted quality assessment in harvesting and storage logistics. Integration into mobile camera systems, log-yard scanning setups, or future harvester- or loader-mounted imaging systems could support continuous quality monitoring along the supply chain. This would enable earlier detection of degradation processes, improved storage strategy decisions, and more data-driven timber allocation — contributing to the digitalization and increased transparency of mechanized wood supply systems.
| Keywords | log-quality; automated-stain-detection; U-net; sensor-assisted-grading |
|---|
