Automatic detection of root rot and resin in felled Scots pine stems using convolutional neural networks

Abstract

Root rot caused by Heterobasidion spp. is the most serious fungal disease of conifer forests in the Northern Hemisphere. In Scots pine (Pinus sylvestris L.) stands infected by H. annosum, root rot reduces sawlog quality due to decay and resin-soaked patches. Automatically detecting the disease during harvesting operations could be used to optimize bucking as well as to efficiently collect data on root-rot incidence within forest stands and at larger geographical scales. In this study, we develop deep learning models based on convolutional neural networks to automatically detect root rot disease and the presence of resinous wood in stem end images of Scots pine. In addition, we study the effect of pre-filtering the images via a classical texture operator prior to model development. Using transfer learning on pre-trained feature extractor networks, we first construct classifiers for detecting severely rotten wood in stem end images. Second, we develop a classifier for detecting the presence of resin outside branch knots. In rot detection, using regular RGB images, our final model reaches a binary classification accuracy of (63 ± 6)% on the independent test data, where the error is the standard error. Pre-processing the images using the classical texture operator increases the final classification accuracy to (70 ± 6)%. To detect only resin using regular RGB images, we find an accuracy of (80 ± 6)%. Finally, we discuss the operational implications and requirements of implementing such computer vision algorithms in the next generation of forest harvesters.