Automated means to classify lab-scale termite damage

摘要

We evaluated convolutional neural networks (CNNs) to visually classify severity of termite attack according to the rating scale from the AWPA E1 standard. The wood wafer dataset contained only 181 uniquely damaged samples. To help compensate for the lack of a large dataset, we performed stratified 5-fold cross validation in which the number of testing samples varied from 33 to 39. Data augmentation was performed on-the-fly on each training set by rotating and flipping samples to generate from 8520 to 8880 images. The CNNs tested included sequential AlexNet, InceptionV3, ResNet50, and InceptionV4-ResNetV2. Each CNN trained relatively well with no sign of overfitting. Sequential AlexNet achieved the best performance, predicting samples with an overall adjusted accuracy of 76.5%, followed by InceptionV3, ResNet50, and InceptionV4-ResNetV2 with 75.0%, 70.8%, and 67.2%, respectively. Sequential AlexNet reached a top-2 accuracy of 97.2%. Mispredictions were more prevalent in the E1 mid-grade ranges, i.e., 4-6 and 8-9. We believe that AlexNet out-performed the other CNNS because of two possible reasons: (1) the other deeper CNNs learned spurious features from the samples that led to more mispredictions; or (2) the other deeper CNNs simply require a training dataset larger than we had available. Future work will focus on developing machine-learning software to grade samples and segment the damage on early/latewood. In addition, we plan to solicit and collect more termite damage samples worldwide in order to enhance our tool and further validate our current findings based only on a limited dataset.