Deep learning combined with Balance Mixup for the detection of pine wilt disease using multispectral imagery.

• The Normalized Difference Vegetation Index can preliminarily screen areas with diseased trees from multispectral images. • Multispectral features learned through 1D convolutional neural networks are efficient in detecting pine wilt disease. • Balance Mixup helps PWDNet achieve total recall more stably without losing precision. Pine wilt disease (PWD) is one of the most destructive infectious diseases affecting coniferous forests. This research proposes a detection method based on deep learning combined with Balance Mixup. At the beginning, regions of interest (ROIs) on multispectral images are automatically generated using normalized difference vegetation index threshold segmentation algorithm. Multispectral data of these ROIs are extracted and labeled as positive (represented by 1) or negative (represented by 0) depending on whether they contained infected trees. During training, the Balance Mixup algorithm selects two multispectral data from a batch of training data according to a certain random strategy. The two data will be mixed into a new multispectral data whose label is the maximum value of the labels (0 or 1) of the original data. The above process will be repeated many times until a new batch of data is generated. A one-dimensional convolutional neural network called pine wilt disease net (PWDNet) is trained by the mixed data and output feature vectors, which are used as input data by a logistic model to determine whether there are PWD in mixed data. By combining Balance Mixup and PWDNet, our method achieved a recall of 1.00 and a precision of 0.90 in testing set. The satisfactory results indicates that the method can provide technical support for the prevention and control of PWD. [ABSTRACT FROM AUTHOR]