**2. Related Work**

### *2.1. Pine Wilt Disease Classification*

The damage to pine forests caused by PWD is a serious social issue, and several attempts have been made to track the disease using orthophotographs. This method of PWD detection is challenging because it is only possible to collect the data within a limited time. An infected pine tree shows its disease symptoms through changes to the colors of its pine needles; the pine needles gradually change in appearance from solid green to yellow and then brown, and the dead tree color finally turns to ash grey.

Reference [**?** ] analyzed the spatial distribution pattern of damaged trees while introducing the Classification and Regression Trees (CART) model. Another study [**?** ] effectively extracted ecological information to predict risk rates of infected trees using a self-organizing map (SOM) and random forest models.

In later studies, spectral sensors were used to capture hyper-spectral images for PWD analysis [**????** ]. Spectral sensors can provide different surface reflections of PWD that can help evaluate the regions of interest better than the naked eye. For example, near infrared (NIR) is a subset of the infrared band which covers the wavelength range from 780 to 1400 nm. The fusion of the NIR and the red band can reflect the changes in photosynthesis [**?** ]. However, multi-spectral cameras are costly and more unstable than general RGB cameras, as their image quality is affected by various environmental conditions.

As PWD becomes a more serious problem worldwide, machine learning technology is increasingly being used for its automatic detection. The most common method for the early diagnosis of the disease depends on UAV images. In this method [**?** ], a high-quality orthophotograph is generally cropped into small pieces, and a conventional supervised classifier is applied to recognize the disease location in a limited region. In another study, ref. [**?** ] introduced a method using simple classifiers like multi-layer perceptron (MLP) and Support Vector Machine (SVM) to distinguish the regions of healthy or PWD-infected trees.
