**4. Experiment**

### *4.1. Pine Wilt Disease-Infected Image Acquisition*

Data acquisition is a primary task in training a robust DNN for classification or regression tasks. For this step, we collected diverse PWD data samples representing various stages of infection, as shown in Figure **??**. Pine wilt can kill a pine tree between 40 days to a few months after infection, and an infected tree shows different symptoms according to its stage of infection. In the early stage, the needles will remain green, but the accumulation of terpenes in xylem tissue results in cavitation, which interrupts water flux in the pine trees. In the second stage, the tree can no longer move water upward, thus causing it to wilt and the needles to turn yellow. The pinewood nematodes then grow in number, and all needles turn yellow-brown or reddish-brown. The disease progresses uniformly branch by branch. After the whole tree is fully dead, the tree needles still remain in place without falling, and show white bare color.

In this work, we obtained UAV images using drone technology. Between August and September, we took orthophotographs with UAVs in the disease-prone region. We tried to capture high-resolution images to best observe the changes in color caused by the disease. We captured high-resolution images from a low altitude using a CMOS camera (Sony Rx R12). The terrain level changes frequently within a hilly region, which affects the resolution quality of an orthophotograph. The images were sequentially taken while overlapping the area. We used drone mapping software (Pixel4Dmapper) to recorrect the GIS coordinates and ensure that the orthophotograph GSD was between 3.2 cm/pixel and 10.2 cm/pixel, where all overlapped patches composed a large orthophotograph ("\*.tif"). Furthermore, each large orthophotograph had an ESRI (Environmental Systems Research Institute) format output with geographic coordinates that specify trees that have been deemed to be potentially infected with PWD by experts. During the training, we converted those geographic coordinates to a bounding box annotation. Then, we conducted a large

orthophotograph crop to small 800 × 800 patches and sent the data to train the network with GT. After the data cleaning, we obtain a total of 4836 images with 6121 PWD damaged tree points as well as 265,694 normal patches (river, roof, field, etc. no PWD images) used to extract the "disease-like" objects. We used five-fold cross-validation to evaluate our proposed system, where each fold included balanced samples of various resolutions. In the real-world scenarios test, we captured another 10 real-world orthophotographs and compared the results with the expert-labeled ground truth points (730 PWD-infected trees).

**Figure 5.** Symptoms of each stage of PWD infection.
