*3.2. Efficient Data Augmentation (EDA)*

Data augmentation is a well-known strategy for significantly increasing the amount of data available for training models without having to collect new data. It acts as a regularizer to reduce bias and generalize the system capability. Real-world UAV images are highly sensitive and collecting them is time-consuming. The image quality of UAV images is dependent on several environmental factors, including reflected light, contrast effects, and camera shake. Meanwhile, detection accuracy is affected by natural weather phenomena such as clouds or thick haze. We used several advanced augmentation methods to improve the performance for locating the PWD in different imaging situations.

**Geometric transformation:** We first cropped 800 × 800 patches from large "\*.tif" images (more than 6 × 10<sup>8</sup> pixels) and applied random horizontal and vertical flip, rotation (0∼90 degree), and resizing (0.9∼2.0 times zooming) to augmen<sup>t</sup> the images and strengthen the model's ability to handle the various resolutions/shooting angles.

**Color space augmentation:** The outward appearance of PWD varies based on the stage of the disease; diseased trees are grayish-green in color in the early stage, with the needles turning brown, and eventually ash grey. In the middle stage of the disease, the color of the diseased leaf (brown) resembles the color of a maple leaf. We carried out random gamma, brightness, and contrast adjustments as well as PCA color augmentation [**?** ] to generate synthetic images from real ones. The trained model with the augmented data including these synthetic images was less sensitive to color and focused more on texture discriminative features.

**Noise injection:** Poor-quality photography devices mounted on UAV often have unavoidable shot noise from unwanted electrical fluctuation that occurs when taking the pictures. We simulated this condition by adding random Gaussian noise (mean of 0∼0.3, std set to 1) to real images. Moreover, the device can be affected by radio interference, and a damaged image may randomly lose information, where this missing data appears as irregular black blocks. To address this problem, we augmented the training data by adopting the robust regularization technique [**?** ] (cutout) to randomly remove various regions of input.

**Other augmentations:** Clouds may block sunlight and create a dark region (shadow) in an image. We added a mask to the original image to change the brightness of the local area of the image so as to simulate shadows. Furthermore, haze is more commonly seen in mountainous areas is another special natural phenomenon that reduces visibility on PWDinfected trees. Changing the opacity of image can generate synthetic haze. We augmented our images to maintain the performance of the network under poor weather conditions.

### *3.3. The Hard Negative Mining Algorithm*

Hard negative mining (HNM) is a bootstrapping method that has been widely used in the classification field and which improves network performance by focusing on hard training samples. We improved the algorithm and applied it to the PWD detection problem. The modified method proceeds according to the following steps: (a) The object detection network is trained on the supervised training dataset. (b) The trained neural network is used to predict the unseen samples (no PWD samples) that are not included in the training samples. (c) The network predicts the objects of interest including "disease-like" objects which can be relabeled into several categories. (d) The "disease-like" objects are merged with genuine PWD-infected objects, and the neural network is retrained on the new dataset. The workflow of selecting "disease-like" objects is shown Figure **??**. Initially, the large-sized orthophotograph ("\*.tif") is divided into small pieces, where the image without GT samples is kept and fed into the trained detector. Then, the model automatically filters out the easily recognizable objects (confidence score < 0.7), and the expert manually relabels the remaining hard objects according to the texture information.

The network with only three categories (background, disease, and hard negative samples) did not perform well due to the ambiguous classification decision boundaries. We manually annotated the hard negative samples into different categories. Figure **??** shows the division of "disease-like" objects and their relationships with the disease in terms of resolution and color. White branch (wb) denotes a dead tree that has a radial umbrella shape. The white-green (wg), yellow, and maple trees have a homogeneous color similar to PWD in its early and middle stages. The oak category indicates the presence of oak tree disease symptoms that resemble PWD-infected trees in UAV images. We also categorize yellow land into a separate category because this can appear similar to PWD, especially in low-resolution images.

**Figure 3.** Extraction of "disease-like" objects.
