Balanced Dataset

the dataset was obtained as a pre-annotated collection introduced in [1]. This dataset can be accessedat https://doi.org/10.5281/zenodo.7528566. The researcherscreated this dataset using high-resolution aerial imagery(10 cm GSD) obtained from the SWISSIMAGE dataset,which is freely provided by the Swiss Federal Office ofTopography (swisstopo) https://www.swisstopo.admin.ch/de/orthobilder-swissimage-10-cm.  In the data preprocessing stage, to reduce class imbalance andimprove object detection model performance, the initial datasetwas reviewed and balanced. Initially, using the available labelfiles, the sample distribution across each class was calculated.Then, to equalize the number of samples in each class to atarget value (2,500 samples), synthetic data were generated forunderrepresented classes using data augmentation techniques.This augmentation process was performed using the imgauglibrary, and included the following transformations:• Rotation: Between −30◦to +30◦to simulate differentimaging angles.• Scaling: Zooming in or out between 80% and 120% ofthe original image size.• Translation: Shifting the image horizontally and vertically up to 10% of the image dimensions.• Horizontal Flip: Applied with a 50% probability.• Vertical Flip: Applied with a 50% probability.• Gaussian Blur: Used to reduce noise and simulatevarying imaging conditions.• Additive Gaussian Noise: Added to enhance model A sample of these augmented images is shown in Figure 9.The coordinates of the bounding boxes were also updated toreflect the applied image transformations. After the balancingstage was completed, the final images and their correspondinglabels were split into three sets: training, validation, andtesting. These were then saved in standard CSV format.This approach, aimed at achieving a more uniform classdistribution, played a significant role in reducing model biasand improving accuracy during both training and evaluationphases. The number of samples in this newly constructeddataset is illustrated in Figure