<![CDATA[Periodic land cover monitoring in the Core Zone of Parangtritis Sand Dunes is crucial for conservation, but manual annotation of UAV imagery is time-consuming and subjective. This study implements a Standard U-Net for automatic semantic segmentation of sand and non-sand classes on UAV images. The training data consists of an integration of manual annotations (463 UAV images from 2019) and semi-automated annotations based on the Segment Anything Model (SAM) on 98 UAV images from 2022, while the test data comprises 80 UAV images from 2022 that the model has not previously encountered. Preprocessing included tiling images and masks into 256×256 patches, with data augmentation via rotation, zoom, brightness/contrast adjustment, and horizontal flip. The model was trained using a combined Binary Cross Entropy and Dice loss, Adam optimizer (learning rate 10⁻⁴), batch size 8, and a fixed threshold of 0.45. Evaluation on 80 independent 2022 test images yielded Pixel Accuracy of 73.31%, Dice Coefficient of 83.28%, and Intersection over Union (IoU) of 71.35%. Qualitatively, the model captured the general shape of sand areas well, though noise appeared in heterogeneous scenes, while the highest accuracy was achieved on homogeneous sand textures. The practical implication is accelerating precise spatial data provision for agencies monitoring sand dune changes. Further research is recommended to improve ground truth quality, integrate shadow correction, explore modern architectures, and apply loss functions handling class imbalance for comprehensive accuracy enhancement.]]>
Copyrights © 2026
