Electronic Theses and Dissertation

Perkembangan teknologi penginderaan jauh berbasis drone telah meningkatkan potensi pemetaan dan ekstraksi informasi geospasial secara otomatis. Namun, proses identifikasi objek pada citra ortofoto masih banyak dilakukan secara manual sehingga memerlukan waktu dan tenaga yang besar. Penelitian ini bertujuan mengembangkan model segmentasi semantik berbasis U-Net++ untuk mendeteksi objek bangunan, jalan, dan badan air secara otomatis pada citra ortofoto drone. Dataset yang digunakan berupa citra ortofoto hasil akuisisi drone yang telah diberi anotasi sebagai ground truth. Tahapan penelitian meliputi tiling, pencocokan citra dan mask, perubahan ukuran citra menjadi 512 × 512 piksel, normalisasi, serta augmentasi data untuk meningkatkan keragaman sampel pelatihan. Model dilatih menggunakan optimizer Adamdengan kombinasi Binary Cross Entropy (BCE) Loss dan Dice Loss, kemudian dievaluasi menggunakan metrik Accuracy, Precision, Recall, F1-Score, dan Intersection over Union (IoU).
Hasil penelitian menunjukkan bahwa model U-Net++ mampu melakukan segmentasi objek dengan performa yang sangat baik. Model memperoleh Accuracy sebesar 92,64%, Precision 94,89%, Recall 92,13%, F1-Score 93,49%, dan IoU rata-rata 87,78%. Berdasarkan evaluasi per kelas, segmentasi jalan memberikan performa terbaik dengan nilai IoU sebesar 92,45%, diikuti bangunan sebesar 88,22%, dan badan air sebesar 83,46%. Hasil visualisasi menunjukkan bahwa model mampu mempertahankan bentuk geometris objek dan menghasilkan batas segmentasi yang relatif akurat. Performa tersebut didukung oleh penggunaan citra beresolusi tinggi, mekanisme nested skip connections pada U-Net++, serta strategi augmentasi data yang meningkatkan kemampuan generalisasi model. Penelitian ini menunjukkan bahwa U-Net++ merupakan pendekatan yang efektif untuk segmentasi otomatis citra ortofoto drone dan berpotensi mendukung pemetaan, perencanaan infrastruktur, serta pengembangan sistem informasi geospasial.

Kata kunci: deep learning, U-Net++, semantic segmentation, ortofoto drone

The advancement of drone-based remote sensing technology has created new opportunities for automated mapping and geospatial information extraction. However, object identification from drone orthophotos is still largely performed manually, making the process time-consuming and labor-intensive. This study aims to develop a semantic segmentation model based on the U-Net++ architecture for the automatic detection of three object classes, namely buildings, roads, and water bodies, from drone orthophotos. The dataset consisted of drone orthophotos manually annotated to produce ground truth masks. Data preprocessing included image tiling, image-mask pairing, resizing to 512 × 512 pixels, normalization, and data augmentation to improve sample diversity. The U-Net++ model was trained using the Adam optimizer with a combination of Binary Cross Entropy (BCE) Loss and Dice Loss. Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, and Intersection over Union (IoU) metrics. The experimental results demonstrate that the proposed U-Net++ model achieved excellent segmentation performance, with an Accuracy of 92.64%, Precision of 94.89%, Recall of 92.13%, F1-Score of 93.49%, and a mean IoU of 87.78%. Class-wise evaluation showed that road segmentation achieved the highest IoU (92.45%), followed by buildings (88.22%) and water bodies (83.46%). Visual analysis indicated that the model effectively preserved object geometry and generated accurate segmentation boundaries. The high performance was attributed to the use of high-resolution orthophotos, the nested skip connection mechanism of U-Net++, and an effective data augmentation strategy that improved model generalization. These findings indicate that U-Net++ is a robust and promising approach for automatic object segmentation from drone orthophotos, supporting geospatial mapping. Keywords: deep learning, U-Net++, semantic segmentation, drone orthophoto