Electronic Theses and Dissertation

Optimalisasi fragmentasi batuan pasca peledakan sangat penting untuk meningkatkan efisiensi operasi tambang. Metode evaluasi fragmentasi saat ini sering menghadapi kendala terkait akurasi dan efisiensi analisis citra lapangan yang kompleks. Penelitian ini memperkenalkan pendekatan baru menggunakan Segment Anything Model (SAM) 2.1 berbasis deep learning untuk mensegmentasi dan menganalisis distribusi ukuran fragmen (PSD) batuan secara otomatis dari citra drone. Penelitian dilakukan di Martabe Gold Mine dengan menganalisis 24 citra udara resolusi tinggi. Citra dibagi dalam beberapa patch untuk mengevaluasi kinerja SAM 2.1 dalam mensegmentasi dan mengukur PSD fragmen batuan, yang hasilnya dibandingkan terhadap ground truth dan software komersial WipFrag. Hasil penelitian menunjukkan bahwa SAM 2.1 mencapai akurasi segmentasi yang baik pada patch berkualitas tinggi dengan rata-rata Intersection over Union (IoU) 0,7452, Dice Coefficient 0,8535, Precision 0,7937, Recall 0,9247, dan F1-Score 0,8535, serta menghasilkan pengukuran PSD dengan error rendah, yaitu rata-rata Mean Absolute Error (MAE) 1,78, Mean Squared Error (MSE) 7,60, dan Root Mean Squared Error (RMSE) 2,71. Pada patch berkualitas rendah, akurasi menurun signifikan dengan rata-rata IoU 0,4827, Dice Coefficient 0,5649, Precision 0,3845, Recall 0,5521, dan F1-Score 0,4304 akibat over-segmentation. Namun, meskipun demikian, SAM 2.1 tetap lebih unggul dibandingkan WipFrag dengan nilai MAE 4,65 vs. 18,49, MSE 39,58 vs. 546,44, dan RMSE 6,15 vs. 23,37. Analisis fragmentasi di Martabe Gold Mine menunjukkan bahwa 99,62% fragmen memiliki ukuran ≤75 cm, yang menandakan boulder kurang dari 1%, mengindikasikan efektivitas desain peledakan. Implementasi SAM 2.1 berhasil meningkatkan kecepatan dan akurasi evaluasi fragmentasi tanpa mengganggu operasi tambang. Meskipun masih ada tantangan dalam segmentasi pada citra kompleks.

Kata Kunci: fragmentasi batuan, segmentasi citra, SAM 2.1. WipFrag, Martabe Gold Mine.

Optimizing post-blast rock fragmentation is critical for enhancing mining efficiency, from loading to material processing. Current fragmentation evaluation methods face challenges in accuracy and efficiency, particularly when analyzing complex field images. This study proposes a novel approach using the deep learning-based Segment Anything Model (SAM) 2.1 to automate rock fragment segmentation and particle size distribution (PSD) analysis from drone imagery. Conducted at Martabe Gold Mine, the research analyzed 24 high-resolution aerial images, divided into smaller patches, to evaluate SAM 2.1’s performance in segmenting rock fragments and measuring PSD. Results were compared against ground truth data and the commercial software WipFrag. Results show SAM 2.1 achieves satisfactory segmentation accuracy on high-quality patches, with an average Intersection over Union (IoU) of 0.7452, Dice Coefficient of 0.8535, Precision of 0.7937, Recall of 0.9247, and F1-Score of 0.8535. PSD measurements reflect low error rates: Mean Absolute Error (MAE) of 1.78, Mean Squared Error (MSE) of 7.60, and Root Mean Squared Error (RMSE) of 2.71. On low-quality patches, accuracy drops sharply due to over-segmentation, with IoU of 0.4827, Dice Coefficient of 0.5649, Precision of 0.3845, Recall of 0.5521, and F1-Score of 0.4304. Still, SAM 2.1 outperforms WipFrag, with MAE of 4.65 vs. 18.49, MSE of 39.58 vs. 546.44, and RMSE of 6.15 vs. 23.37. Analysis at Martabe reveals 99.62% of fragments are ≤75 cm, indicating effective blasting design with less than 1% boulders. SAM 2.1 enhances evaluation speed and precision without disrupting mining operations, despite challenges with complex images. Keywords: rock fragmentation, image segmentation, SAM 2.1, WipFrag, Martabe Gold Mine.