Perbandingan Kinerja Model U-Net dan YOLOv11 dalam Klasifikasi Batupasir pada Citra Sayatan Tipis

Azkia Pramesti Trapsila Siwi, Dr.rer.nat. Ir. I Wayan Warmada, IPM; Ir. Esti Handini, S.T., M.Eng., D.Sc., IPM.

2026 | Skripsi | TEKNIK GEOLOGI

Analisis petrografi sayatan tipis merupakan metode fundamental untuk mengidentifikasi komposisi mineral kerangka batuan dan menentukan klasifikasi batupasir. Namun, metode manual seperti point counting membutuhkan waktu lama dan rentan terhadap subjektivitas pengamat. Penelitian ini bertujuan untuk mengevaluasi dan membandingkan kinerja model deep learning U-Net dan YOLOv11 dalam melakukan segmentasi mineral kuarsa, feldspar, dan fragmen litik pada citra sayatan tipis batupasir, serta menguji penerapannya dalam klasifikasi batuan menurut diagram Pettijohn (1973). Data penelitian mencakup citra sayatan tipis dari Formasi Kebo-Butak dan Formasi Ngrayong yang diambil menggunakan mikroskop polarisasi pada kondisi pencahayaan Plane Polarized Light (PPL) dan Cross Polarized Light (XPL). Dataset terdiri atas 1.200 citra berpasangan. Seluruh citra melalui tahapan alignment, cropping, dan penggabungan citra sebelum digunakan dalam proses pelatihan model. Evaluasi model dilakukan menggunakan metrik Pixel Accuracy, Mean Intersection over Union (mIoU), Precision, Recall, dan F1-score. Hasil segmentasi kemudian digunakan untuk menghitung komposisi mineral dan melakukan klasifikasi batupasir berdasarkan diagram Pettijohn (1973). Hasil pengujian menunjukkan bahwa U-Net menghasilkan kinerja segmentasi yang lebih unggul dibandingkan YOLOv11 pada seluruh metrik evaluasi, dengan Pixel Accuracy sebesar 0,7868, mIoU sebesar 0,6209, Precision sebesar 0,7419, Recall sebesar 0,7746, dan F1-score sebesar 0,7562, sedangkan YOLOv11 memperoleh Pixel Accuracy 0,7624, mIoU 0,5554, Precision 0,6990, Recall 0,7130, dan F1-score 0,7042. Kedua model mampu merekonstruksi klasifikasi Quartz Arenite pada sampel kaya kuarsa dan Feldspathic Wacke pada sampel Kebo-Butak. Meskipun demikian, kedua model belum berhasil mengidentifikasi kuarsa pada sampel Kebo-Butak akibat kontras optik yang rendah dan ketidakseimbangan data latih. Secara keseluruhan, U-Net terbukti lebih baik dalam mempertahankan detail batas spasial antarbutir mineral sehingga paling sesuai diterapkan untuk otomatisasi analisis petrografi batupasir.

Thin-section petrographic analysis is a fundamental method for identifying framework mineral compositions and determining sandstone classification. However, manual methods such as point counting are time-consuming and prone to observer subjectivity. This study aims to evaluate and compare the performance of U-Net and YOLOv11 deep learning models in segmenting quartz, feldspar, and lithic fragments from sandstone thin-section images, as well as to test their application in rock classification according to the Pettijohn (1973) diagram. The research data comprises thin-section images from the Kebo-Butak Formation and Ngrayong Formation acquired using a polarizing microscope under Plane Polarized Light (PPL) and Cross Polarized Light (XPL) conditions. The dataset consists of 1,200 paired images. All images underwent alignment, cropping, and image merging stages prior to model training. Model evaluation was conducted using Pixel Accuracy, Mean Intersection over Union (mIoU), Precision, Recall, and F1-score metrics. The segmentation results were subsequently used to calculate mineral compositions and perform sandstone classification based on the Pettijohn (1973) diagram. The test results show that U-Net achieved superior segmentation performance compared to YOLOv11 across all evaluation metrics, yielding a Pixel Accuracy of 0.7868, mIoU of 0.6209, Precision of 0.7419, Recall of 0.7746, and F1-score of 0.7562, whereas YOLOv11 obtained a Pixel Accuracy of 0.7624, mIoU of 0.5554, Precision of 0.6990, Recall of 0.7130, and F1-score of 0.7042. Both models successfully reconstructed the classification of Quartz Arenite on the quartz-rich sample and Feldspathic Wacke on the Kebo-Butak sample. Nevertheless, neither model succeeded in identifying quartz in the Kebo-Butak sample due to low optical contrast and training data imbalance. Overall, U-Net proved superior in preserving intergranular spatial boundary details, making it the most suitable model for automated petrographic analysis of sandstone.

Kata Kunci : Batupasir, Segmentasi, U-Net, YOLOv11, Klasifikasi Pettijohn