ATROUNET: ARSITEKTUR U-NET BERBASIS KONVOLUSI ATROUS DAN RESNET-34 UNTUK SEGMENTASI TUTUPAN LAHAN

Delfia Nur Anrianti Putri, Wahyono, S. Kom., Ph.D.

2025 | Skripsi | ILMU KOMPUTER

Segmentasi tutupan lahan dengan citra satelit dapat membantu  memahami kondisi dan perubahan lahan, tetapi sering menghadapi tantangan dalam variasi spasial dan spektral, degradasi fitur, dan faktor atmosfer. Penelitian ini meninjau penggunaan arsitektur U-Net, backbone ResNet-34, modul Atrous Spatial Pyramid Pooling (ASPP), dan teknik image enhancement pada kualitas segmentasi tutupan lahan pada dataset DeepGlobe Land Cover Classification.

Hasil penelitian menunjukkan bahwa model AtroUNet, U-Net dengan tambahan modul ASPP tanpa image enhancement menunjukkan performa terbaik dengan akurasi 79,83?n mIoU 77,00%. Dalam pelatihan, pretrained backbone ResNet-34 mempercepat konvergensi dan peningkatan metrik. Penambahan contrast stretching dan edge enhancement tidak meningkatkan performa, menunjukkan bahwa efektivitas image enhancement bergantung pada karakteristik dataset. Performa U-Net dengan modul ASPP lebih tinggi sebab modul ini dapat mengekstrak fitur lokal hingga global dengan baik. Selain itu, penggunaan Lovasz-Softmax loss efektif dalam mengoptimasi mIoU meski dataset mengalami ketidakseimbangan ekstrem.

Hal ini menegaskan bahwa U-Net dengan ResNet-34, modul ASPP, serta Lovasz-Softmax loss dapat meningkatkan segmentasi tutupan lahan secara signifikan dalam citra satelit resolusi tinggi. Tetapi, masih dibutuhkan eksplorasi lanjutan untuk teknik image enhancement pada berbagai kondisi citra satelit.

Land cover segmentation from satellite imagery aids in analyzing land conditions and changes but often faces challenges such as spatial and spectral variations, feature degradation, and atmospheric factors. This study examines the use of U-Net, ResNet-34 backbone, Atrous Spatial Pyramid Pooling (ASPP) module, and image enhancement techniques, for land cover segmentation on the DeepGlobe Land Cover Classification dataset.

The results show that the AtroUNet model, U-Net with ASPP module without image enhancement, achieved the best performance with an accuracy of 79.83% and an mIoU of 77.00%. During training, the pretrained ResNet-34 backbone accelerated convergence and improved model metrics. However, contrast stretching and edge enhancement did not improve the results, indicating that the effectiveness of image enhancement depends on the dataset characteristics. The superior performance of U-Net with ASPP module is attributed to its ability to effectively extract both local and global features. Furthermore, the use of Lovasz-Softmax loss proved effective in optimizing mIoU despite the dataset’s extreme class imbalance.

These findings confirm that U-Net with ResNet-34, the ASPP module, and Lovasz-Softmax loss can significantly improve land cover segmentation in high-resolution satellite imagery. However, further exploration of image enhancement techniques is needed to optimize segmentation across various satellite image conditions.

Kata Kunci : Atrous Spatial Pyramid Pooling (ASPP), Deep Learning, Transfer Learning, Image Enhancement, Geospatial Analysis, Land Cover Classification, Remote Sensing