<![CDATA[Tanah longsor merupakan bencana yang sering terjadi di wilayah tropis dengan kerugian besar terhadap aspek sosial, ekonomi, dan lingkungan. Sulawesi Tenggara termasuk salah satu wilayah dengan tingkat kerawanan tinggi akibat kondisi topografi berbukit, curah hujan tahunan yang tinggi, serta aktivitas manusia seperti deforestasi dan pertambangan. Penelitian ini bertujuan untuk mengembangkan model klasifikasi daerah rawan longsor menggunakan metode Deep Learning berbasis Convolutional Neural Network (CNN) dengan memanfaatkan data radar Sentinel-1 GRD. Data penelitian mencakup 54 lokasi InaRisk BNPB dari enam kabupaten di Sulawesi Tenggara. Proses penelitian meliputi akuisisi data Sentinel-1, pra-pemrosesan (speckle filtering, kalibrasi radiometrik, koreksi topografi), ekstraksi patch multi-skala, pembangunan dan pelatihan CNN menggunakan TensorFlow/Keras, serta evaluasi model dengan metrik Accuracy, Precision, Recall, F1-score, dan AUC-ROC. Hasil penelitian menunjukkan model terbaik diperoleh pada patch 128×128 dengan akurasi 85,71%, presisi 85,81%, recall 85,71%, F1-score 85,46%, dan AUC-ROC 0,9807. Temuan ini menunjukkan potensi CNN dalam mendukung pemetaan kerawanan longsor secara akurat untuk mitigasi bencana di Sulawesi Tenggara. Landslides are disasters that frequently occur in tropical regions, causing severe impacts on social, economic, and environmental aspects. Southeast Sulawesi is among the regions with high susceptibility due to its hilly topography, high annual rainfall, and human activities such as deforestation and mining. This study aims to develop a landslide susceptibility classification model using a Deep Learning approach based on Convolutional Neural Networks (CNN) by utilizing Sentinel-1 GRD radar data. The research dataset consists of 54 InaRisk BNPB locations across six districts in Southeast Sulawesi. The research process includes Sentinel-1 data acquisition, preprocessing (speckle filtering, radiometric calibration, topographic correction), multi-scale patch extraction, CNN model construction and training using TensorFlow/Keras, and model evaluation with metrics such as Accuracy, Precision, Recall, F1-score, and AUC-ROC. The results show that the best model was achieved using 128×128 patches, reaching an accuracy of 85.71%, precision of 85.81%, recall of 85.71%, F1-score of 85.46%, and AUC-ROC of 0.9807. These findings demonstrate the potential of CNN to support accurate landslide susceptibility mapping for disaster mitigation in Southeast Sulawesi.]]>
