Model Prediksi Moda Transportasi Menggunakan Kecerdasan Buatan Multimodal: Penelitian Kasus Bus Komunitas dan Bus Layanan Responsif Permintaan di Kyushu

Syah, Raden Muhammad Azhar Ardlin (2026) Model Prediksi Moda Transportasi Menggunakan Kecerdasan Buatan Multimodal: Penelitian Kasus Bus Komunitas dan Bus Layanan Responsif Permintaan di Kyushu. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6012231023-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Aksesibilitas terbatas terhadap transportasi umum di daerah pedesaan dan pinggiran kota merupakan tantangan utama untuk kota dengan populasi yang menua dan kepadatan penduduk yang rendah, seperti Jepang. Untuk mengatasi masalah ini, pemerintah daerah telah memperkenalkan layanan alternatif seperti community bus (CB) dan demand-responsive bus (DRB). Namun, mengidentifikasi pola regional yang terkait dengan kedua jenis layanan ini secara objektif dan terukur masih sulit. Penelitian ini mengembangkan kerangka kerja multimodal Artificial Intelligence (AI) yang mengintegrasikan data tabular dengan citra spasial berdasarkan Geographic Information System (GIS) untuk mengklasifikasikan pemerintah daerah ke dalam kategori CB dan DRB. Penelitian ini menggunakan data dari 106 pemerintah daerah di wilayah Kyushu, Jepang (41 CB dan 65 DRB). Percobaan dilakukan dalam dua tahap. Pertama, lima strategi multimodal fusion (concatenation fusion, gated fusion, decision fusion, bilinear fusion, dan full bilinear fusion) dibandingkan. Selanjutnya, model multimodal dengan kinerja tertinggi dibandingkan dengan model hanya citra dan model hanya tabular. Kinerja model dievaluasi menggunakan 5-fold cross-validation. Metrik evaluasi utama adalah matthews correlation coefficient (MCC), dilengkapi dengan macro F1-score, macro precision, accuracy, dan macro recall. Hasilnya, bilinear fusion menunjukkan kinerja deskriptif terbaik, mencapai MCC rata-rata tertinggi (0,4608), macro F1-score (0,7244), dan precision (0,7359), meskipun perbedaan antar-fusion tidak signifikan secara statistik. Dalam perbandingan modality, model multimodal berkinerja lebih baik daripada model hanya gambar dan hanya tabular, tetapi perbedaan keseluruhan antar modality tidak signifikan secara statistik. Analisis SHAP dan Grad-CAM lebih lanjut mengungkapkan bahwa model tersebut bergantung pada indikator regional numerik dan wilayah citra spasial yang bermakna. Ketika diterapkan pada 127 kota di Kyushu yang tidak berlabel, model yang dipilih memprediksi 51 sebagai CB dan 76 sebagai DRB. Sedangkan ketika diterapkan pada area yang sudah memiliki label di Kyushu, model dapat mengklasifikasikan seluruh kota atau kabupaten dengan benar. Hasil ini menunjukkan bahwa AI multimodal menyediakan pendekatan yang layak dan dapat diinterpretasikan untuk klasifikasi CB dan DRB tingkat kota atau kabupaten.
=====================================================================================================================================
Limited accessibility to public transportation in rural and suburban areas is a major challenge for cities with aging populations and low population densities, such as Japan. To address this issue, local governments have introduced alternative services such as community bus (CB) and demand-responsive bus (DRB). However, identifying regional patterns associated with these two types of services objectively and measurably remains difficult. This study develops a multimodal Artificial Intelligence (AI) framework that integrates tabular regional indicators with Geographic Information System (GIS)-based spatial imagery to classify local governments into CB and DRB categories. The study uses data from 106 local governments in the Kyushu region of Japan (41 CBs and 65 DRBs). The experiment was conducted in two stages. First, five multimodal fusion strategies (concatenation fusion, gated fusion, decision fusion, bilinear fusion, and full bilinear fusion) were compared. Next, the highest-performing multimodal model was compared with image-only and tabular-only models. Model performance was evaluated using 5-fold cross-validation. The primary evaluation metrics were the Matthews correlation coefficient (MCC), complemented by macro F1-score, macro precision, accuracy, and macro recall. The bilinear fusion demonstrated the best descriptive performance, achieving the highest average MCC (0.4608), macro F1-score (0.7244), and precision (0.7359), although the differences between fusions were not statistically significant. In the modality comparison, the multimodal model outperformed the image-only and tabular-only models, but the overall difference between modalities was not statistically significant. Further SHAP and Grad-CAM analyses revealed that the model relied on numerical regional indicators and meaningful spatial image regions. When applied to 127 unlabeled cities in Kyushu, the selected model predicted 51 as CB and 76 as DRB. Meanwhile, when applied to areas in Kyushu with existing labels, the model correctly classified all municipalities. These results indicate that multimodal AI provides a feasible and interpretable approach for municipality-level CB and DRB classification.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Klasifikasi Moda Transportasi, AI Multimoda, Bus Komunitas, Bus Layanan Responsif Permintaan, Transport Mode Classification, Multimodal AI, Community Bus, Demand-Responsive Bus
Subjects:	T Technology > TE Highway engineering. Roads and pavements > TE7 Transportation--Planning
Divisions:	Faculty of Civil, Planning, and Geo Engineering (CIVPLAN) > Civil Engineering > 22101-(S2) Master Thesis
Depositing User:	Raden Muhammad Azhar Ardlin Syah
Date Deposited:	03 Aug 2026 07:14
Last Modified:	03 Aug 2026 07:14
URI:	http://repository.its.ac.id/id/eprint/142213

Actions (login required)

View Item