Design and Development of a Human-in-the-Loop Review Interface for the Seaseek Marine Annotation Platform

Adnyana, Dimas Gistha (2026) Design and Development of a Human-in-the-Loop Review Interface for the Seaseek Marine Annotation Platform. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026221057-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Underwater ecological surveys increasingly generate image collections whose interpretation requires time-consuming manual annotation. Image degradation, uncertain visual categories, and repeated observations across adjacent frames make fully manual processing difficult to scale and prevent automated predictions from being treated as verified ecological data. This Final Project designs, develops, and evaluates a human-in-the-loop review interface for the Seaseek marine annotation platform. The evaluated scope covers two role-specific workflows. Line A supports technician review of fish and sea cucumber detections through bounding-box inspection, acceptance, rejection, reclassification, redrawing, and manual annotation creation. Line B supports ecologist review of sampled benthic points through coral, sand, and unknown labels, label overrides, synchronized image and list selection, and coverage presentation. The study follows a six-stage Design Science Research Methodology supported by iterative prototyping. Evaluation combines functional and black-box testing, repository inspection, internal walkthroughs, and two rounds of formative user testing. The resulting interface presents model outputs as provisional, retains confidence and origin information where available, distinguishes review states, and provides loading, empty-state, and failure feedback. Six unit tests and three dashboard smoke tests passed successfully, while the production build completed with a warning. Linting identified seven errors and twenty warnings. Complete live Line A and Line B correction, persistence, and reload sequences were not reproduced during the final evidence audit. The results indicate that two distinct review workflows can share navigation, state language, feedback patterns, and frontend components without merging their annotation models. The resulting artifact remains a non-production prototype. The evaluation does not establish model accuracy, ecological validity, reduced annotation time, or usability with practicing marine researchers. Further work is required to support authenticated review histories, production-level security, comprehensive integration testing, accessibility, and formal domain-specific evaluation.
====================================================================================================================================
Survei ekologi bawah air semakin menghasilkan koleksi citra yang interpretasinya memerlukan anotasi manual yang memakan waktu. Degradasi citra, ketidakpastian kategori visual, dan pengamatan berulang pada frame yang berdekatan menyebabkan proses manual sulit diskalakan serta menghambat prediksi otomatis untuk dapat diperlakukan sebagai data ekologi yang telah terverifikasi. Tugas Akhir ini merancang, mengembangkan, dan mengevaluasi antarmuka tinjauan *human-in-the-loop* untuk platform anotasi kelautan Seaseek. Ruang lingkup evaluasi mencakup dua alur kerja berbasis peran. Line A mendukung tinjauan teknisi terhadap deteksi ikan dan teripang melalui pemeriksaan kotak pembatas, penerimaan, penolakan, reklasifikasi, penggambaran ulang, dan pembuatan anotasi manual. Line B mendukung tinjauan oleh ekolog terhadap titik bentik yang disampel melalui pelabelan karang, pasir, dan tidak diketahui, perubahan label, sinkronisasi pemilihan pada citra dan daftar, serta penyajian persentase tutupan. Penelitian menggunakan *Design Science Research Methodology* enam tahap yang didukung oleh pengembangan prototipe secara iteratif. Evaluasi menggabungkan pengujian fungsional dan *black-box*, pemeriksaan repositori, penelusuran internal, serta dua putaran pengujian pengguna formatif. Antarmuka yang dihasilkan menyajikan keluaran model sebagai usulan sementara, mempertahankan informasi tingkat keyakinan dan asal data ketika tersedia, membedakan status tinjauan, serta menampilkan umpan balik untuk proses pemuatan, kondisi tanpa data, dan kegagalan. Enam pengujian unit dan tiga *dashboard smoke test* berhasil dijalankan, sedangkan proses *build* produksi selesai dengan satu peringatan. Pemeriksaan *lint* menemukan tujuh galat dan dua puluh peringatan. Rangkaian lengkap koreksi, penyimpanan, dan pemuatan ulang pada Line A dan Line B menggunakan layanan aktif belum berhasil direproduksi dalam audit bukti akhir. Hasil penelitian menunjukkan bahwa dua alur kerja tinjauan yang berbeda dapat berbagi navigasi, bahasa status, pola umpan balik, dan komponen *frontend* tanpa menyatukan model anotasinya. Artefak yang dihasilkan masih berupa prototipe nonproduksi. Evaluasi belum membuktikan akurasi model, validitas ekologi, pengurangan waktu anotasi, maupun tingkat kegunaan bagi peneliti kelautan. Pengembangan lebih lanjut diperlukan untuk mendukung riwayat tinjauan yang terautentikasi, keamanan tingkat produksi, pengujian integrasi yang lebih menyeluruh, aksesibilitas, dan evaluasi formal berbasis domain.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	anotasi citra bawah air, human-in-the-loop, antarmuka tinjauan, interaksi manusia-AI, design science research, underwater image annotation, review interface, human-AI interaction
Subjects:	Q Science > QA Mathematics > QA76.9.U83 Graphical user interfaces. User interfaces (Computer systems)--Design.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Dimas Gistha Adnyana
Date Deposited:	03 Aug 2026 06:26
Last Modified:	03 Aug 2026 06:26
URI:	http://repository.its.ac.id/id/eprint/142191

Actions (login required)

View Item