We present the results of an Artificial Intelligence (AI)–based galaxy morphology classification study using morphology classification data produced by citizen scientists through the project “Everyone’s Galaxy Research Lab(모두의 은하 연구소)” conducted as part of the program “Science with Citizens(시민과 함께 과학)”. To do that, we construct a high-confidence dataset of 20,854 galaxies by aggregating crowdsourced labels and cross-validating them with previous classifications by experts for large-survey data. We train and compare VGGNet- and ResNet-based convolutional neural networks under different input strategies for the three RGB color channels, including models that integrate all color channels into a single input and ensemble models that process each color channel separately. We find that all models achieve consistently high performance (F1 ≈ 0.89) for primary morphological classes such as elliptical, disk, and spiral galaxies. However, detecting fine substructures including rings, bars, tidal tails, and irregular features is found to require deeper architectures and morphology-dependent input configurations. These results show that computationally efficient channel-separated ensemble models improve precision for geometrically well-defined structures, and that integrated RGB models more effectively capture faint, diffuse structures. Based on these results, we propose an adaptive and cost-efficient AI framework in which we tailor network depth and input configuration to the structural scale and complexity of target galaxies. Our study establishes a scalable deep learning pipeline that bridges citizen science datasets and automated analysis for next-generation large galaxy surveys.

ABSTRACT
1. 서론
2. 연구 방법
    2.1. 연구 설계
    2.2. 목록 통합
    2.3. 형태 분류
    2.4. 자료 신뢰도 검증
    2.5. 기계 학습 및 추론
3. 연구 결과
    3.1. 이전 분류 결과와 비교
    3.2. 모형 학습 결과
    3.3. 추론 결과
4. 논의
    4.1. 모두의 은하 연구소의 자료 신뢰도 및 가이드라인 개선 제안
    4.2. 기존 목록의 대한 논의
    4.3. 유형별 분류 성능 분석
    4.4. 정밀도와 재현율 분석을 통한 오차 요인 진단
    4.5. 희소 구조의 유형 불균형 원인 분석
5. 결론
ACKNOWLEDGMENTS
APPENDIX A. 실험 서버 사양
REFERENCES

• Ha Yoon Park(Daejeon Science High School for the Gifted, Daejeon 34142, Republic of Korea)
• Bomi Park(Astronomy Program, Department of Physics and Astronomy, Seoul National University, Seoul 08826, Republic of Korea, SNU Astronomy Research Center, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Republic of Korea)
• Ho Seong Hwang(Astronomy Program, Department of Physics and Astronomy, Seoul National University, Seoul 08826, Republic of Korea, SNU Astronomy Research Center, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Republic of Korea, Institute for Data Innovation in Science, Seoul National University, Seoul 08826, Republic of Korea) Corresponding author
• Changbom Park(Korea Institute for Advanced Study, 85 Hoegi-ro, Dongdaemun-gu, Seoul 02455, Republic of Korea)