SICDet: An Oriented Object Detection Model for Complex Ice Crystal Images

Abstract Ice crystal morphology recognition is fundamental to cloud microphysics. Natural ice crystals exhibit random orientations, large scale variations, and aggregation or adhesion, while image-level single-label classifiers cannot localize their constituent particles. To address these issues, we propose SICDet, an oriented object detection model. SICDet comprises an Orientation-Adaptive Strip Backbone (OASB), a Content-Adaptive Feature Pyramid Network (CA-AFPN), and an Ice-Crystal Morphology-Aware Regression Loss (IC-MARL), which perform multidirectional feature extraction, dynamic scale–detail fusion, and morphology-aware regression weighting, respectively. On NASCENT19, SICDet achieves accuracies of 95.00%, 90.26%, 84.48%, and 78.07% for aggregation-status, basic-habit, microphysical-process, and composite-label recognition. Compared with the YOLO11 oriented detection head, it improves basic-habit, microphysical-process, and composite-label accuracy by 3.54, 5.12, and 2.25 percentage points, respectively. Without fine-tuning on CLOUDLAB, SICDet achieves accuracies of 99.95% for the basic habit of columnar ice crystals and 74.90% for their microphysical processes. The results show that SICDet improves the localization and recognition of constituent particles in complex ice crystal images and generalizes across datasets. It can support automated ice crystal analysis and the interpretation of cloud microphysical processes.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11469187/v1
Primary Topic
Advanced Neural Network Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

SICDet: An Oriented Object Detection Model for Complex Ice Crystal Images

Linjie Deng, Jinrong Hu, Wang Yuhao, Li Lizhi
2 links
Research Square
Advanced Neural Network Applications
preprint

SICDet: An Oriented Object Detection Model for Complex Ice Crystal Images

Linjie Deng, Jinrong Hu, Wang Yuhao, Li Lizhi
preprint en

Abstract

Abstract Ice crystal morphology recognition is fundamental to cloud microphysics. Natural ice crystals exhibit random orientations, large scale variations, and aggregation or adhesion, while image-level single-label classifiers cannot localize their constituent particles. To address these issues, we propose SICDet, an oriented object detection model. SICDet comprises an Orientation-Adaptive Strip Backbone (OASB), a Content-Adaptive Feature Pyramid Network (CA-AFPN), and an Ice-Crystal Morphology-Aware Regression Loss (IC-MARL), which perform multidirectional feature extraction, dynamic scale–detail fusion, and morphology-aware regression weighting, respectively. On NASCENT19, SICDet achieves accuracies of 95.00%, 90.26%, 84.48%, and 78.07% for aggregation-status, basic-habit, microphysical-process, and composite-label recognition. Compared with the YOLO11 oriented detection head, it improves basic-habit, microphysical-process, and composite-label accuracy by 3.54, 5.12, and 2.25 percentage points, respectively. Without fine-tuning on CLOUDLAB, SICDet achieves accuracies of 99.95% for the basic habit of columnar ice crystals and 74.90% for their microphysical processes. The results show that SICDet improves the localization and recognition of constituent particles in complex ice crystal images and generalizes across datasets. It can support automated ice crystal analysis and the interpretation of cloud microphysical processes.

Research Square
Chengdu University of Information Technology (CN)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon