Enhancing PhenoCam Annotation Efficiency via Transfer Learning: Focus on Snow and Image Quality
(2025) In Student thesis series INES NGEK01 20251Dept of Physical Geography and Ecosystem Science
- Abstract
- Globally, automated ecological cameras (PhenoCam) are widely used to monitor vegetation and seasonal changes. Snow is usually easy to identify, but image quality flags (like haze, glare) can make it hard to detect in large datasets. Manual quality control becomes prohibitively time-intensive for large-scale phenological studies, creating a critical need for robust automated snow detection methods. To improve the efficiency and accuracy of image annotation, this study uses a transfer learning approach with the help of a lightweight deep learning model, MobileNetV2, to train the model to recognize the presence of snow and combine it with image quality annotation to make a comprehensive judgment. Using over 5,000 images from two contrasting... (More)
- Globally, automated ecological cameras (PhenoCam) are widely used to monitor vegetation and seasonal changes. Snow is usually easy to identify, but image quality flags (like haze, glare) can make it hard to detect in large datasets. Manual quality control becomes prohibitively time-intensive for large-scale phenological studies, creating a critical need for robust automated snow detection methods. To improve the efficiency and accuracy of image annotation, this study uses a transfer learning approach with the help of a lightweight deep learning model, MobileNetV2, to train the model to recognize the presence of snow and combine it with image quality annotation to make a comprehensive judgment. Using over 5,000 images from two contrasting Swedish sites (Lönnstorp and Röbäcksdalen), we constructed a balanced training dataset spanning 2024 full-year and diverse snow/weather conditions. The training and evaluation of the MobileNetV2-based snow detection model achieved a validation accuracy of 71.98% and an AUC of 0.88, indicating strong discriminatory power in classifying snow presence. Confidence intervals of prediction probabilities showed the model's outputs were conservative, with a narrow range centered around 0.5-0.6. Cross-site validation achieved 71.98% accuracy, demonstrating effective generalization across different vegetation types and lighting conditions. Fine-tuning the model’s sensitivity boosted precision from 0.86 to 1.0, but more testing is needed across different climates and weather patterns. This automated approach reduces manual annotation time by 61% while maintaining 71.98% detection accuracy, enabling more efficient large-scale phenological analyses. Integration with vegetation indices (e.g., Green Chromatic Coordinate) could further improve snow-covered vegetation assessment for climate change monitoring. The methods developed in this thesis, particularly the use of transfer learning and standardized regions of interest filtering, could benefit other ecological monitoring tasks. These include detecting phenological transitions, monitoring vegetation damage caused by extreme weather events, and filtering image quality issues for long-term analysis. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9201327
- author
- Guo, Kexin LU
- supervisor
-
- José Beltran LU
- organization
- course
- NGEK01 20251
- year
- 2025
- type
- M2 - Bachelor Degree
- subject
- keywords
- Physical Geography, Ecosystem Analysis, PhenoCam, Transfer learning, MobileNetV2, Snow detection, Image quality, Phenological monitoring
- publication/series
- Student thesis series INES
- report number
- 695
- language
- English
- id
- 9201327
- date added to LUP
- 2025-06-17 15:49:14
- date last changed
- 2025-06-17 15:49:14
@misc{9201327,
abstract = {{Globally, automated ecological cameras (PhenoCam) are widely used to monitor vegetation and seasonal changes. Snow is usually easy to identify, but image quality flags (like haze, glare) can make it hard to detect in large datasets. Manual quality control becomes prohibitively time-intensive for large-scale phenological studies, creating a critical need for robust automated snow detection methods. To improve the efficiency and accuracy of image annotation, this study uses a transfer learning approach with the help of a lightweight deep learning model, MobileNetV2, to train the model to recognize the presence of snow and combine it with image quality annotation to make a comprehensive judgment. Using over 5,000 images from two contrasting Swedish sites (Lönnstorp and Röbäcksdalen), we constructed a balanced training dataset spanning 2024 full-year and diverse snow/weather conditions. The training and evaluation of the MobileNetV2-based snow detection model achieved a validation accuracy of 71.98% and an AUC of 0.88, indicating strong discriminatory power in classifying snow presence. Confidence intervals of prediction probabilities showed the model's outputs were conservative, with a narrow range centered around 0.5-0.6. Cross-site validation achieved 71.98% accuracy, demonstrating effective generalization across different vegetation types and lighting conditions. Fine-tuning the model’s sensitivity boosted precision from 0.86 to 1.0, but more testing is needed across different climates and weather patterns. This automated approach reduces manual annotation time by 61% while maintaining 71.98% detection accuracy, enabling more efficient large-scale phenological analyses. Integration with vegetation indices (e.g., Green Chromatic Coordinate) could further improve snow-covered vegetation assessment for climate change monitoring. The methods developed in this thesis, particularly the use of transfer learning and standardized regions of interest filtering, could benefit other ecological monitoring tasks. These include detecting phenological transitions, monitoring vegetation damage caused by extreme weather events, and filtering image quality issues for long-term analysis.}},
author = {{Guo, Kexin}},
language = {{eng}},
note = {{Student Paper}},
series = {{Student thesis series INES}},
title = {{Enhancing PhenoCam Annotation Efficiency via Transfer Learning: Focus on Snow and Image Quality}},
year = {{2025}},
}