Time‐series digital camera photos combined with machine learning algorithms can realize accurate observation of flowering phenology
- 1Key Laboratory of Vegetation and Environmental Change Institute of Botany, Chinese Academy of Sciences Beijing 100093 China
Abstract
Digital cameras are widely used for documenting phenological observations, and numerous images have been collected. However, intelligent approaches are required to extract valuable phenological information from time‐series images. In this study, we used machine learning (ML) algorithms, including convolutional neural network (CNN)‐based You Only Look Once (YOLO) object detection and semantic segmentation methods to identify flowers in images, establish curves of flower count and flower cover, and extract the phenophases of first, peak and end flowering. Random forests (RF) was performed to recognize flower pixels to calculate the flower cover, construct the flower cover curve and extract the same phenophases as those of the YOLO methods. Furthermore, flowering phenophases were also extracted through manual visual identification. We used a generalized additive model (GAM) to fit curves for flower count and flower cover, and extracted flowering phenophases by calculating the inflection points of the fitted curves. We found that (1) YOLO‐based methods could effectively identify flowers, and the variation in flower count and flower cover obtained from the YOLO object detection and semantic segmentation models reflected the trend of flowering phenology. The flower count and flower cover curves effectively supported the extraction of first and peak flowering. The difference between the YOLO‐identified and manually identified flowering phenophases ranged from 1 day to 3 days using the optimal thresholds. For end flowering, except for the end flowering identified based on flower count derived from YOLO object detection, the date difference in phenophases between the YOLO‐identified and manually identified ranged from 1 day to 8 days. (2) There are apparent outliers in the RF‐calculated flower cover values, particularly during the post‐peak‐flowering period. However, the identified flowering phenophases based on the RF‐derived flower cover curve after omitting outliers were consistent with those of manual visual identification and YOLO‐based methods (except end flowering identified based on flower count derived from YOLO object detection), with the date difference in phenophases ranging from 0 to 8 days. (3) The GAM performed well in fitting the trends of the normalized cumulative flower count and flower cover. Using the threshold generated by second derivate method, the identified end flowering was close to that of “late flowering” stage identified by manual visual identification, and the date difference ranged from 0 to 6 days. (4) Due to the variation in flowering rhythm and progression across different plant species, fixed thresholds are not fully optimal for all plants, and the thresholds used to extract flowering phenology require targeted adjustments based on specific observed species. Our study showed that a time‐lapse digital camera combined with ML algorithms can help improve the objectivity of phenology observations, indicating the possibility of using ML algorithms to identify flowering phenology.
Related
- Influence of flowering phenology and strobili amount on mating patterns in an Abies nordmanniana clonal seed orchard
- Aerial Phenotyping of Coffee Flowering Using Deep Learning for Early Yield Prediction
- Advancing grapevine phenology predictions: a comparison of regression-based statistical learning approach and conventional models in mediterranean environment.
- Synthetic-augmented multimodal deep learning fuses dual-angle RGB images and phenology to unlock genotype-informative canopy structural trait in wheat
- ML4SD: Leveraging Machine Learning and High-Throughput Search Algorithms for an Iterative Growth-Coupled Design Innovation
BibTeX
@article{song2026timeseries,
title = {{Time‐series digital camera photos combined with machine learning algorithms can realize accurate observation of flowering phenology}},
author = {Chuangye Song and Yuan Jia and Lin Zhang and Dongxiu Wu},
journal = {Remote Sensing in Ecology and Conservation},
year = {2026},
month = oct,
doi = {10.1002/rse2.70069},
url = {https://doi.org/10.1002/rse2.70069},
note = {CC BY},
}
RIS
TY - JOUR
TI - Time‐series digital camera photos combined with machine learning algorithms can realize accurate observation of flowering phenology
AU - Chuangye Song
AU - Yuan Jia
AU - Lin Zhang
AU - Dongxiu Wu
PY - 2026
DA - 2026/10/31
JO - Remote Sensing in Ecology and Conservation
DO - 10.1002/rse2.70069
UR - https://doi.org/10.1002/rse2.70069
AB - Digital cameras are widely used for documenting phenological observations, and numerous images have been collected. However, intelligent approaches are required to extract valuable phenological information from time‐series images. In this study, we used machine learning (ML) algorithms, including convolutional neural network (CNN)‐based You Only Look Once (YOLO) object detection and semantic segmentation methods to identify flowers in images, establish curves of flower count and flower cover, and extract the phenophases of first, peak and end flowering. Random forests (RF) was performed to recognize flower pixels to calculate the flower cover, construct the flower cover curve and extract the same phenophases as those of the YOLO methods. Furthermore, flowering phenophases were also extracted through manual visual identification. We used a generalized additive model (GAM) to fit curves for flower count and flower cover, and extracted flowering phenophases by calculating the inflection points of the fitted curves. We found that (1) YOLO‐based methods could effectively identify flowers, and the variation in flower count and flower cover obtained from the YOLO object detection and semantic segmentation models reflected the trend of flowering phenology. The flower count and flower cover curves effectively supported the extraction of first and peak flowering. The difference between the YOLO‐identified and manually identified flowering phenophases ranged from 1 day to 3 days using the optimal thresholds. For end flowering, except for the end flowering identified based on flower count derived from YOLO object detection, the date difference in phenophases between the YOLO‐identified and manually identified ranged from 1 day to 8 days. (2) There are apparent outliers in the RF‐calculated flower cover values, particularly during the post‐peak‐flowering period. However, the identified flowering phenophases based on the RF‐derived flower cover curve after omitting outliers were consistent with those of manual visual identification and YOLO‐based methods (except end flowering identified based on flower count derived from YOLO object detection), with the date difference in phenophases ranging from 0 to 8 days. (3) The GAM performed well in fitting the trends of the normalized cumulative flower count and flower cover. Using the threshold generated by second derivate method, the identified end flowering was close to that of “late flowering” stage identified by manual visual identification, and the date difference ranged from 0 to 6 days. (4) Due to the variation in flowering rhythm and progression across different plant species, fixed thresholds are not fully optimal for all plants, and the thresholds used to extract flowering phenology require targeted adjustments based on specific observed species. Our study showed that a time‐lapse digital camera combined with ML algorithms can help improve the objectivity of phenology observations, indicating the possibility of using ML algorithms to identify flowering phenology.
SN - 2056-3485
ER -