Sentinel-1 SAR time series for forest phenology: multimodal deep learning with meteorological fusion across deciduous broadleaf and evergreen needleleaf forests
Abstract
Forest phenology is essential for evaluating the health and productivity of forest ecosystems and their responses to climate change. Vegetation indices derived from optical remote sensing are widely used to monitor large-scale forest phenology, but cloud contamination creates gaps during key transitions. Synthetic aperture radar (SAR) offers cloud-penetrating observations of the earth surface, yet its ability to retrieve forest phenology across forest types remains insufficiently evaluated. Here, we developed a multimodal deep learning (MDL) framework that fuses Sentinel-1 time series with meteorological variables (temperature, soil moisture, radiation and precipitation) to estimate the start (SOS) and end (EOS) of the growing season in deciduous broadleaf (DB) and evergreen needleleaf (EN) forests. Four architectures, including Convolutional Neural Network-CNN, Temporal Convolutional Network-TCN, Transformer, and U-Net were evaluated under Sentinel-1-only and multimodal settings using PhenoCam-derived reference dates across North America and Europe. Under leave-one-year-out temporal cross validation, Sentinel-1-only models retrieved phenology with moderate skill, with U-Net and Transformer generally providing the strongest performance. Meteorological fusion further improved retrievals most clearly in DB forests, where the best MDL model achieved RMSEs of 9.5 d for SOS and 10.8 d for EOS. Improvements were more modest in EN forests, particularly for EOS, reflecting weaker seasonal SAR signals in evergreen canopies. These results show that deep learning markedly improves SAR-based phenology retrieval compare with conventional threshold method, and that integrating meteorological information can further enhance accuracy, particularly in DB forests. The framework facilitates operational phenology monitoring in persistently cloudy regions where optical imagery is often unavailable.
Related
- Divergent mercury accumulation dynamics in tropical dry and moist broadleaf forests
- Litter VOC emissions inventory from deciduous urban trees in the mediterranean region
- Functional traits and acclimation of Central European deciduous forests during the Piacenzian
- Decoupled carbon assimilation and growth responses to aridity in temperate deciduous oaks
- Atmospheric dryness effects on canopy chlorophyll fluorescence and Gross Primary Production (GPP) in a deciduous forest during heat waves
BibTeX
@article{xiao2026sentinel1,
title = {{Sentinel-1 SAR time series for forest phenology: multimodal deep learning with meteorological fusion across deciduous broadleaf and evergreen needleleaf forests}},
author = {Yi Xiao and Frédéric Frappart and Yongshuo Fu and Jean-Pierre Wigneron and Marc Peaucelle},
journal = {Agricultural and Forest Meteorology},
year = {2026},
month = oct,
doi = {10.1016/j.agrformet.2026.111373},
url = {https://hal.science/hal-05781506v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Sentinel-1 SAR time series for forest phenology: multimodal deep learning with meteorological fusion across deciduous broadleaf and evergreen needleleaf forests
AU - Yi Xiao
AU - Frédéric Frappart
AU - Yongshuo Fu
AU - Jean-Pierre Wigneron
AU - Marc Peaucelle
PY - 2026
DA - 2026/10/07
JO - Agricultural and Forest Meteorology
DO - 10.1016/j.agrformet.2026.111373
UR - https://hal.science/hal-05781506v1
AB - Forest phenology is essential for evaluating the health and productivity of forest ecosystems and their responses to climate change. Vegetation indices derived from optical remote sensing are widely used to monitor large-scale forest phenology, but cloud contamination creates gaps during key transitions. Synthetic aperture radar (SAR) offers cloud-penetrating observations of the earth surface, yet its ability to retrieve forest phenology across forest types remains insufficiently evaluated. Here, we developed a multimodal deep learning (MDL) framework that fuses Sentinel-1 time series with meteorological variables (temperature, soil moisture, radiation and precipitation) to estimate the start (SOS) and end (EOS) of the growing season in deciduous broadleaf (DB) and evergreen needleleaf (EN) forests. Four architectures, including Convolutional Neural Network-CNN, Temporal Convolutional Network-TCN, Transformer, and U-Net were evaluated under Sentinel-1-only and multimodal settings using PhenoCam-derived reference dates across North America and Europe. Under leave-one-year-out temporal cross validation, Sentinel-1-only models retrieved phenology with moderate skill, with U-Net and Transformer generally providing the strongest performance. Meteorological fusion further improved retrievals most clearly in DB forests, where the best MDL model achieved RMSEs of 9.5 d for SOS and 10.8 d for EOS. Improvements were more modest in EN forests, particularly for EOS, reflecting weaker seasonal SAR signals in evergreen canopies. These results show that deep learning markedly improves SAR-based phenology retrieval compare with conventional threshold method, and that integrating meteorological information can further enhance accuracy, particularly in DB forests. The framework facilitates operational phenology monitoring in persistently cloudy regions where optical imagery is often unavailable.
SN - 0168-1923
ER -