Time-series SAR scattering coefficients over woodland: trend analysis and explainable modelling

In the era of large models, massive amounts of Synthetic Aperture Radar (SAR) scattering data need to be synthesized to meet the demand for interpretation training, which calls for clear temporal patterns of time-series SAR for sequence generation. However, the temporal evolution trends of SAR scattering coefficients have been neither comprehensively studied nor explicitly modelled. To address the issue, this paper takes the long-sequence temperate woodlands as the research object for analysis and explicit modelling, where the trend analysis provides explainable motivations for model design. Using Sentinel-1A ground range detected data with a 12-day revisit cycle, two SAR image sequences are constructed, each consists of VV or VH intensity images of 174 consecutive moments spanning from April 2019 to December 2024. By classifying geographically matched multi-temporal optical images through a fine-grained multi-scale convolutional neural network, the woodland area is identified, and 9.48 million VV/VH scattering coefficient sequences are extracted. The seasonal Mann-Kendall test evaluates the annual changes in scattering intensity, while seasonal-trend decomposition using LOESS provides seasonal patterns. Correlation analysis shows a high correlation between the average temperature and the average scattering intensity. Based on the analysis, a scattering intensity model is constructed using a modified Transformer network, which predicts scattering intensity sequences for woodlands. The evaluation of the synthetic sequence for year 2024 indicates minor deviation of the average intensity prediction, which confirms the effective modelling and the necessary analysis.