Soil moisture is a fundamental climate variable sustaining the terrestrial biosphere. Whereas flood-recharged soil moisture (FRSM) is an important input flux in semi-arid floodplain ecosystems, its spatio-temporal dynamics is not fully understood due to lack of adequate field data. While existing active remotely sensing data are valuable to understand soil moisture, a trade-off between high temporal against coarse spatial resolutions limit their utility at local scales. In this study, we extracted linear backscatter coefficient sigma nought (σ ₀) and time series data from 91 pre-processed 10 m multi-temporal dual Sentinel-1 images. The data was collected from both inside and outside the flooded zone in a semi-arid area in northern Zimbabwe. To characterize FRSM anomaly, lag and memory, we built a hybrid deep learning long short-term memory autoencoder (LSTMAE) model based on a recent flood event which was subsequently evaluated using mean absolute error (MAE) and root mean squared error (RMSE) loss metrics using an independent validation dataset. Validation results showed that both VV and VH-polarized data effectively detected FRSM positive anomaly with very small MAE (0.0799σ ₀; 0.0191σ ₀) and small RMSE (0.0967σ ₀, 0.0250σ ₀) respectively. In the flood zone, the LSTMAE model detected three positive anomalies for both polarizations. Also, our study established that the VV LSTMAE model was effective in detecting subtle positive anomalies while VH depicted the longest lag and memory at a local scale. The study concludes that the extraction of σ ₀ on Sentinel-1 time series data offers a good understanding of localised FRSM characteristics within semi-arid floodplains.