Integration of multi-temporal sentinel-1 and sentinel-2 data for paddy rice crop height estimation and uncertainty assessment using quantile regression forests

Abstract

PurposeThis study estimates rice crop height using multi-temporal Sentinel-1 and Sentinel-2 data, collected alongside field measurements conducted in T & uuml;rkiye and Bulgaria during 2023 and 2024.MethodTo evaluate the efficacy of the datasets in estimating the rice height, we developed three Quantile Regression Forest (QRF) models. The QRF, an extension of Random Forest Regression (RFR), provides conditional quantiles for epistemic uncertainty estimation. Specifically, the first model (M1) utilized Sentinel-1 dual-polarimetric data, their ratios, and Radar Vegetation Index. The second model (M2) incorporated Sentinel-2 spectral bands and a range of spectral indices, while the third model (M3) combined Sentinel-1 and Sentinel-2 data. To address variability in flooding and drainage periods across growth stages and management practices, the models were trained and evaluated using the complete, flooded, and non-flooded datasets.ResultsThe results indicated that M3 yielded the most accurate predictions, with an RMSE of 12.35 cm on the flooded test dataset. Notably, the models with the flooded datasets generally exhibited lower uncertainty and more consistent predictions. However, all models struggled with underestimations for heights exceeding 100 cm, indicating limited predictive capability for extreme values. Interestingly, while M1 and M2 models showed different results for complete and flooded datasets, the M3 model gave similar results for both conditions, which is a practical advantage that eliminates the need to distinguish the flooded and non-flooded samples.ConclusionIn summary, combining radar and optical data with machine-learning improves rice height estimation, while uncertainty estimates enhance reliability for agricultural and environmental applications.