A generalized model for mapping sunflower areas using Sentinel-1 SAR data

Abstract

Existing crop mapping models, rely heavily on reference (calibration) data obtained from remote sensing observations. However, the transferability of such models in space and time, without the need for additional extensive datasets remains a significant challenge. There is still a large gap in developing generalized classification models capable of mapping specific or multiple crops with minimal calibration data. In this study, we present a generalized automatic approach for sunflower mapping at 20-m spatial resolution, using the C-band Sentinel-1 (S1) synthetic aperture radar (SAR) data driven by previously developed phenological metrics. These metrics characterize the directional behavior of the sunflower head, capturing distinct backscattering responses in SAR data acquired from ascending and descending orbits. Specifically, we utilize SAR-derived backscatter values in VH and VV polarization, as well as their ratio VH/VV, as input features to a random forest classifier that was calibrated for the year 2022 in Ukraine. This model is further directly applied to selected sites for multiple years in Ukraine (generalization in time) and other major sunflower producing countries (generalization in space): Ukraine for 2018–2020, and Hungary, France, Russia and USA for 2018. Our results reveal that the model based on features acquired from descending orbits outperforms its ascending orbit counterpart because of the directional behavior of sunflower: user's accuracy (UA) of 96%, producer's accuracy (PA) of 97% and F-score of 97% (descending) compared to UA of 90%, PA of 89% and F-score of 90% (ascending). When generalized to other years and countries, our model achieves an F-score exceeding 77% for all cases, with the highest F-scores (>91%) observed in Ukraine and Russia sites and the lowest (77%) for the US site. We further utilize the produced maps (pixel-based) for the selected regions and years to estimate sunflower planted areas using a statistical sampling-based approach. Our estimates yield the relative root mean square error (RMSE) as 19.7% of the mean area, when compared to the reference data from official statistics and reference maps. These findings demonstrate the robustness of our proposed approach across space and time in generating accurate sunflower maps, its ability to mitigate cloud cover issues through spaceborne SAR data acquisitions, and its potential for obtaining estimates of sunflower planted areas. This research emphasizes the importance of developing interpretable and domain-specific machine learning models that can be readily extended to multiple geographical regions with little to no labelled datasets.