Areej Alwahas, Kasper Johansen, Jorge Rodriguez, Matthew F McCabe
The configuration combining self-supervised pretraining, pseudo-label augmentation, and optical-radar inputs achieved an mIoU of 0.80, an F1-score of 0.88, and an overall accuracy of 0.97. In contrast, using optical information alone substantially reduced accuracy, with an mIoU of 0.35, an F1-score of 0.32, and an overall accuracy of 0.64. Fallow and forage produced the highest mapping accuracies, with mIoU values of 0.97 and 0.85, respectively, while vegetables had the lowest accuracy among the four crop groups, with an mIoU of 0.52.
INTRODUCTION: In arid and hyper-arid regions, agriculture depends heavily on irrigation, making crop type monitoring important for water allocation, monitoring crop management policies, and providing the information required to forecast food supply. However, field labels are often scarce, and crop calendars can shift due to locally managed planting, harvest, and irrigation decisions, complicating mapping at field-scale.
METHODS: We present a seasonal crop type mapping approach applied to Wadi Al-Dawasir, Saudi Arabia, generating maps for 2020-2024 from biweekly optical and radar satellite time series. The method learns representations from unlabeled imagery through self-supervised pretraining and fine-tunes a segmentation model using a small set of field observations with pseudo-label augmentation from unsupervised clustering. We mapped four classes: fallow, cereal, vegetables, and forage, and evaluated performance using overall accuracy, F1-score, and mean intersection-over-union (mIoU).
RESULTS: The configuration combining self-supervised pretraining, pseudo-label augmentation, and optical-radar inputs achieved an mIoU of 0.80, an F1-score of 0.88, and an overall accuracy of 0.97. In contrast, using optical information alone substantially reduced accuracy, with an mIoU of 0.35, an F1-score of 0.32, and an overall accuracy of 0.64. Fallow and forage produced the highest mapping accuracies, with mIoU values of 0.97 and 0.85, respectively, while vegetables had the lowest accuracy among the four crop groups, with an mIoU of 0.52.
DISCUSSION: These results show that self-supervised temporal pretraining combined with pseudo-label augmentation can support efficient multi-season, field-scale crop mapping using limited labels in irrigation-driven arid regions. The lower accuracy for vegetables highlights the continued challenge of mapping heterogeneous crop groups with overlapping phenological patterns.