Abstract

Accurate estimation of harvested paddy area plays a fundamental role in the development of effective food security policies, as these estimates are derived from regular monitoring of paddy phenological phases on a monthly basis. Reliable information on crop growth dynamics not only informs national production forecasting but also supports decision-making in relation to regional self-sufficiency, trade, and food supply stability. Previous work undertaken by Statistics Indonesia (BPS), known as the “Mixed Method Project,” has demonstrated the potential of integrating Sentinel-1 synthetic aperture radar (SAR) imagery with machine learning methods to detect paddy phenological phases. This approach has opened opportunities for modernizing the agricultural statistics system toward a more cost-effective and scalable data collection framework.

Despite these advances, several challenges remain unresolved, with the most critical being classification errors and inconsistencies in the temporal sequence of predicted phenological phases. To overcome these limitations, this study introduces a refined workflow that emphasizes both pre-processing and post-processing strategies. In the pre-processing stage, self-organizing map (SOM) filtering is applied to generate more representative training samples by reducing noise in the original SAR signals. Complementary to this, data augmentation techniques based on over- and undersampling are introduced to mitigate the imbalance commonly observed in phenological class distributions, thereby improving model generalization. In the post-processing stage, classification outputs are adjusted using prior information from previously predicted phases, ensuring logical continuity of crop development and reducing phase-switching errors.

To evaluate performance, the proposed framework is validated through two complementary approaches: (1) accuracy comparison against baseline machine learning models without the enhanced workflow, and (2) phase estimation using dot sampling analysis cross-checked with published references. Results from both validation steps demonstrate notable improvements in accuracy and consistency, thereby confirming the effectiveness of the method. Overall, the proposed workflow enhances the reliability of paddy monitoring systems and provides a robust foundation for advancing data-driven food security assessment and policy formulation.