From time-series generation and model selection to transfer learning: A review and comparative analysis of pixel-wise approaches for large-scale crop mapping

Accurate and timely large-scale crop mapping is essential for crop management and food security. Over the past decade, a wide range of pixel-wise crop classification approaches have emerged for workflow elements including image preprocessing, sampling methods, feature selection, model architectures, and transfer learning strategies. To guide users in constructing robust workflows, we reviewed methods for supervised and transfer-learning-based workflow elements and propose selecting a crop-mapping workflow based on target-site sample size and domain shift from a sample-abundant source site. To comprehensively evaluate the reviewed approaches, we systematically analyzed how different time-series construction strategies, training sample sizes, variable sets, model architectures, and transfer-learning approaches affect performance for large-scale crop classification. The literature review and comparative analysis led to three key insights. First, in our multi-site experiments, generating time series with 7-day linear resampling combined with Transformer models consistently yielded the highest accuracies for both fully supervised and transfer-learning workflows, while Random Forests offered competitive accuracy with substantially faster training. Second, transfer learning increased workflow adaptability: unsupervised domain adaptation was effective for homogeneous crop classes, while fine-tuning was more robust across diverse transfer scenarios. Third, workflow selection depended strongly on the availability of labeled samples: below a certain ground-truth sample threshold, transfer-learning workflows were more viable for reliable crop mapping, whereas fully supervised workflows became preferable once sufficient labeled data were available. To support transparency and reproducibility, we provide all code, configurations, and results in an open GitHub repository: A review of pixel-wise approaches for large-scale crop mapping.