Answers are grounded in this paper and cite their sources.
Continuous multiyear crop mapping is essential for understanding agricultural dynamics under climate change, yet its operational scale-up is fundamentally bottlenecked by the reliance on extensive, year-specific field sampling. In order to address this issue, this study suggests an innovative multiyear crop mapping framework that integrates the concept of temporal model transfer with automated sample purification system to reduce the dependence on newly collected target-year ground truth data during the training procedure. The framework will have its main innovation of isolating high-confidence samples out of the existing 30-m historical crop maps through spatial filtering and Isolation Forests to train an initial 10-m resolution model. This model is then directly transferred to subsequent years to generate initial maps, from which temporally updated, pure samples are autonomously derived to retrain the classifier for the target year. Based on the Sentinel-2 time-series images, we firstly compared five classification models, which concluded the best base model used for interannual feature extraction was 1D-CNN. Sample purification strategy has been tested in two different types of agriculture settings Wangkui, China, and Emmet, USA, in three successive years (2021–2023) and demonstrated good performance in reducing the accuracy loss that occurs during the direct model transfer, producing highly accurate crop maps (Overall Accuracy: 0.9534–0.9725; Macro-F1: 0.9520–0.9697). This solution provides a scalable and cost-effective approach for continuous crop monitoring, supports the analysis and management of crop rotation patterns.
No takes yet. Share an insight, caveat, or question.
Feng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: