Transferability of spatial and temporal learning models for winter wheat mapping in data-scarce environments: A case study in Armenia

Winter wheat constitutes a key component of Armenia’s agricultural economy and national food security. Accurate and timely identification and mapping of winter wheat fields can support informed policymaking, infrastructure planning, resource allocation, and crop monitoring. However, large-scale crop mapping using traditional field-based surveys is labor-intensive, spatially limited, and often impractical. In this study, we present a machine learning framework for large-scale winter wheat field mapping in Armenia under limited ground truth availability using multi-source satellite images. We evaluate the possibility of cross-region generalization through image-based semantic segmentation models (3D Unet and SegNet) and pointwise temporal classification models (Random Forest, one-dimensional convolutional neural network, and long short-term memory (LSTM)) using multi-temporal Sentinel-2 and PlanetScope images. We train these models using large, well-labeled winter wheat datasets from United States counties and a limited set of known winter wheat fields in Armenia, and subsequently transfer and test across five independent test regions in Armenia. Models trained on Sentinel-2 imagery generalize well within the United States, achieving test accuracy and F1-scores of 0.96 and 0.86, respectively. However, their performance degrades when transferred to fragmented agricultural landscapes in Armenia, particularly for image-based semantic segmentation models. In contrast, the LSTM-based temporal model demonstrates superior transferability, achieving accuracy and F1-scores of 0.96 and 0.96, respectively, while effectively suppressing non-crop features and identifying both known and previously unmapped winter wheat fields. Using the best-performing model, we generate provincial-scale winter wheat maps for Shirak Province for the years 2023, 2024, and 2025, estimating cultivated areas of 25,641 ha, 21,088 ha (approximately 3% higher than the USDA Foreign Agricultural Service estimate), and 28,909 ha, respectively. These results highlight the potential of temporal machine learning models for scalable winter wheat mapping in data-scarce regions, offering a practical pathway toward nationwide crop inventory generation and agricultural decision support.