Hcropland30: A hybrid 30-m global cropland map generated by leveraging global land cover products and Landsat data via deep learning

Hu, Q., Cai, Z., You, L., Fritz, S. ORCID: https://orcid.org/0000-0003-0420-8549, Zhang, X., Yin, H., Wei, H., Yang, J., Li, Z., Yu, Q., Wu, H., Xu, B., & Wu, W. (2026). Hcropland30: A hybrid 30-m global cropland map generated by leveraging global land cover products and Landsat data via deep learning. Remote Sensing of Environment 344 e115552. 10.1016/j.rse.2026.115552.

Full text not available from this repository.

Abstract

While global cropland mapping has seen notable improvements over the past decade, with several 10–30 m global land use and land cover (LULC) products now available, these products often suffer from considerable classification errors and spatial inconsistencies in cropland information. In this study, we developed a hybrid 30 m global cropland map for 2020, namely Hcropland30, to advance cropland mapping at a global scale. We started our approach with a hierarchical sampling strategy that used the simulated annealing method to select representative 1° × 1° grids globally and identify point-level samples within these grids, considering factors like cropland heterogeneity, spectral feature, climate, and topography. We then employed an ensemble learning technique that integrated multiple classifiers, including support vector machine (SVM), random forest (RF), multi-layer perceptron (MLP), and gradient boosted decision tree (GBDT), to expand these sparse point-level samples into area-level pseudo-labels. These pseudo-labels were used to train a U-Net model for predicting global cropland, using Landsat-derived seasonal NDVI composites and six LULC datasets as inputs. Our results revealed considerable spatial discrepancies among existing 10–30 m LULC products, with areas exhibiting ‘high’ to ‘very high’ inconsistency significantly outnumbering those of ‘medium’ to ‘low’ inconsistency. Our Hcropland30 map achieved an overall accuracy of 89.8% and a cropland F1-score of 0.85, higher than other LULC datasets, especially in regions with high landscape heterogeneity. Although LULC datasets provide useful prior information, the proposed framework is robust to variations in the number, type, and combination of LULC inputs. Furthermore, given the availability of historic Landsat data and LULC datasets, this methodology can be readily extended to create multi-temporal Hcropland30 maps and adapted for global mapping of other land cover types. Overall, this study not only generated a high-accuracy global cropland dataset but also introduced a robust methodology for diverse global mapping applications by leveraging deep learning and varied prior knowledge.

Item Type:	Article
Uncontrolled Keywords:	Global cropland mapping; Hierarchical sampling; Ensemble learning; Spatial inconsistency; Deep learning
Research Programs:	Advancing Systems Analysis (ASA) Advancing Systems Analysis (ASA) > Novel Data Ecosystems for Sustainability (NODES) Strategic Initiatives (SI)
Depositing User:	Luke Kirwan
Date Deposited:	10 Jul 2026 07:49
Last Modified:	10 Jul 2026 07:49
URI:	https://pure.iiasa.ac.at/21719

Actions (login required)

View Item