Ascertain the construction and sampling strategy of the Zindi Africa landslide detection dataset

Ascertain the dataset construction process and sampling strategy used to create the Zindi Africa "Classification for Landslide Detection" competition dataset employed in this study, including how image patches were selected and labeled for training and testing.

Background

The study relies on a patch-based, multi-modal dataset provided through the Zindi Africa competition, which lacks key metadata and documentation. While the dataset enables benchmarking of multi-modal landslide detection models, the absence of detailed information on how samples were constructed and selected limits deeper analysis and fair assessment of generalization and operational performance.

Clarifying the dataset’s construction and sampling protocol would enable reproducibility, facilitate comparison with other methods and datasets, and support robust evaluation under varied geographic and temporal conditions.

References

In addition, the dataset used in this study originates from a ML competition and was provided without accompanying metadata, such as acquisition dates, geographic locations, or detailed satellite product specifications, limiting deeper analysis of temporal and regional factors. Furthermore, the dataset construction process and sampling strategy remain unknown.

— Multi-Modal Landslide Detection from Sentinel-1 SAR and Sentinel-2 Optical Imagery Using Multi-Encoder Vision Transformers and Ensemble Learning  (2604.05959 - Nasios, 7 Apr 2026) in Subsection: Limitations