MRH-ViT Desert Locust Movement Prediction Dataset: Multi-Resolution Heterogeneous Vision Transformer with 59-Band Multi-Spectral Satellite Imagery for Ethiopia (2015-2023)

This dataset supports research on desert locust movement prediction using deep learning and remote sensing. It contains 18,871 georeferenced multi-spectral satellite observations from Ethiopia spanning 2015-2023, compiled from FAO Desert Locust Information Service reports. Each sample comprises 59 environmental bands (41×41 pixels at 250m resolution, covering ~10×10 km) including vegetation indices (NDVI, EVI, VHI, NDWI), climate data (LST, TCI, TVDI, precipitation from CHIRPS), soil properties (moisture from SMAP, texture, sand content), atmospheric conditions (u/v wind components from ERA5), evapotranspiration (MOD16A2), and static terrain features (elevation, slope, aspect, land cover). The dataset includes 12,114 locust presence and 6,757 absence records, processed with temporal lags (30, 60, 90 days) to capture environmental conditions influencing locust development and migration patterns.

The dataset was developed for the MRH-ViT (Multi-Resolution Heterogeneous Vision Transformer) model, achieving 82.87% accuracy, 86.09% precision, 87.47% recall, 86.77% F1-score, and 89.01% AUC-ROC in predicting desert locust presence. This represents a 6.4% improvement in F1-score over existing state-of-the-art methods. The data supports early warning systems for agricultural protection in the Horn of Africa, where locust swarms threaten food security for vulnerable farming populations.

Data extraction utilized Google Earth Engine with custom Python pipelines (available at github.com), integrating data from MODIS, Landsat, Sentinel-2, SMAP, SRTM, CHIRPS, and ERA5. The dataset is suitable for research in computer vision, ecological modeling, remote sensing applications, and transformer-based spatiotemporal prediction. This dataset was created as part of a Master of Science thesis at Addis Ababa University under the supervision of Dr. Solomon Gizaw. The research addresses critical agricultural threats from desert locust swarms, which affected over 800,000 farming households during the 2020 East Africa outbreak.

Data processing implemented robust error handling for infrastructure challenges including power outages and connectivity issues. The extraction pipeline includes automatic restart capabilities, progress tracking, and comprehensive logging mechanisms.

The dataset maintains the natural class imbalance (1.79:1 presence to absence ratio) observed in FAO reports to reflect realistic operational conditions. Stratified random splitting (70:15:15) was used for training, validation, and test sets.

GitHub repository with data extraction tools and model implementation: github.com

For questions or collaborations, please contact the author.