A forecasting model for desert locust presence during recession period, using real-time satellite imagery

Software and data availability section: All the raw datasets used in this study were downloaded from open FAO locustHub database available upon request on FAO Locust Monitoring Data. Remote sensing images are available on the earth data NASA platform upon the creation of an account: urs.earthdata.nasa.gov. The precipitation data were obtained from the International Research Institute for Climate and Society available on iridl.ldeo.columbia.edu and the soil sand content map from the International Soil Reference and Information Center can be downloaded on files.isric.org. The automatic downloading of satellite images and forecasting system along with the scripts needed to replicate the model training are available in the following dataverse directory: doi.org. The study was conducted using R 4.1.2 (R Core Team, 2022) and Python.Corrigendum to: This article has been updated: doi.org International audience Highlights: • We built an operational forecasting system for Desert locust preventive management. • We used random forest model for real-time forecasting of locust presence and update every decade. • Pest distribution was explained by sand cover, ecoregions, temperature, precipitations and vegetation cover. • Field evaluation revealed a strong correlation between predicted probabilities and observed locust densities.Abstract: Desert locust (Schistocerca gregaria) is a major agricultural pest that poses significant socioeconomic challenges to food security. This study aims to enhance preventive management of desert locusts in Western and Northern Africa by improving an operational model developed by Piou et al. (2019). The model employs satellite remote sensing data and machine learning to forecast locust occurrence at a 1 km 2 resolution every ten days. Objectives include identifying environmental risk factors, training random forest models with high-predictive power and providing updated forecasts via a web interface. It is the first implementation of a statistical forecasting model for this species within an automated system, delivering updated locust presence probabilities every ten days. Validated through field surveys with a positive error rate of 23%, the forecasting tool shows a strong correlation between predicted probabilities and observed locust densities. This operational tool can guide survey teams, optimize resource allocation, and mitigate environmental impacts efficiently. We believe continuous evaluation and integration of the forecast system will enhance its effectiveness in preventing locust outbreaks, thereby safeguarding food security in the region.