LAUSR

Abstract Quantitatively evaluating the potential marginal lands for cassava (Manihot esculenta Crantz) in previous studies has mostly been conducted at regional and national scales. However, at present, there is no reliable cassava marginal land assessment data on the global scale. By explicitly incorporating known samples with multicovariates, a boosted regression tree (BRT) modeling procedure was used to generate the first global potential marginal land distribution map of cassava with a 5 × 5 km spatial resolution grid. The total amount of potential marginal land resources suitable for cassava is approximately 1696.28 million hectares, mainly distributed across Central Africa, South America, Southeast Asia and North America. The estimated map reveals that the ecological zones suitable for cassava lie between the 30° parallels and are even distributed at higher latitudes. Validation statistics indicate that the fitted BRT models obtain high performance with a tenfold cross-validation area under the curve (AUC) value of 0.962 [95% confidence interval (CI) 0.957–0.967]. The investigated patterns show that the relative contribution of land cover is 46.47% [95% CI 43.46–49.49%], followed by annual cumulative precipitation (35.96% [95% CI 33.29–38.62%]), minimum annual temperature (10.23% [95% CI 9.61–10.85%]), mean solar radiation (3.13% [95% CI 2.77–3.49%]) and maximum annual temperature (2.31% [95% CI 2.15–2.47%]). The results provide a baseline for understanding the potential distribution of marginal land resources viable for cassava worldwide.