Projecting migration is challenging, due to the context-specific and discontinuous relations between migration and the socioeconomic and environmental conditions that drive this process. Here, we investigate the usefulness of Machine Learning (ML) Random Forest (RF) models to develop three net migration scenarios in South Asia by 2050 based on historical patterns (2001–2019). The model for the direction of net migration reaches an accuracy of 75%, while the model for the magnitude of migration in percentage reaches an R value of 0.44. The variable importance is similar for both models: temperature and built-up land are of primary importance for explaining net migration, aligning with previous research. In all scenarios we find hotspots of in-migration North-western India and hotspots of out-migration in eastern and northern India, parts of Nepal and Sri Lanka, but with disparities across scenarios in other areas. These disparities underscore the challenge of obtaining consistent results from different approaches, which complicates drawing firm conclusions about future migration trajectories. We argue that the application of multi-model approaches is a useful avenue to project future migration dynamics, and to gain insights into the uncertainty and range of plausible outcomes of these processes.

中文翻译：

---

环境和社会经济变化中南亚移民模式的情景预测


由于移民与推动这一过程的社会经济和环境条件之间存在特定背景和不连续的关系，因此预测移民具有挑战性。在这里，我们研究了机器学习 (ML) 随机森林 (RF) 模型的有用性，以根据历史模式（2001-2019 年）制定 2050 年南亚的三种净移民情景。净迁移方向模型的准确度达到 75%，而迁移幅度百分比模型的 R 值达到 0.44。两个模型的变量重要性相似：温度和建筑用地对于解释净迁移至关重要，这与之前的研究一致。在所有情景中，我们都发现了印度西北部的迁入热点以及印度东部和北部、尼泊尔和斯里兰卡部分地区的迁出热点，但其他地区的情景之间存在差异。这些差异凸显了从不同方法获得一致结果的挑战，这使得对未来移民轨迹得出明确的结论变得复杂。我们认为，多模型方法的应用是预测未来移民动态并深入了解这些过程的不确定性和合理结果范围的有用途径。