Maize Genomics and Genetics 2025, Vol.16, No.5, 239-250 http://cropscipublisher.com/index.php/mgg 244 drought (Vergopolan et al., 2020; Wang et al., 2023). Furthermore, some hybrid methods are also being used, such as combining biophysical crop models with ML, training the models with data across years and regions, and the resulting predictions often hold water in different drought scenarios (Attia et al., 2022; Wang et al., 2022; Li et al., 2023). 5.3 Evaluation of prediction performance under drought It is not enough to merely judge whether a model runs accurately in a certain year or a certain place. A truly reliable model needs to be able to stand the test of many years, multiple regions, and various drought intensities. This is also why many current studies are conducting cross-environment model evaluations. Some models use genetic algorithms or feature selection to optimize their structure, and the prediction results are indeed quite impressive. For instance, the predicted R²for production can reach 0.92, and the resilience index (such as STI) also has a level of 0.82, which is already quite good. The introduction of drought-specific factors such as SIF, CDI, and soil moisture has also helped the model maintain high stability in extreme years (Shuai and Basso, 2022; Luo et al., 2024). However, to be fair, not all models can handle extreme situations perfectly. Some models still have the problem of "underestimating production loss due to extreme drought", which is quite likely to be exposed in practice (Amiri et al., 2022; Bueechi et al., 2023). This indicates that we still need to continue to strengthen the mechanism construction of the model in terms of drought response. 6 Case Study: Integrated Prediction System for Drought-Tolerant Maize in Sub-Saharan Africa 6.1 Background and objectives Sub-saharan Africa (SSA) is not short of sunlight or arable land, but the food problem has never been solved. The reasons are very complex. Among them, drought is the most direct and common limiting factor affecting corn yield. Especially in some areas where water resources are already tight, poor harvests for several consecutive seasons are the norm. Therefore, drought-resistant corn varieties, along with a reliable yield prediction system, become particularly crucial. The SSA region is promoting an integrated prediction framework, with a clear goal: to integrate genomic, remote sensing, meteorological and environmental data and use machine learning to predict corn yields under drought conditions (Ndlovu et al., 2024). This not only serves breeding projects but also aims to provide some more practical reference information for farmers and policymakers. 6.2 Model development and integration Not all models can be truly implemented. Many solutions run very fast in the laboratory, but once they are put into practical application, they fail to adapt to the local environment. But what the SSA team did this time was a bit different. They did not start from the top-level design but from the local demands, prioritizing the issue of "locality" first. One of the core data sources of the model is Earth Observation (EO) data. Such as rainfall, water availability, extreme temperatures, number of drought days... All these indicators have been integrated into the system, and the time scale has been refined to the sub-month level. Don't underestimate this detail. The significant improvement in prediction performance is largely supported by it. As for the genetic aspect, they used the RR-BLUP model in the multi-environment corn experiment and identified many key quantitative trait nucleotides (QTN) and candidate genes related to drought resistance. Furthermore, they also incorporated the results of GWAS for verification, which helps improve prediction accuracy and makes the model output more biologically explanatory (Amadu et al., 2025). It is worth mentioning that they did not stop at making predictions. Instead, they used big data platforms and spatial modeling to precisely "deliver" these drought-resistant materials to high-risk drought areas. In this way, the materials cultivated will not be "selected well but not put to use". 6.3 Outcomes and implications Ultimately, whether a system is worth promoting depends on whether it can truly be "put into use". At least in terms of on-site performance, this system of the SSA team has passed the test. Just in terms of yield prediction, the model's accuracy during the corn growing season is quite good, with a Nash-Sutcliffe efficiency value exceeding 0.6 and an average relative error controlled within 20% (Lee et al., 2022). This level of accuracy is not only sufficient for research but can also be used for early warning and making decisions on grain dispatching. In terms
RkJQdWJsaXNoZXIy MjQ4ODYzNA==