军事知识服务平台
基于FPL-DEM的吉林省大豆产量预测模型构建
Soybean Yield Forecasting Model in Jilin Province Based on Fluctuation Perception Layered-Dynamic Ensemble Model
摘要 针对吉林大豆在东—中—西地形梯度下呈现的空间异质性,以及极端气候背景下产量预测中年际波动强、极端年份样本稀缺而易引发模型偏差的问题,以吉林省 35 个大豆主产市县 2005-2023 年的产量为目标变量,同期 16个气象因子、土壤温湿度及归一化植被指数(Normalized Difference Vegetation Index,NDVI)为特征.结合大豆生育期对显著偏移特征进行校正,根据产量波动异质性,经三级样本分层、特征增强筛选 Top80 关键特征,构建 RF、XGBoost和 GBR 共3 个基础模型.经 Optuna 优化与 TimeSeriesSplit 验证,构建嵌入高误差样本修正的波动感知分层—动态集成模型(FPL-DEM),以 R2、RMSE 和 RE 作为评价指标.动态集成模型预测效果显著优于基模型:2022-2023 年预测集 R2 达到0.947 4,逐年平均 RE 绝对值为10.963%;2019-2023 年预测集 R2 达到0.931 3,逐年平均 RE 绝对值为13.333 5%.生育期校正特征偏移精准识别关键因子,三级分层与动态集成提升模型异质样本适配性.该模型适用于吉林省大豆产量预测,赋能区域农业管理与种植主体决策,助力大豆稳产增收及粮食安全保障.
Abstract:
To address spatial heterogeneity in soybean production across Jilin Province along the east-central-west topographic gradient,and the pronounced interannual yield variability under extreme climatic conditions where the limited representation of extreme-year observations can induce prediction bias.This study used the yields of 35 main soybean producing cities and counties in Jilin Province from 2005 to 2023 as target variables,and 16 meteorological factors,soil temperature and humidity and Normalized Difference Vegetation Index(NDVI)as characteristics during the same period.Combined with the significant offset characteristics of soybean growth period,according to the heterogeneity of yield fluctuation,Top80 key features were screened by three-level sample stratification and feature enhancement,and three base models of RF,XGBoost and GBR were constructed.After Optuna optimisation and TimeSeriesSplit validation,the fluctuation perception layered-dynamic(FPL-DEM)ensemble model embedded with high error sample correction was constructed,and R2,RMSE and RE were used as evaluation indexes.The prediction effect of the dynamic integration model is significantly better than that of the base model:the prediction set R2 reaches 0.947 4 in 2022-2023,and the yearly average RE absolute value was 10.963%;the prediction set R2 reached 0.931 3 in 2019-2023,and the yearly average RE absolute value was 13.333 5%.The fertility correction feature bias accurately identified the key factors,and the three-level stratification and dynamic integration improved the adaptability of the model to heterogeneous samples.The model is suitable for soybean yield prediction in Jilin Province,which can empower regional agricultural management and planting decision-making,and help to ensure stable soybean yield and food security.
刊 名 大豆科学 ISTICPKU
年,卷(期): 2026, 45(5)
机标分类号: TE328 S513 S165.27
在线出版日期: 2026年10月12日
基金项目: SJ2025008-5:吉林省三江实验室科技创新项目,