1.College of Information,Xi’an University of Finance and Economics,Xi’an 710100,China
2.Intelligent Financial Collaborative Trusted Computing Key Laboratory of Shaanxi Province for Higher Education Institutions,Xi’an 710100,China
3.State Key Laboratory of Eco-hydraulics in Northwest Arid Region,Xi’an University of Technology,Xi’an 710048,China
4.School of Agricultural Engineering,Jiangsu University,Zhenjiang 212013,China
5.Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences,Chongqing 400714,China
Show less
文章历史+
Received
Accepted
Published
2026-05-07
2026-06-19
2026-09-01
Issue Date
2026-09-04
PDF (4209K)
摘要
针对西北干旱半干旱引黄灌区玉米产量受遥感、气象、土壤及灌溉环境等多源因素共同影响,变量间普遍存在共线性、冗余信息和潜在伪相关,限制了传统相关性驱动模型的预测稳定性与解释能力的问题。本研究构建了融合因果森林、时序卷积网络和时间注意力机制(Causal forest-temporal convolutional network-temporal attention,CF-TCN-TA)的估产方法。该方法首先利用因果森林估计多源变量对玉米产量的潜在影响,筛选出具有较强解释意义的关键驱动因子,以降低冗余变量和潜在伪相关对模型训练的干扰;进而将筛选后的关键变量输入时序卷积网络,提取玉米生长季内的时序变化特征,并通过时间注意力机制自适应增强关键月份的信息贡献。基于宁夏回族自治区2019—2023年县域玉米产量及多源遥感、气象和土壤数据开展验证,结果表明:CF-TCN-TA模型获得了最优预测性能,决定系数(Coefficient of determination,R²)为0.830 3,均方根误差(Root mean square error,RMSE)为882.54 kg/hm²,平均绝对误差(Mean absolute error,MAE)为706.60 kg/hm²,平均绝对百分比误差(Mean absolute percentage error,MAPE)为10.80%,整体优于传统机器学习模型和常规深度学习模型。因果效应评估结果显示,土壤pH、深层田间持水量、气温及光合有效辐射吸收比例(Fraction of photosynthetically active radiation,FPAR)、增强型植被指数(Enhanced vegetation index,EVI)等遥感指标是影响研究区县域玉米产量差异的重要驱动因子。因果启发式特征筛选与时序注意力建模的融合,能够在提高估产精度的同时增强模型解释性,可为西北干旱半干旱引黄灌区县域玉米产量监测与农业生产管理提供方法参考。
Abstract
In the arid and semi-arid Yellow River irrigation districts of northwestern China, maize yield is jointly influenced by multiple factors including remote sensing, meteorological, soil, and irrigation conditions. The widespread presence of multicollinearity, information redundancy, and potential spurious correlations among these variables limits the predictive stability and interpretability of conventional correlation-driven models. To address this issue, this study developed a framework integrating Causal forest, Temporal convolutional network, and Temporal attention mechanism (CF-TCN-TA) for county-level maize yield prediction. The framework first employs Causal forest to estimate the potential effects of multi-source variables on maize yield and to identify key driving factors with strong explanatory significance, thereby mitigating the impact of redundant variables and potential spurious correlations on model training. Subsequently, the selected key variables are fed into a Temporal convolutional network to extract temporal variation characteristics during the maize growing season, and a Temporal attention mechanism is incorporated to adaptively enhance the information contribution of critical months. Validation was conducted using county-level maize yield data and multi-source remote sensing, meteorological, and soil data from the Ningxia Hui Autonomous Region from 2019 to 2023. The results indicated that the CF-TCN-TA model achieved the optimal predictive performance, with a coefficient of determination (R2) of 0.830 3, a root mean square error (RMSE) of 882.54 kg/hm², a mean absolute error (MAE) of 706.60 kg/hm², and a mean absolute percentage error (MAPE) of 10.80%, consistently outperforming conventional machine learning models and standard deep learning models. Causal effect evaluation further revealed that soil pH, deep-layer field water-holding capacity, air temperature, and remote sensing indicators such as the fraction of photosynthetically active radiation (FPAR) and the enhanced vegetation index (EVI) were among the important driving factors influencing spatial variations in county-level maize yield within the study area. The integration of causal-driven feature selection and temporal attention modeling can enhance model interpretability while improving yield estimation accuracy, offering a methodological reference for county-level maize yield monitoring and agricultural production management in the arid and semi-arid Yellow River irrigation districts of northwestern China.
为更细致地刻画玉米冠层结构和光能利用特征,本研究引入MODIS MCD15A3H产品中的叶面积指数(Leaf area index,LAI)和光合有效辐射吸收比例(Fraction of photosynthetically active radiation,FPAR)数据。该产品空间分辨率为500 m,时间分辨率为4 d,经月尺度聚合后与其他遥感变量进行匹配[16]。
该框架包括4个核心环节。第一,基于宁夏县级行政边界和玉米种植区掩膜,提取2019—2023年玉米生长季内的遥感、气候和土壤变量,并与县域统计产量数据进行时空匹配,构建“县域-年份-月份-变量”的月尺度特征矩阵。第二,采用因果森林估计各变量对玉米产量的潜在影响,并结合条件平均处理效应(Conditional average treatment effect,CATE)排序、平均处理效应(Average treatment effect,ATE)显著性检验和群体平均处理效应(Group average treatment effects,GATES)异质性检验筛选具有解释意义的关键驱动因子。第三,将筛选后的关键变量输入TCN-TA模型,由TCN提取玉米生长季内相邻月份和跨月份的动态变化特征,并通过时间注意力机制学习不同月份对产量预测的相对贡献。第四,基于预测结果开展模型精度评价、消融分析、年度泛化验证和空间制图分析,形成从多源数据构建、因果启发筛选、时序建模到可解释评价的完整流程。该框架的核心思想是:因果森林用于提升输入特征的稳健性和解释性,TCN用于刻画玉米生长季内的时序依赖关系,时间注意力机制用于识别关键月份信息贡献。三者共同构成因果启发的县域玉米产量可解释预测方法。
DongL X, ZhangB, LiG, YanZ G, LuY L. Comparison of spring wheat yield prediction models in dryland based on machine learning and the ARIMA model[J]. Journal of Triticeae Crops, 2024, 44(12): 1551-1559 (in Chinese)
[3]
YangX, HuaZ K, LiL, HuoX H, ZhaoZ Q. Multi-source information fusion-driven corn yield prediction using the Random Forest from the perspective of Agricultural and Forestry Economic Management [J]. Scientific Reports, 2024, 14: 4052
FengM C, SuY, LinT, YuX, SongY, JinX L. High throughput cotton yield estimation based on multi-source remote sensing data from unmanned aerial vehicles and machine learning[J]. Transactions of the Chinese Society for Agricultural Machinery, 2025, 56(3): 169-179 (in Chinese)
[6]
DuJ L, ZhangY, WangP X, TanseyK, LiuJ M, ZhangS Y. Enhancing winter wheat yield estimation with a CNN-transformer hybrid framework utilizing multiple remotely sensed parameters[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 4405213
WangL L, LyuJ, ZhangJ Y. Explicating the role of agricultural socialized services on chemical fertilizer use reduction: Evidence from China using a double machine learning model[J]. Agriculture, 2024, 14(12): 2148
[9]
IatrouM, MourelatosS, KarydasC. Comparing XGBoost and double machine learning for predicting the nitrogen requirement of rice[J]. Remote Sensing, 2026, 18(3): 420
[10]
JiaoL C, WangY H, LiuX, LiL L, LiuF, MaW P, GuoY W, ChenP H, YangS Y, HouB. Causal inference meets deep learning: A comprehensive survey[J]. Research, 2024, 7: 0467
[11]
SrivastavaK, SawarkarA, NawaleS, AgrawalA R. Influence of causal inference for crop prediction[C]//In: 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT). Kamand: Institute of Electrical and Electronics Engineers, 2024: 1-6
[12]
CuiY F, KosorokM R, SverdrupE, WagerS, ZhuR Q. Estimating heterogeneous treatment effects with right-censored data via causal survival forests[J]. Journal of the Royal Statistical Society Series B: Statistical Methodology, 2023, 85(2): 179-211
[13]
ShiJ M, ShiF H, HuangX X. Prediction of maturity date of leafy greens based on causal inference and convolutional neural network[J]. Agriculture, 2023, 13(2): 403
[14]
RungeJ, NowackP, KretschmerM, FlaxmanS, SejdinovicD. Detecting causal associations in large nonlinear time series datasets[EB/OL]. [--].
[15]
CihlarJ, St.-Laurent L, Dyer J A. Relation between the normalized difference vegetation index and ecological variables [J]. Remote Sensing of Environment, 1991, 35(2/3): 279-298
[16]
HueteA, DidanK, MiuraT, RodriguezE P, GaoX, FerreiraL G. Overview of the radiometric and biophysical performance of the MODIS vegetation indices[J]. Remote Sensing of Environment, 2002, 83(1/2): 195-213
[17]
PettorelliN, VikJ O, MysterudA, GaillardJ M, TuckerC J, StensethN C. Using the satellite-derived NDVI to assess ecological responses to environmental change[J]. Trends in Ecology & Evolution, 2005, 20(9): 503-510
[18]
MyneniR B, HoffmanS, KnyazikhinY, PrivetteJ L, GlassyJ, TianY, WangY, SongX, ZhangY, SmithG R, LotschA, FriedlM, MorisetteJ T, VotavaP, NemaniR R, RunningS W. Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data[J]. Remote Sensing of Environment, 2002, 83(1-2): 214-231
[19]
AbatzoglouJ T, DobrowskiS Z, ParksS A, HegewischK C. TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958-2015 [J]. Scientific Data, 2018, 5: 170191
[20]
FuH K, LuJ, LiJ, ZouW L, TangX H, NingX Y, SunY. Winter wheat yield prediction using satellite remote sensing data and deep learning models[J]. Agronomy, 2025, 15(1): 205
[21]
PoggioL, de SousaL M, BatjesN H, HeuvelinkG B M, KempenB, RibeiroE, RossiterD. SoilGrids 2.0: Producing soil information for the globe with quantified spatial uncertainty[J]. Soil, 2021, 7(1): 217-240
[22]
HeinemannH, SeidelF, HirteJ, WelckerC, ChapuisR, GelotC, BeauchêneK, LuchaireN, BouidghaghenJ, DonA. Variability in maize root biomass distribution under drought can contribute to climate change adaptation[J]. Plant and Soil, 2026, 519(1): 1025-1044
[23]
NiuJ C, TangH Z, ZhangY Y, GaoB B, JiW J, YuQ Y, HuangY F. Optimization of crop yield prediction models based on stratified soil properties[J]. Smart Agricultural Technology, 2025, 12: 101466
[24]
宁夏回族自治区统计局. 宁夏统计年鉴(2019-2023)[EB/OL].[--].
[25]
Ningxia Hui autonomous region bureau of statistics. ningxia statistical yearbook (2019-2023)[EB/OL].[--]. in Chinese)
[26]
AtheyS, WagerS. Estimating treatment effects with causal forests: An application[J]. Observational Studies, 2019, 5(2): 37-51
[27]
ChernozhukovV, ChetverikovD, DemirerM, DufloE, HansenC, NeweyW, RobinsJ. Double/debiased machine learning for treatment and structural parameters[J]. The Econometrics Journal, 2018, 21(1): C1-C68
[28]
WangS, LiY, LiuZ Y, LiuX Y. Intelligent faulty diagnosis on photovoltaic farm based on spatial attention mechanism and temporal convolutional network[J]. Next Research, 2025, 2(3): 100604
[29]
LinL, WuJ L, FuS, ZhangS H, TongC S, ZuL Z. Channel attention & temporal attention based temporal convolutional network: A dual attention framework for remaining useful life prediction of aircraft engines[J]. Advanced Engineering Informatics, 2024, 60: 102372
[30]
WangZ M, WuH B, LiuJ H, WangW K, JiangC, LiQ M. Deterioration trend prediction of spalling damage for heavy-haul railway based on PCA and TCN-attention [C]//In: 2(CAC)023 China Automation Congress. Chongqing: Institute of Electrical and Electronics Engineers, 2023: 7425-7430
[31]
BaiS, KolterJ Z, KoltunV. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling[EB/OL]. [--].
KakimotoS, MienoT, TanakaT S T, BullockD S. Causal forest approach for site-specific input management via on-farm precision experimentation [J]. Computers and Electronics in Agriculture, 2022, 199: 107164
XuX Y, ShaoC X, SunZ G, LongB J, DongW L. Research progress on the effect of heat stress on physiological characteristics of maize at key growth stage and the yield [J]. Journal of Maize Sciences, 2021, 29(2): 81-88, 96 (in Chinese)
LiuH, MaZ H, LiuW F, WanM H, MaF L, WuN, LiuJ L. Effects of different tillage practices with organic fertilizers on rhizosphere soil microbial communities of maize in saline-alkali soils[J]. Chinese Journal of Eco-Agriculture, 2025, 33(1): 25-39 (in Chinese)
[39]
WagerS, AtheyS. Estimation and inference of heterogeneous treatment effects using random forests[J]. Journal of the American Statistical Association, 2018, 113(523): 1228-1242
[40]
ChernozhukovV, DemirerM, DufloE, Fernández-ValI. Fisher–Schultz lecture: Generic machine learning inference on heterogeneous treatment effects in randomized experiments, with an application to immunization in India [J]. Econometrica, 2025, 93(4): 1121-1164
JiangY, MaT H. Prediction of Henan winter wheat yield using CNN-LSTM-attention network [J]. Journal of Triticeae Crops, 2024, 44(10): 1352-1359 (in Chinese)
[43]
SinghK, HuangY B, YoungW, HarveyL, HallM, ZhangX, LobatonE, JenkinsJ, ShankleM. Sweet potato yield prediction using machine learning based on multispectral images acquired from a small unmanned aerial vehicle[J]. Agriculture, 2025, 15(4): 420
[44]
JiangY B, LiH Y, MaW Q, YuW J, ChenJ X, GaoY L, QiG P, YinM H, KangY X, MaY L, WangJ H, XuL T. A meta-analysis of the effects of nitrogen fertilizer application on maize ( Zea mays L.) yield in Northwest China [J]. Frontiers in Plant Science, 2025, 15: 1485237
[45]
LiW Y, LiW P, ZhengY J, WangX S, QiX F. Dynamics of soil moisture and its response to meteorological factors at different depths in an arid land, northwest China[J]. Atmosphere, 2026, 17(3): 232
[46]
LiX H, ZhugeS L, DuJ Y, ZhangP, WangX Y, LiuT J, LiD H, MaH R, LiX Z, NieY X, LiaoC J, DingH P, ZhangZ M. The molecular mechanism by which heat stress during the grain filling period inhibits maize grain filling and reduces yield[J]. Frontiers in Plant Science, 2025, 15: 1533527
[47]
MaJ Q, ZhaoF, CuiB F, LiuL, HaoX P, ZhaoY, DingY, ChenY J. TCN-attention model-based prediction of reference crop evapotranspiration in northern Henan Province [J]. Agronomy, 2026, 16(4): 435