MS-YieldStackNet: multi-source data fusion for wheat yield estimation using a stacked ensemble neural network

Machine learning methods for wheat yield prediction

Machine learning has transformed wheat yield prediction by modeling complex agroecological relationships. Traditional models like linear regression and decision trees have been widely used (Wang et al., 2020), but advanced methods, such as random forests and gradient boosting, offer improved accuracy, with R² values up to 0.78 in semi-arid regions (Pandey & Singh, 2021). Deep learning approaches, including CNNs and long short-term memory (LSTM) models, have shown promise, achieving RMSEs as low as 522.3 kg/ha using MODIS data (Gao et al., 2022). Hybrid ML approaches, combining multiple algorithms, further enhance performance; for instance, Agarwal & Tarar (2021) integrated random forests and LSTMs for precise forecasts in Indian agriculture. Recent advancements in automated ML (AutoML) have streamlined model selection and hyperparameter tuning, with Cai et al. (2019) reporting an R² of 0.75 for wheat yield using AutoML with satellite and climate data. The study Wang et al. (2020) fuse multi-source data with AdaBoost, achieving an R² of 0.86 and RMSE of 0.51 t/ha for winter wheat in the U.S., utilizing high-frequency data, which is less accessible in low-resource settings. The development of an automated machine learning approach is proposed in Kheir et al. (2024) for robust and fast crop yield estimation using a fusion of soil, remote sensing, and weather dataset 20 models in the new approach, proving AutoML’s outperformance over conventional ML. The Three Decision Support System used in Kheir et al. (2025), and DSSAT wheat models slightly overestimated wheat yield but accurately predicted nitrogen content. The hybrid PBM-MLRS approach closely estimated Fe and Zn content with a root mean square error (RMSE) of 0.42 t/ha for yield and 0.89% for nitrogen content.

However, these studies often rely on single-model architectures or limited data modalities, failing to capture the full complexity of yield-influencing factors. Stacking ensemble methods, which combine multiple models (e.g., neural networks and random forests), remain underexplored for wheat yield prediction despite their potential to improve robustness and accuracy. This study addresses this gap by integrating three feed-forward neural networks with a random forest meta-learner, achieving an RMSE of 78.19 kg/ha and MAE of 59.07 kg/ha, surpassing prior benchmarks.

Data types in yield prediction models

The choice of data sources significantly impacts model performance. Remote sensing data, including vegetation indices like NDVI and GNDVI, are widely used for crop monitoring, with studies reporting RMSEs of 180–850 kg/ha when combined with weather data (Pang, Chang & Chen, 2022; Ruan et al., 2022). Soil data, such as nutrient levels and pH, enhance prediction accuracy by capturing ground-level variability (Shen et al., 2022). However, most studies focus on one or two data types, limiting their ability to model complex agroecological interactions. For example, Franch et al. (2019) used remote sensing data alone, achieving an R² of 0.7, while Arshad et al. (2023) improved accuracy (R² = 0.85) by assimilating Sentinel-2 data into crop growth models. The lack of comprehensive multi-source integration remains a critical gap. Our study addresses this by fusing multispectral satellite imagery (NDVI, DVI), in-situ soil analytics, and agrometeorological data within a unified feature space, processed by MS-YieldStackNet, to achieve high precision in regional yield forecasting.

Region-specific applications in agroecological zones

Wheat yield prediction models must account for regional agroecological variability. In semi-arid regions like Pakistan, Wang et al. (2020) developed a two-branch deep learning model for winter wheat, reporting an RMSE of 721 kg/ha. In Australia, Cai et al. (2019) fused satellite and climate data and achieved high performance with R² = 0.75. In Brazil, Schwalbert et al. (2020) applied neural networks for soybean yield. These studies highlight the challenge of achieving high accuracy in diverse agroecological zones, particularly in data-scarce regions like Faisalabad, Pakistan. The proposed system outperforms these benchmarks by leveraging multi-source data and a stacked ensemble approach, achieving an R² of 0.81, RMSE of 78.19 kg/ha, and MAE of 59.07 kg/ha in Faisalabad, demonstrating superior performance in a similar agroecological context.

Novelty of this work

While prior studies have advanced wheat yield prediction, they often rely on single data modalities or conventional ML models, limiting their scalability and precision in complex agroecological settings. The proposed framework introduces a novel stacked ensemble neural architecture that integrates multispectral satellite imagery, in-situ soil analytics, and meteorological variables, processed through three parallel feed-forward neural networks and a random forest meta-learner. Unlike hybrid models (Agarwal & Tarar, 2021) our method optimizes multi-source data fusion and ensemble learning, achieving superior performance metrics (R² = 0.81, RMSE = 78.19 kg/ha) that surpass all previous studies (see Table 1). This work advances the field by providing a scalable, high-precision framework for regional wheat yield forecasting, addressing critical gaps in data integration and model complexity.

Study area

The study is carried out at approximately city level, i.e., Faisalabad, Pakistan. As Faisalabad is an agricultural city and a variety of crops are sown and cultivated here. The reason for selecting this area is that there are many wheat crop lands and every year these crops are cultivated. Wheat crop is considered the backbone of any country in perspective of food and economy. Also, it is considered as a cash crop as it plays a vital role in the Gross Domestic Product (GDP) of Pakistan. Agriculture University and many agriculture research centers are the plus point with respect to information gathering. The depiction of the study area using satellite imagery is represented in Fig. 1.

Data sources

The dataset comprises wheat yield records and different agrometeorological parameters recorded seasonally for the past several years. Additionally, remote sensing data are curated to calculate vegetation indices, which, combined with agrometeorological and soil data, are utilized in constructing models for yield estimation.

Data fusion and alignment

Data from remote sensing (NDVI, DVI from Landsat-8), weather (seasonal averages of precipitation, minimum temperature, maximum temperature, mean temperature, relative humidity, wind speed), soil (organic matter, EC, pH, available phosphorus, potassium, saturation, texture), and yield (per-acre records) were aligned temporally and spatially. Temporally, data were synchronized to the wheat growing season (November–April) using seasonal aggregates over 5 years, with remote sensing images selected to match these intervals. Spatially, data were georeferenced to specific coordinates in Faisalabad using GPS equipment (aligned via Google Earth Engine for remote sensing data). Soil and yield data, collected at the field level, were mapped to the same coordinates. After preprocessing, features were concatenated into a single input vector per field and season, creating a unified feature space that captures agroecological interactions for the MS-YieldStackNet model.

Handling missing data

Missing data were minimal, with only one null value each in DVI and NDVI. These rows were dropped using Python’s pandas library to maintain data integrity. No missing values were reported in weather, soil, or yield data, and the dataset contained no duplicate records.

Correlation analysis and feature selection

Subsequently, we performed a comprehensive analysis of the agrometeorological parameters that impact crop yield, as determined through Pearson correlation analysis. The correlation coefficients, ranging from −1 to 1, provide insights into the strength and direction of the relationships between each parameter and the yield. The agrometeorological parameter’s correlation with the yield is represented in Fig. 2. A threshold of $|r|>0.3$ was selected to retain features with moderate-to-strong correlations, balancing statistical significance $(p<0.05)$ and agroecological relevance for semi-arid regions like Faisalabad, as supported by prior studies (Pandey & Singh, 2021). For example, wind speed (r = 0.375825) and precipitation (r = 0.308777) showed positive correlations, reflecting their importance for wheat yield. Temperature attributes (maximum, minimum, mean) were also positively correlated, with minimum temperature exhibiting a strong relationship. Relative humidity also shows a positive relationship with yield, with a relatively lower correlation value, indicating that it affects the amount of wheat yield, but that the variation in humidity in the region does not significantly affect wheat growth and yield (Ismail et al., 2023). Finally, features were standardized using Z-score normalization, transforming each to have a mean of zero and a standard deviation of one, ensuring compatibility with the regression models.

Stacking ensemble for enhanced yield prediction

The MS-YieldStackNet model addresses the research gap of limited integration of multi-source data (remote sensing, weather, soil) with advanced ensemble methods in semi-arid regions like Faisalabad, Pakistan. Its novelty lies in combining heterogeneous data with a stacking ensemble of three FFNNs and a random forest meta-learner, achieving superior performance (RMSE of 78.19 kg/ha, MAE of 59.07 kg/ha) compared to single-model approaches (Kattenborn et al., 2021; Lu, Tan & Jiang, 2021).

Three identical FFNNs were selected as base learners for their simplicity, computational efficiency, and ability to capture non-linear patterns in multi-source agricultural data. Empirical validation showed that three FFNNs optimized diversity and computational cost, yielding the lowest MSE (6,114.30 on the test set) compared to configurations with one, two, or four base learners. Alternative models, such as CNNs and LSTMs, underperformed due to the dataset’s limited spatial and temporal complexity, while random forest produced a higher RMSE (e.g., 180 kg/ha). The random forest meta-learner was chosen over a linear regressor for its robustness to overfitting and ability to model non-linear relationships among base learner predictions (see Fig. 4).

The dataset was split into 80% training and 20% testing subsets, with random shuffling applied to ensure unbiased model evaluation. Data leakage was ruled out by isolating the test set prior to training and hyperparameter tuning, ensuring it was not used to influence model development. Despite the dataset’s temporal structure (seasonal averages, November-April), random shuffling was used to align with standard machine learning practices, as the stacking ensemble effectively captured seasonal patterns.

The proposed framework for wheat yield estimation employs a comprehensive collection of multi-source data fusion and modeling strategy, formally outlined in Algorithm 1. This methodology integrates heterogeneous curated agricultural datasets comprising geospatial records ( ${\mathcal{D}}_s$), satellite-based vegetation indices ( ${\mathcal{D}}_r$), in-situ meteorological measurements ( ${\mathcal{D}}_w$), soil sample analytics ( ${\mathcal{D}}_{soil}$), and corresponding ground-truth yield observations ( ${\mathcal{D}}_y$). In the initial phase, spectral indices such as NDVI and DVI are derived from satellite imagery. Concurrently, key meteorological variables including temperature, precipitation, relative humidity, and solar radiation are collected alongside critical soil parameters (e.g., pH, potassium, saturation levels). These features are concatenated into a unified vector representation ${\mathbf{x}}_i\in {\mathbb{R}}^d$ following spatio-temporal synchronization across all data modalities, resulting in a fused dataset ${\mathcal{D}}_f$.

Subsequent preprocessing steps involve noise filtering through statistical thresholding and normalization of feature values. The dataset is then partitioned into training and testing subsets. A Stacked Ensemble Neural Network is employed for model training, where multiple base learners $M_k:{\mathbf{x}}_i\to {\hat{y}}_i^{(k)}$ are independently trained to generate preliminary yield predictions. Their outputs are aggregated into a meta-feature vector ${\hat{\mathbf{y}}}^{(base)}$, which is then passed to a meta-learner $M_{meta}$ to refine the final yield estimate ${\hat{y}}_i$. Model evaluation is conducted using standard metrics, including root mean square error (RMSE) and the coefficient of determination $R^2$, ensuring predictive accuracy and generalizability. Finally, the trained model ${\mathcal{M}}_{SENet}$ is deployed through a web-based platform to facilitate real-time, scalable yield forecasting, thereby empowering farmers, agronomists, and policymakers with actionable insights.

Neural architecture with hyperparameters

The structure starts with an input layer with 64 neurons, each corresponding to a feature inside the dataset. These neurons feed their activations through Rectified Linear Unit (ReLU) activation functions, which introduce non-linearity into the model. ReLU is chosen for its simplicity and effectiveness in preventing the vanishing gradient problem, commonly encountered in deep neural networks (Elsken, Metzen & Hutter, 2019; Feurer & Hutter, 2019). Following the primary hidden layer, there is a second hidden layer with 32 neurons. This layer serves to further extract and abstract features from the data, allowing the model to learn more complex patterns. Again, ReLU activation functions are applied. The output layer consists of a single neuron, configured with a linear activation function, which is suitable for continuous value regression tasks (Wistuba, Rawat & Pedapati, 2019). Following the neural network layers, a random forest model is used as a meta-learner for prediction. The choice to use Random Forest after the neural network layers is due to its robustness to overfitting, ability to handle large datasets with high dimensionality, and capability to capture complex non-linear relationships. Additionally, random forest provides interpretability through feature importance analysis, complementing the deep learning model’s predictions. The description pf model is illustrated in Table 4.

Evaluation metrics

Evaluation metrics provide numerical measures of a model’s performance. Evaluation metrics aid in decision-making processes such as model selection, feature engineering, hyperparameter tuning, and assessing the impact of different algorithms or techniques. While numerous metrics exist for evaluating regression models, the two most common are MSE and MAE.

MAE provides a simple understanding of the magnitude of errors in the predictions. MSE boosts large errors due to the squaring operation, making it sensitive to outliers. MAPE is particularly useful in time series analysis, where you are trying to predict future values based on historical data. By calculating the MAPE, you get a sense of how far off your forecasts are from the actual outcomes expressed as a percentage. R² is useful for understanding how well the model explains the variability in the data and comparing different models’ performances. Both MSE and MAE measure the differences between predicted and actual values. However, MSE squares these differences, amplifying large errors and potentially biasing results. MAE avoids this by using absolute values, providing a more robust measure for datasets with outliers. RMSE balances emphasis on smaller and larger errors by taking the square root of MSE. Detailed descriptions and formulas for these metrics are provided in the appendix. The predictive performance of the regression algorithms is presented in Table 5.

Experimental results and discussion

The comparison in Table 5. demonstrates that the stacking ensemble exhibits superior predictive performance compared to baseline models, justifying its selection as the final model. The stacking ensemble achieved an R² of 0.81, implying it explains nearly 81% of the observed yield variance, a strong indicator of suitability for yield prediction in this context. Performance measurements like RMSE (78.19 kg/ha), MSE (6,114.30), MAE (59.07 kg/ha), and MAPE were calculated to measure the error between predicted and actual yield values, with the stacking ensemble providing the lowest values among tested models.

Baseline models were included to provide a robust comparison. Linear regression, as a simple baseline, achieved RMSE of approximately 173 kg/ha, reflecting its limitations with the complex, non-linear relationships in the multi-source data. Random Forest Regression, the second-best performing algorithm, had an RMSE of approximately 172 kg/ha, while LSTM, ranked third, showed competitive performance based on MSE, RMSE, MAE, and MAPE (see Fig. 5). The comparison in Table 5 indicates that the stacking ensemble’s R² of 0.81 is a reasonable value for yield prediction, reflecting its ability to capture the underlying patterns in the fused dataset.

A critical evaluation of our work against relevant and recent state-of-the-art studies confirms its performance and practical utility. For example, comparing with multi-source data and AdaBoost model by Wang et al. (2020), our method marks a reduction in prediction error. After converting their reported errors to consistent units (kg/ha) reveals that our work reduces RMSE by approximately 85% (78.19 vs. 510 kg/ha) and MAE by 85% (59.07 vs. 390 kg/ha).

Next, Kheir et al. (2024) demonstrated the effectiveness of an AutoML framework evaluating 20 models, our purpose method achieves a higher R² value with a more streamlined and computationally efficient architecture Finally, MS-YieldStackNet also outperforms the complex hybrid process-based ML approach by Kheir et al. (2025), which reported an RMSE of 420 kg/ha. Our RMSE of 78.19 kg/ha represents an 81% reduction in error, demonstrating that a well-designed data-driven model can surpass the yield prediction accuracy of more intricate hybrid methodologies. Beyond raw accuracy, a key advantage of our approach is its design for practicality; unlike approaches reliant on high-frequency data or multiple model integrations, it is engineered for high performance with aggregated seasonal data, enhancing its feasibility and scalability in resource-constrained agricultural environments.

In an ideal scenario, the residual plot should display a random and uniform distribution of values around the identity line, serving as a critical tool for assessing the magnitude of errors and identifying observations contributing to these errors. The plot in Fig. 6 reveals heteroscedasticity, particularly for high-yield observations in the range of 2,000 to 2,200, which correspond to the most recent records. This pattern may be attributed to precision farming techniques adopted in recent years and changing agricultural methods not fully captured by the dataset attributes. These factors likely contribute to the larger residuals observed in this range, reflecting reduced correlation with historical data.

The scatter plot visualizes both anticipated and ground-reality yield values, with a regression line indicating the high-quality becoming relationship among them. A modest vertical hole among the regression line and observations suggests the version’s robustness. The graph in Fig. 7 gives a quick assessment of the model’s predictive overall performance and its alignment with real facts points. The following graph shows it carefully;

In conclusion, the stacking ensemble method showcases exceptional predictive performance, as evidenced by its high R² value and low error metrics. This suggests its suitability for accurate yield prediction, outperforming other regression algorithms evaluated in the study.

This study analyzed and discussed different past studies and showed their pros and cons. Additionally, these existing studies were compared with the proposed study (see Table 6) and proved that how our system is beneficial, advanced and highly accurate than the existing ones.

Conventional regression algorithms

Deep learning algorithms

Evaluation metrics