A hybrid deep learning approach for winter wheat yield prediction: evidence from leveraging multi-source data Authors: Manogna R. L, Shaanil Punglia, Siddhant Tushar Joshi Published: 2026 ABSTRACT Accurate district-level wheat yield forecasts are critical for food security planning, supply-chain management, and agricultural policy in India, the world’s second-largest wheat producer. We benchmark nine model classes for this task on a 23-year (2001–2023) dataset of 275 districts across India’s seven largest wheat-producing states, which together account for ∼95% of national production. The benchmark covers Random Forest, XGBoost, LightGBM, a 1D-CNN, an LSTM, a BiLSTM, a single-stream Transformer encoder, the recently proposed Parallel CNN-LSTM-Attention design, and our hybrid CNN-BiLSTM-Attention with modality-specific routing (a 1D-CNN over the vertically structured soil profile and a BiLSTM with self-attention over the meteorological and remote-sensing time series). The proposed model is the best entry, achieving a Mean Absolute Error (MAE) of 273.2 kg/ha and an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} of 0.795 on the held-out test set — a 43.8% MAE reduction over the Random Forest baseline, a ∼28% reduction over the gradient-boosted baselines, and a ∼4% reduction over the next-best deep model. A simple persistence forecast (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) however, achieves an MAE of 274.9 kg/ha, essentially tying the proposed model on average. We show that the architectural value-add concentrates in anomalous years: in the dry 2023 sowing season the model improves MAE by 7.2% and RMSE by 11.2% over persistence, and SHAP attribution localises the temporal contribution to the February–March grain-filling window led by EVI and NDVI signal — consistent with the well-documented sensitivity of wheat grain-filling to moisture and temperature stress in that window. Together, these results position persistence-aware, modality-specific deep learning as a practical framework for stress-sensitive yield forecasting in data-scarce agricultural regions. Introduction Accurate and timely crop yield prediction is essential for global food security, enabling effective policymaking, supply chain management, and market stabilization1,2. Reliable production signals also propagate through the wider agricultural economy: they shape price discovery and the lead–lag efficiency of spot and futures markets3,4, and they mediate how the financialization of agricultural commodities feeds back into food security5,6. These linkages are especially consequential in emerging economies, where food security interacts with energy use, trade openness, and access to finance7,8. This is critical in India, one of the world’s largest agricultural producers with a significant economic dependence on the sector. As a staple crop, wheat is crucial to India’ s agriculture, with the country’ s core wheat belt producing nearly 14% of the global supply9,10. However, forecasting yield in this region is exceptionally difficult due to its vast and diverse agro-climatic zones, complex environmental interactions, and high vulnerability to climate variability. Yield-forecasting methods have shifted from process-based simulations such as DSSAT and APSIM toward data-driven models that learn non-linear input–yield relationships from observational data11,12. Yet both approaches expose complementary limitations at district scale: process models require intensive local calibration, while traditional machine-learning baselines flatten temporal, spatial, and soil-depth structure into one-dimensional feature vectors. Deep learning offers a way to retain this structure, with recurrent networks capturing meteorological and vegetation sequences13 and convolutional networks extracting patterns from structured soil profiles or satellite imagery14,15; recent work therefore increasingly favours hybrid architectures that combine these strengths16. Despite this proliferation of architectures, the deep-learning literature for India’ s wheat belt remains methodologically thin and is dominated by single-architecture studies that compare a proposed model only against a Random Forest or Linear Regression reference. We address this gap with two complementary contributions. First, we conduct a controlled benchmark of nine model classes — Random Forest, XGBoost, LightGBM, 1D-CNN, LSTM, BiLSTM, a single-stream Transformer encoder, the recently proposed parallel CNN-LSTM-Attention design17, and our hybrid CNN-BiLSTM-Attention — under identical feature access. Second, we examine modality-specific routing of inputs as a deliberate design choice: a 1D-CNN over the vertically structured soil profile and a BiLSTM with self-attention over the meteorological and remote-sensing time series, in contrast to designs that concatenate or replicate inputs across heterogeneous branches. We validate this benchmark on a 23-year dataset (2001–2023) covering 275 districts across India’ s seven core wheat-producing states, which account for nearly 95% of national production18. Rather than treating marginal leaderboard gains as the sole measure of progress, we evaluate the proposed architecture against both modern deep baselines and an explicit persistence forecast, asking when learned meteorological, remote-sensing, and soil signals add information beyond inter-annual yield autocorrelation. The results show that the hybrid model’ s advantage is most meaningful in anomalous seasons, where SHAP attributions connect its gains to agronomically interpretable EVI and NDVI signal during the February–March grain-filling window. Ablation studies further confirm that the soil-routing branch, bidirectional temporal pathway, and self-attention mechanism each contribute non-redundant predictive capacity. Because the model is trained for early prediction and forecasts district-level yields one month before harvest completion, the benchmark is designed around the lead time needed for practical policy and supply-chain interventions. The remainder of this paper is organised as follows. Section Literature Review reviews related work in crop yield prediction and deep learning architectures for agricultural applications. Section Data describes the study area, data sources, and preprocessing procedures. Section Methodology presents the proposed CNN-BiLSTM-Attention architecture, the training protocol, and the eight comparison baselines. Section Results and Discussion reports the experimental results, including the nine-model benchmark, the ablation studies, and a SHAP-based feature-level interpretability analysis that contextualises the architectural contribution against an explicit persistence baseline. Finally, Section Conclusion concludes with a discussion of key findings and directions for future research. Literature review This review situates the proposed benchmark within three strands of prior work: the transition from process-based crop simulations to empirical machine-learning models, the rise of deep sequence and convolutional architectures for structured agricultural data, and the remaining gap between global deep-learning benchmarks and district-scale wheat forecasting in India. The progression from process-based to machine learning models Historically, the modeling of agricultural output relied heavily on biophysical simulation frameworks. These process-based models, such as the Decision Support System for Agrotechnology Transfer (DSSAT) and the Agricultural Production Systems sIMulator (APSIM), offer a rigorous, scientifically transparent approach to understanding the physiological development of crops under specific environmental stimuli19,20. They utilize fundamental equations of thermodynamics, fluid dynamics, and plant biology to simulate daily biomass accumulation. While these mechanistic models are highly effective at the field scale, their practical utility at a regional or district level is inherently constrained. Process-based models require massive amounts of precisely calibrated, localized input data; ranging from highly specific soil pedology to granular, daily field management interventions like exact nitrogen application rates and irrigation volumes. Such high-fidelity data is rarely available at scale, particularly in the Indian subcontinent. Consequently, when these models are extrapolated across vast, heterogeneous geographic areas, the uncertainty inherent in coarse soil or climatic approximations frequently eclipses the predictive signal, rendering large-scale forecasts unreliable and overly sensitive to parameter misspecification21. To circumvent the stringent limitations of explicit biophysical simulation, researchers increasingly turned to data-driven empirical methodologies, specifically traditional machine learning algorithms. Techniques such as Random Forest, Multiple Linear Regression, Support Vector Machines, and Gradient Boosting have demonstrated considerable efficacy in modeling the non-linear, complex relationships between agricultural yields and environmental covariates11,12,22. These algorithms excel at finding statistical correlations within historical datasets without requiring explicit biological programming. However, conventional machine learning approaches possess a fundamental architectural limitation that severely hinders their application to advanced agricultural forecasting: they necessitate the flattening of multidimensional data into simplistic, one-dimensional feature vectors. This mandatory transformation systematically strips away the intrinsic structural integrity of the agricultural data. For instance, the chronological progression of a growing season, from the onset of the monsoon to the dry harvest, is lost when time-series weather data is flattened, just as the vertical stratification of soil profiles is obliterated when depth-dependent variables are aggregated into a single spatial vector. Deep learning architectures and the shift to hybrid models The advent of Deep Learning has catalyzed a paradigm shift in precision agriculture and regional yield forecasting. Deep neural networks are uniquely capable of learning hierarchical, abstract representations directly from raw, highly structured data without the need for manual feature engineering. By deploying specialized architectures such as Recurrent Neural Networks for sequential meteorological data and Convolutional Neural Networks for spatial grids and vertical soil profiles, researchers can now natively process the complex, multi-modal data streams generated by Earth observation satellites, automated weather stations, and global pedological databases. The contemporary frontier of this domain is defined by hybrid deep learning models that intricately fuse diverse neural network branches to simultaneously capture spatial, temporal, and static agricultural determinants. Recurrent Neural Networks (RNNs) and their advanced variants, particularly Long Short-Term Memory (LSTM) networks, became foundational for crop yield prediction due to their superior capability in handling sequential data and capturing long-term temporal dependencies inherent in climate and vegetation time series23. Studies using LSTM models have demonstrated enhanced predictive performance over traditional Random Forest and LASSO approaches, particularly in estimating yields under extreme weather conditions24–26. Beyond yield itself, the same sequence-learning advantages have been documented for related agricultural forecasting tasks in India, where LSTM- and GRU-based models, including hybrid LSTM-GARCH formulations, consistently outperform classical econometric baselines in forecasting volatile commodity prices27,28. Concurrently, Convolutional Neural Networks (CNNs) have proven effective for feature extraction across structured agricultural data. One-dimensional CNNs have demonstrated strong performance on sequential streams, including for yield estimation in the Indian wheat belt15, while higher-dimensional variants have been applied to multispectral imagery14. A more recent line of work replaces or supplements the recurrent backbone with self-attention. Single-stream Transformer encoders applied to monthly meteorological and remote-sensing sequences have reached state-of-the-art accuracy for winter-wheat yield in Chinese provinces29; we replicate this Transformer-encoder core as one of our deep baselines. Hybrid designs that fuse attention with recurrent and convolutional branches have also emerged: in particular, the Parallel CNN-LSTM-Attention (TPCLA) architecture of Song et al.17 routes the temporal stream through both a CNN and an LSTM branch in parallel before merging the representations with attention. We include TPCLA as a baseline because it directly contrasts with the modality-specific routing pursued in this work. The state-of-the-art recurrent backbone has also advanced to Bidirectional LSTMs (BiLSTMs). A BiLSTM processes the temporal sequence of weather and vegetation indices in two parallel directions — chronologically forward from sowing to harvest, and reverse-chronologically backward from harvest to sowing — so that the hidden state at any month carries both historical context and downstream consequences. This bidirectionality is particularly valuable in crop modelling: the physiological impact of a mid-season heatwave is much better understood when the hidden state simultaneously encodes the moisture conditions that preceded it and the accelerated, stress-induced senescence that immediately followed it13,30,31. Survey of current state-of-the-art and identification of research gap While traditional machine learning methods, such as Random Forest and Gradient Boosted Regression have been rigorously evaluated for crop yield prediction in India32–34, the application of more advanced deep learning approaches remains methodologically underdeveloped compared to global benchmarks. Table 1 synthesizes the current state of deep learning in crop yield prediction, contrasting global state-of-the-art models with existing approaches applied to the Indian context.Table 1Comparative analysis of deep learning architectures for crop yield prediction (2020–2025).StudyRegion & cropModel architectureWolanin et al. (2020)India (wheat belt)1D-CNNBali & Singla (2021)India (Punjab, wheat)RNN-LSTMSaini et al. (2023)India (sugarcane)CNN-BiLSTM hybridKhaki et al. (2020)US Corn Belt (corn/soy)CNN-RNNShahhosseini et al. (2021)US Corn Belt (corn)Ensemble CNN-DNNWang et al. (2022)China (winter wheat)LSTMGuo et al. (2024)China (winter wheat)SSA-LSTM-TransformerYe et al. (2024)China (winter wheat)ASTGNNMirhoseini Nejad et al. (2024)US (soybean)ConvLSTM-ViTSong et al. (2025)China (wheat)TPCLA Table 1Comparative analysis of deep learning architectures for crop yield prediction (2020–2025).StudyRegion & cropModel architectureWolanin et al. (2020)India (wheat belt)1D-CNNBali & Singla (2021)India (Punjab, wheat)RNN-LSTMSaini et al. (2023)India (sugarcane)CNN-BiLSTM hybridKhaki et al. (2020)US Corn Belt (corn/soy)CNN-RNNShahhosseini et al. (2021)US Corn Belt (corn)Ensemble CNN-DNNWang et al. (2022)China (winter wheat)LSTMGuo et al. (2024)China (winter wheat)SSA-LSTM-TransformerYe et al. (2024)China (winter wheat)ASTGNNMirhoseini Nejad et al. (2024)US (soybean)ConvLSTM-ViTSong et al. (2025)China (wheat)TPCLA Comparative analysis of deep learning architectures for crop yield prediction (2020–2025). A critical review reveals that while sophisticated architectures like CNN-RNNs and Vision Transformers are well-established in the US and China, their direct application to India is hindered by specific data constraints. The CNN-RNN framework by35 is heavily predicated on the availability of highly structured, uniform management data, regularly ingesting variables such as precise planting dates and standardized fertilizer rates. In contrast, Indian agriculture is characterized by deeply fragmented, unorganized smallholder farming where such management practices are largely undocumented and driven by localized socio-economic micro-factors. Consequently, direct transfer of US models fails because the foundational management matrices they require do not exist in India, forcing architectures to derive these latent variables purely from remote sensing proxies and historical yield lags. Furthermore, reliable digitised district-level records for the Indian wheat belt span only ∼20 years, which constrains the applicability of parameter-heavy, data-hungry architectures (e.g. Vision Transformers trained on 30–40 years of US/Chinese records) and favours data-efficient hybrid designs. To address these gaps, this work pursues four contributions:Modality-specific routing of inputs. The proposed CNN-BiLSTM-Attention model routes each input through an architecturally appropriate branch — a 1D-CNN over the vertically structured soil profile and a BiLSTM with self-attention over the meteorological and remote-sensing time series — rather than concatenating heterogeneous inputs into a single tensor or replicating them across parallel branches.Rigorous benchmark comparison. We evaluate nine model classes (Random Forest, XGBoost, LightGBM, 1D-CNN, LSTM, BiLSTM, Transformer encoder, Parallel CNN-LSTM-Attention, and our hybrid) under an identical-feature-access protocol. We additionally compare against an explicit persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_t = y_{t-1}$$\end{document}) to delineate where deep features add value over the strong inter-annual yield autocorrelation.Agronomic interpretability. We complement the temporal self-attention weights with a SHAP attribution analysis at the feature, soil-depth, and feature-by-month levels, demonstrating that the model’s decisions concentrate on agronomically critical periods.Spatial limits of district-scale yield forecasting. We use residual maps and state-wise error diagnostics to show that high aggregate accuracy can still mask systematic spatial biases. Modality-specific routing of inputs. The proposed CNN-BiLSTM-Attention model routes each input through an architecturally appropriate branch — a 1D-CNN over the vertically structured soil profile and a BiLSTM with self-attention over the meteorological and remote-sensing time series — rather than concatenating heterogeneous inputs into a single tensor or replicating them across parallel branches. Modality-specific routing of inputs. The proposed CNN-BiLSTM-Attention model routes each input through an architecturally appropriate branch — a 1D-CNN over the vertically structured soil profile and a BiLSTM with self-attention over the meteorological and remote-sensing time series — rather than concatenating heterogeneous inputs into a single tensor or replicating them across parallel branches. Rigorous benchmark comparison. We evaluate nine model classes (Random Forest, XGBoost, LightGBM, 1D-CNN, LSTM, BiLSTM, Transformer encoder, Parallel CNN-LSTM-Attention, and our hybrid) under an identical-feature-access protocol. We additionally compare against an explicit persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_t = y_{t-1}$$\end{document}) to delineate where deep features add value over the strong inter-annual yield autocorrelation. Rigorous benchmark comparison. We evaluate nine model classes (Random Forest, XGBoost, LightGBM, 1D-CNN, LSTM, BiLSTM, Transformer encoder, Parallel CNN-LSTM-Attention, and our hybrid) under an identical-feature-access protocol. We additionally compare against an explicit persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_t = y_{t-1}$$\end{document}) to delineate where deep features add value over the strong inter-annual yield autocorrelation. Agronomic interpretability. We complement the temporal self-attention weights with a SHAP attribution analysis at the feature, soil-depth, and feature-by-month levels, demonstrating that the model’s decisions concentrate on agronomically critical periods. Agronomic interpretability. We complement the temporal self-attention weights with a SHAP attribution analysis at the feature, soil-depth, and feature-by-month levels, demonstrating that the model’s decisions concentrate on agronomically critical periods. Spatial limits of district-scale yield forecasting. We use residual maps and state-wise error diagnostics to show that high aggregate accuracy can still mask systematic spatial biases. Spatial limits of district-scale yield forecasting. We use residual maps and state-wise error diagnostics to show that high aggregate accuracy can still mask systematic spatial biases. Taken together, these contributions position the study not simply as an accuracy benchmark, but as a testbed for separating learnable agro-climatic signal from the management, irrigation, and distribution-shift factors that make district-scale Indian wheat forecasting intrinsically difficult. Data This section first motivates the selection of the study region and characterizes its geographical challenges, then details the acquisition of meteorological, soil, remote sensing, and district-level yield datasets, and finally outlines the preprocessing and temporal alignment procedures employed prior to model development. Study area The study focuses on a contiguous region spanning seven states across northern, western, and central India which collectively produced 107.49 million tonnes of wheat in 2023-24, representing 94.88% of national and 13.69% of global wheat production10,18. Despite this productive capacity, recent yield stagnation9 has emerged from multiple interacting stressors: heat stress during critical growth stages36, disease pressure under humid conditions37, and erratic rainfall and hailstorms38. These challenges, combined with contrasting environments ranging from the fertile Indo-gangetic plains to the semi-arid conditions of Rajasthan, make the region an ideal testbed for evaluating yield prediction models. Figure 1 shows a satellite map of the study area with district boundaries overlaid.Fig. 1Satellite map of the study region spanning seven core wheat-producing states across northern, western, and central India. Thick white lines denote the outer boundaries of the region, while thinner white lines indicate district boundaries. The inset map (bottom right) shows the location of the study area within India highlighted in red. Map created using QGIS version 3.40 LTR (https://qgis.org). Satellite map of the study region spanning seven core wheat-producing states across northern, western, and central India. Thick white lines denote the outer boundaries of the region, while thinner white lines indicate district boundaries. The inset map (bottom right) shows the location of the study area within India highlighted in red. Map created using QGIS version 3.40 LTR (https://qgis.org). Data acquisition The dataset for this study was compiled from multiple sources and comprises yield data, remote sensing data, meteorological data, and soil data for the period 2000 to 2023. The first step in the data acquisition process was sourcing district-level administrative boundary shapefiles from the Survey of India39. These geospatial files served as the primary reference for all subsequent spatial aggregation of data, ensuring that all predictor and target variables corresponded to the same geographical units. Yield data Wheat yield data was sourced from the Department of Agriculture and Farmers’ Welfare portal40, focusing specifically on rabi (winter) wheat, which is sown after the monsoon at the onset of winter. The target variable, annual district-level winter wheat yield measured in kg/ha, was obtained for 275 of the 296 districts across the seven study states for 2000 to 2023. Data for the remaining 21 districts was unavailable, likely due to insufficient wheat cultivation. Yield values from 2001 to 2023 serve as ground truth for model training and evaluation, while the 2000 yields are used as a lagged predictive feature. Meteorological data and vegetation indices Meteorological observations were collected at daily resolution from the NASA POWER API41. The data, available at 0.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document} latitude by 0.625\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document} longitude resolution, was aggregated to the district level using zonal statistics based on the Survey of India shapefiles. Eleven meteorological variables were extracted, including maximum temperature, minimum temperature, daily precipitation, and relative humidity. This selection was guided by agronomic domain knowledge42,43 and prior studies35,44, with the complete list provided in Appendix A. To monitor vegetation health and growth dynamics throughout the season, two remote-sensing vegetation indices, the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI), were sourced. These indices were derived from the MODIS Terra MOD13Q1 product, accessed via the Google Earth Engine platform45,46, which provides 16-day temporal composites at 250 m spatial resolution. NDVI uses the red and near-infrared reflectance bands, while EVI additionally incorporates the blue band to reduce atmospheric and soil-background effects:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{NDVI}&= \frac{\rho _{\textrm{NIR}} - \rho _{\textrm{red}}}{\rho _{\textrm{NIR}} + \rho _{\textrm{red}}} \end{aligned}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{EVI}&= 2.5 \times \frac{\rho _{\textrm{NIR}} - \rho _{\textrm{red}}}{\rho _{\textrm{NIR}} + 6\rho _{\textrm{red}} - 7.5\rho _{\textrm{blue}} + 1} \end{aligned}$$\end{document}where ρ denotes the surface reflectance in the respective spectral bands (near-infrared, red, and blue). NDVI and EVI serve as robust proxies for vegetation density, health, and photosynthetic activity, which are directly correlated with crop biomass and potential yield47,48. The gridded data was aggregated to the district level using the same zonal statistics approach applied to meteorological data. Soil data Static soil characteristics were sourced from the ISRIC SoilGrids250m WCS database49, which provides global soil property maps at 250 m spatial resolution. Eight key soil features, known to influence soil fertility and water retention capacity43, were extracted at six standard depths (see Appendix A for the complete list). For each district, mean values of each soil property were calculated across its geographical area at each depth, preserving the vertical soil profile structure and yielding 48 distinct features per district (8 properties × 6 depths). These soil properties were treated as temporally static throughout the 2001-2023 study period. This is a reasonable assumption given that these inherent soil properties change over pedological or geological timescales, which are far longer than the 20-year timeframe of this analysis50. Data preprocessing Daily meteorological variables and 16-day vegetation indices were aggregated to monthly resolution by computing mean values. All features were then z-score normalized (3) using training set statistics to ensure comparable scales. Finally, to predict annual winter wheat yield for each district, temporal features were structured as sequences of monthly feature vectors spanning June to April, capturing both the preceding monsoon season and the subsequent winter wheat growing season.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} z = \frac{x - \mu }{\sigma } \end{aligned}$$\end{document} Feature correlation analysis To characterize the structure of the predictor space prior to model development, we computed the pairwise Pearson correlation matrix across the thirteen temporally varying inputs and the lagged yield feature. For each district-year, the monthly time series of every meteorological and remote-sensing variable was first collapsed to its June–April growing-season mean; the previous year’s yield (lagged_yield) was carried through unchanged. Pearson correlation is invariant under per-feature affine transforms, so the values reported here on the standardized inputs are numerically identical to those that would be obtained on the raw observations. The resulting matrix, computed across all 5716 district-year samples (the union of the chronological train, validation, and test splits described in Section Data Splitting), is shown in Figure 2.Fig. 2Pearson correlation matrix of the thirteen growing-season-averaged input features and the lagged yield (lagged_yield), computed across all district-year samples (2001–2023). Cells are coloured on a diverging scale; the matrix is symmetric. Pearson correlation matrix of the thirteen growing-season-averaged input features and the lagged yield (lagged_yield), computed across all district-year samples (2001–2023). Cells are coloured on a diverging scale; the matrix is symmetric. Several structural patterns are visible. First, the temperature triplet (T2M_MAX, T2M_MIN, T2M) is internally consistent but not redundant: the daily mean correlates strongly with both extremes (r = 0.92 with the max and 0.86 with the min), while the two extremes are only moderately related (r = 0.60). Second, relative humidity (RH2M) and specific humidity (QV2M) are nearly collinear (r = 0.95), both correlate strongly with precipitation (PRECTOTCORR; r = 0.83 and 0.74), and both vapour-pressure deficit measures (max_vpd, min_vpd) anti-correlate with this group at r between -0.80 and -0.95. The heat-stress axis is embedded in the same cluster, with T2M_MAX co-varying with max_vpd (r = 0.86) and opposing PRECTOTCORR and RH2M (r ≈ -0.67). Thus, thermal stress and moisture deficit are intrinsically difficult to disentangle. We also notice that NDVI and EVI are near-collinear (r = 0.95), consistent with their shared sensitivity to canopy greenness over dense wheat stands. Also, surface pressure (PS) is largely decorrelated from other weather variables (|r| ≤ 0.33), except for moderate associations with T2M (r = 0.48) and T2M_MIN (r = 0.61), suggesting that it mainly captures orographic and elevation differences across districts. The lagged yield (lagged_yield) is only weakly correlated with growing-season weather, but is moderately correlated with vegetation indices (NDVI r = 0.35, EVI r = 0.49). These collinearities would warrant pruning for a linear baseline. However, we retain all features for every benchmarked model in Section Comparative Analysis because tree-based and attention-based architectures tolerate correlated predictors, and identical feature access is required for a fair comparison. Methodology Proposed model architecture: CNN-BiLSTM with self-attention For a given district and year, we denote the monthly sequence of meteorological features as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_{\text {met}}$$\end{document}, the monthly sequence of vegetation indices (NDVI and EVI) as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_{\text {VI}}$$\end{document}, the vertically structured soil features as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_{\text {soil}}$$\end{document}, and the previous year’s yield as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_{\text {lagged\_yield}}$$\end{document}. The task is to combine these multi-modal inputs to predict the current year’s yield \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}$$\end{document}. The proposed model architecture integrates a 1D CNN for soil feature extraction, a bidirectional LSTM (BiLSTM) with self-attention for temporal processing, and a multi-layer perceptron (MLP) for final prediction. The complete architecture is illustrated in Figure 3.Fig. 3Overview of the proposed hybrid CNN-BiLSTM-Attention architecture. The model features two specialized branches: a BiLSTM with self-attention for processing temporal meteorological and VI data, and a 1D CNN for extracting features from vertical soil profiles. The feature vectors from these branches are then fused (concatenated) with lagged yield data for final prediction. Overview of the proposed hybrid CNN-BiLSTM-Attention architecture. The model features two specialized branches: a BiLSTM with self-attention for processing temporal meteorological and VI data, and a 1D CNN for extracting features from vertical soil profiles. The feature vectors from these branches are then fused (concatenated) with lagged yield data for final prediction. CNN branch for soil feature extraction The CNN branch processes vertically-structured soil features through a 1D convolutional layer that slides its kernels along the depth dimension, capturing spatial relationships between adjacent soil layers across multiple properties51. The output of the convolutional layer passes through a ReLU activation function, followed by a max-pooling layer that squeezes the depth dimension by selecting the maximum activations across the soil profile. Batch normalization is applied to stabilize training and accelerate convergence52. This network produces a fixed-size feature vector encoding a compact, learned representation of the district’s soil characteristics (4):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_{\text {soil}} = f_{\text {soil}}(\textbf{x}_{\text {soil}}) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\text {soil}}(\cdot )$$\end{document} represents the complete CNN branch (convolution, ReLU, max-pooling, and batch normalization), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{h}_{\text {soil}} \in \mathbb {R}^{d_s}$$\end{document} is the learned soil representation with dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_s$$\end{document}. BiLSTM with self-attention for temporal processing The BiLSTM with self-attention branch processes the monthly sequence of meteorological and remote sensing features. Weather variables and vegetation indices are concatenated along the feature dimension (5) to form a unified input tensor representing 11 months from June through April:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{x}_{\text {temporal}} = [\textbf{x}_{\text {met}}; \textbf{x}_{\text {VI}}] \in \mathbb {R}^{T \times d} \end{aligned}$$\end{document}where T = 11 is the sequence length. Bidirectional LSTM (BiLSTM). The temporal sequence is processed by a BiLSTM, which extends the traditional LSTM by incorporating both forward and backward temporal contexts30. The BiLSTM employs two parallel LSTM networks: one processes the sequence from June to April, while the other processes it in reverse. This bidirectional architecture captures temporal dependencies from both past and future time steps simultaneously, producing richer contextual representations. The resulting hidden states pass through layer normalization53, which normalizes activations across features to stabilize gradients and improve training dynamics. As shown in (6), 11 hidden states are generated, corresponding to each month:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{H} = f_{\text {temporal}}(\textbf{x}_{\text {temporal}}) = [\textbf{h}_1, \textbf{h}_2, \ldots , \textbf{h}_T] \in \mathbb {R}^{T \times d_h} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\text {temporal}}(\cdot )$$\end{document} represents the combined transformation through the BiLSTM and layer normalization, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{h}_t \in \mathbb {R}^{d_h}$$\end{document} represents the hidden state at time step t, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_h$$\end{document} is the hidden dimension of the BiLSTM. Self-attention mechanism. Recognizing that not all months contribute equally to the final yield, a self-attention mechanism is employed to weigh the importance of each hidden state. The mechanism uses the final BiLSTM hidden state to generate a Query vector, while all 11 hidden states generate Key and Value vectors. Attention scores are computed via the scaled dot product between Query and Keys54 (7), quantifying each month’s relevance to the final pre-harvest state:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{e} = \textbf{Q} \textbf{K}^{\top } = \frac{\textbf{h}_T \textbf{W}_q (\textbf{H} \textbf{W}_k)^{\top }}{\sqrt{d_k}} \in \mathbb {R}^{1 \times T} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_q \in \mathbb {R}^{d_h \times d_k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_k \in \mathbb {R}^{d_h \times d_k}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_v \in \mathbb {R}^{d_h \times d_v}$$\end{document} are learnable projection matrices. These scores are normalized via softmax (8) to produce attention weights:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{\alpha } = \operatorname {softmax}(\textbf{e}) = \frac{\exp (\textbf{e})}{\sum _{i=1}^{T} \exp (e_i)} \in \mathbb {R}^{1 \times T} \end{aligned}$$\end{document}The attention weights are applied to the Value vectors to compute a context vector (9), which summarizes the entire growing season with emphasis on the most influential periods:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{c} = \boldsymbol{\alpha } \textbf{V} = \sum _{i=1}^{T} \alpha _i \textbf{v}_i \in \mathbb {R}^{d_v} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{v}_i$$\end{document} is the i-th row of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V}$$\end{document}. This context vector is added to the final hidden state to produce an attention-enhanced representation of temporal dynamics:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_{\text {temporal}} = \textbf{h}_T + \textbf{c} \in \mathbb {R}^{d_h} \end{aligned}$$\end{document} Prediction head The final stage of the model fuses the learned representations from the branches and generates the yield prediction. The soil feature vector from the CNN branch, the attention-enhanced vector from the BiLSTM branch, and the previous year’s yield are concatenated into a single, comprehensive feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{h}_{\text {fused}}$$\end{document} (11):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_{\text {fused}} = [\textbf{h}_{\text {soil}}; \textbf{h}_{\text {temporal}}; \textbf{x}_{\text {lagged\_yield}}] \in \mathbb {R}^{d_s + d_h + 1} \end{aligned}$$\end{document}This vector is then passed through a multi-layer perceptron (MLP) consisting of fully-connected layers with ReLU activation and dropout layers for regularization. The MLP acts as the regression head, mapping the high-level features to a final, continuous value for the predicted yield \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}$$\end{document} (12).12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{y} = \operatorname {MLP}(\textbf{h}_{\text {fused}}) \end{aligned}$$\end{document} Experimental setup Data splitting To ensure the model is evaluated on its ability to predict future unseen years, the dataset was partitioned chronologically. While data spans 2000-2023, year 2000 was excluded from training as it lacks the prior year’ s yield feature, resulting in an effective dataset of 2001–2023 split as follows (Figure 4):Training Set: Data from 2001 to 2018 was used for model training.Validation Set: Data from 2019 to 2021 was used for hyperparameter optimization and to implement early stopping.Test Set: Data from 2022 to 2023 was held out as a final, unseen test set to evaluate the performance of the final models. Training Set: Data from 2001 to 2018 was used for model training. Training Set: Data from 2001 to 2018 was used for model training. Validation Set: Data from 2019 to 2021 was used for hyperparameter optimization and to implement early stopping. Validation Set: Data from 2019 to 2021 was used for hyperparameter optimization and to implement early stopping. Test Set: Data from 2022 to 2023 was held out as a final, unseen test set to evaluate the performance of the final models. Test Set: Data from 2022 to 2023 was held out as a final, unseen test set to evaluate the performance of the final models. Fig. 4Chronological partition of the dataset into training (2001–2018), validation (2019–2021), and test (2022–2023) sets. Chronological partition of the dataset into training (2001–2018), validation (2019–2021), and test (2022–2023) sets. This chronological split was applied consistently across all districts, preventing temporal data leakage and ensuring realistic evaluation of predictive performance. Implementation details The proposed CNN-BiLSTM with self-attention and all deep-learning baselines (1D-CNN, LSTM, BiLSTM, Transformer encoder, and parallel CNN-LSTM-Attention) were implemented in PyTorch55. The Random Forest baseline was implemented using Scikit-learn56, while the gradient-boosted baselines used the official XGBoost57 and LightGBM58 libraries. Hyperparameter optimisation across all model families was orchestrated with Optuna59, and data preprocessing and geospatial manipulation were handled using GeoPandas. All model training and experimentation were conducted on an NVIDIA T4 GPU with 16 GB of VRAM. Hyperparameter optimization and model training Hyperparameter optimization was conducted using Optuna59 with Bayesian optimization via the Tree-structured Parzen Estimator (TPE) sampler. The objective for each trial was to minimize the Mean Squared Error (MSE) on the validation set. Key hyperparameters, their search spaces, and the proposed model’s selected configuration are detailed in Appendix B. The proposed model and every benchmarked baseline (Section Comparative Analysis) were trained using the Adam optimizer60 with the Mean Squared Error (MSE), or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{2}$$\end{document}, loss as the training objective (13):13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {L}_{\text {MSE}} = \frac{1}{N} \sum _{i=1}^{N} (y_i - \hat{y}_i)^2 \end{aligned}$$\end{document}where N is the number of samples, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i$$\end{document} is the ground truth yield, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_i$$\end{document} is the predicted yield. To prevent overfitting, early stopping was employed with a patience of 15 epochs, halting training if validation loss did not improve for 15 consecutive epochs. Comparative analysis To contextualize the performance of the proposed architecture, we benchmarked it against eight competing model classes that span the principal families of approaches used in district-level yield prediction: a tree-based grid-search baseline, two gradient-boosted decision-tree (GBM) baselines, four standalone deep-sequence baselines (1D-CNN, LSTM, BiLSTM, and a Transformer encoder), and one published hybrid (parallel CNN-LSTM with attention). All eight baselines, together with the proposed model, are evaluated on the same chronological train/validation/test split (Section Data Splitting) and reported in Section Results and Discussion; their architectural specifications, best hyperparameter values, and per-baseline scatter plots are provided in Appendix C. In addition to these eight architecturally competing baselines, we report a non-ML persistence reference (the trivial forecast \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_t = y_{t-1}$$\end{document}, i.e. the lagged yield) as a sanity check against the strong inter-annual autocorrelation of district-level wheat yield identified in Section Feature Correlation Analysis. Persistence is not a competing model and is therefore not included in the comparative table; it is introduced and analyzed in Section Results and Discussion alongside the SHAP-based interpretability analysis. Ablation studies To validate our design choices and quantify the contribution of each key component within the proposed architecture, we constructed and evaluated eight model variants, starting with the full model and progressively removing or replacing its core components. To isolate the impact of each architectural change, the hyperparameters were held constant across all experiments, using the optimal set discovered for the full model. This controlled approach ensures that performance differences are directly attributable to the architectural modifications. For statistical robustness, each ablation experiment was conducted five times using different random seeds. The mean of the test metrics are reported. The following model variants were evaluated: CNN-BiLSTM-Attn (Full Model): The complete proposed architecture.CNN-BiLSTM: The full model without the self-attention mechanism.CNN-LSTM-Attn: The BiLSTM layer is replaced with a unidirectional LSTM.BiLSTM-Attn: The CNN branch and soil features are removed.BiLSTM: Both the CNN branch and the attention mechanism are removed.CNN-LSTM: The BiLSTM is replaced with a unidirectional LSTM and attention is removed.LSTM-Attn: The CNN branch is removed and a unidirectional LSTM is used.LSTM: The simplest temporal model, with only a unidirectional LSTM processing meteorological and remote sensing data. CNN-BiLSTM-Attn (Full Model): The complete proposed architecture. CNN-BiLSTM-Attn (Full Model): The complete proposed architecture. CNN-BiLSTM: The full model without the self-attention mechanism. CNN-BiLSTM: The full model without the self-attention mechanism. CNN-LSTM-Attn: The BiLSTM layer is replaced with a unidirectional LSTM. CNN-LSTM-Attn: The BiLSTM layer is replaced with a unidirectional LSTM. BiLSTM-Attn: The CNN branch and soil features are removed. BiLSTM-Attn: The CNN branch and soil features are removed. BiLSTM: Both the CNN branch and the attention mechanism are removed. BiLSTM: Both the CNN branch and the attention mechanism are removed. CNN-LSTM: The BiLSTM is replaced with a unidirectional LSTM and attention is removed. CNN-LSTM: The BiLSTM is replaced with a unidirectional LSTM and attention is removed. LSTM-Attn: The CNN branch is removed and a unidirectional LSTM is used. LSTM-Attn: The CNN branch is removed and a unidirectional LSTM is used. LSTM: The simplest temporal model, with only a unidirectional LSTM processing meteorological and remote sensing data. LSTM: The simplest temporal model, with only a unidirectional LSTM processing meteorological and remote sensing data. Evaluation metrics The performance of all models was assessed using three standard regression metrics, namely, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}). This combination of metrics was chosen for its widespread use and its ability to capture different aspects of predictive performance, from average error magnitude to the proportion of variance explained. Let n be the number of samples, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i$$\end{document} the true yield value, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_i$$\end{document} the predicted yield value, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{y}$$\end{document} the mean of the true yield values. The metrics are defined in (14)–(16):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \operatorname {MAE} = \frac{1}{n} \sum _{i=1}^{n} |y_i - \hat{y}_i |\end{aligned}$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \operatorname {RMSE} = \sqrt{\frac{1}{n} \sum _{i=1}^{n} (y_i - \hat{y}_i)^{2}} \end{aligned}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R^{2} = 1 - \frac{\sum _{i=1}^{n} (y_i - \hat{y}_i)^{2}}{\sum _{i=1}^{n} (y_i - \bar{y})^{2}} \end{aligned}$$\end{document} Results and discussion This section presents the empirical results, beginning with a benchmark of the proposed model against eight machine-learning and deep-learning baselines under identical feature access, followed by a detailed analysis of the ablation studies. We visualize the predictions of the best-performing seed on the test set, interpret the model’ s self-attention mechanism, and close with a SHAP-based feature-level attribution analysis that contextualises the architectural contribution against an explicit persistence baseline. Comparative analysis: benchmark across nine model classes Table 2 presents the performance of the proposed CNN-BiLSTM-Attention model against eight machine-learning and deep-learning baselines on the held-out test set (sowing years 2022–2023). All models share identical feature access and similar hyperparameter-tuning strategies.Table 2Performance comparison on the held-out test set (2022–2023). We report the mean ± std across five seeds; best values per column in bold. Row order is best-to-worst on MAE.ModelMAE (kg/ha)RMSE (kg/ha)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}CNN-BiLSTM-Attn273.17 ± 5.93369.96 ± 8.160.795 ± 0.009Transformer encoder284.61 ± 9.50388.58 ± 11.440.773 ± 0.014LSTM292.29 ± 20.16390.26 ± 32.200.770 ± 0.0371D-CNN303.50 ± 17.27383.77 ± 20.450.779 ± 0.023BiLSTM312.93 ± 55.72423.03 ± 73.840.725 ± 0.100TPCLA320.49 ± 27.31404.16 ± 29.290.754 ± 0.035XGBoost377.04 ± 6.43478.05 ± 8.280.657 ± 0.012LightGBM381.33 ± 4.06492.34 ± 2.790.636 ± 0.004Random Forest486.43 ± 10.78601.26 ± 12.960.458 ± 0.023 Table 2Performance comparison on the held-out test set (2022–2023). We report the mean ± std across five seeds; best values per column in bold. Row order is best-to-worst on MAE.ModelMAE (kg/ha)RMSE (kg/ha)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}CNN-BiLSTM-Attn273.17 ± 5.93369.96 ± 8.160.795 ± 0.009Transformer encoder284.61 ± 9.50388.58 ± 11.440.773 ± 0.014LSTM292.29 ± 20.16390.26 ± 32.200.770 ± 0.0371D-CNN303.50 ± 17.27383.77 ± 20.450.779 ± 0.023BiLSTM312.93 ± 55.72423.03 ± 73.840.725 ± 0.100TPCLA320.49 ± 27.31404.16 ± 29.290.754 ± 0.035XGBoost377.04 ± 6.43478.05 ± 8.280.657 ± 0.012LightGBM381.33 ± 4.06492.34 ± 2.790.636 ± 0.004Random Forest486.43 ± 10.78601.26 ± 12.960.458 ± 0.023 Performance comparison on the held-out test set (2022–2023). We report the mean ± std across five seeds; best values per column in bold. Row order is best-to-worst on MAE. The benchmark reveals a clear cleavage among the three model classes. The deep architectures cluster in the 273–321 kg/ha MAE range, the gradient-boosted baselines (XGBoost, LightGBM) cluster around 377–381 kg/ha, and Random Forest sits at 486 kg/ha — substantially behind both clusters. The proposed CNN-BiLSTM-Attention model achieves the lowest error in every column. Within the deep cluster, three contrasts merit attention. The Transformer encoder (replicating the encoder core of Guo et al.29) is the closest competitor at 284.6 kg/ha MAE — only ≈4% behind the proposed model. The Parallel CNN-LSTM-Attention architecture of Song et al.17, which routes the temporal stream through both CNN and LSTM branches in parallel rather than reserving the CNN branch for the soil tensor, lags by 14.8% MAE (320.5 kg/ha); we read this as direct empirical support for the modality-specific routing argument made in Section Survey of current state-of-the-art and Identification of Research Gap. The standalone LSTM with flattened soil features attains 292.3 kg/ha and is the strongest pure-RNN baseline and only ≈7% behind our model, bounding the marginal architectural value above a well-tuned LSTM. A cautionary note on training stability: the standalone BiLSTM with flat-soil inputs exhibits unusually high seed-variance (MAE std 55.7 versus 5.9 for our model), suggesting that combining a deep BiLSTM stack with a flat-soil input is difficult to train reliably at this data scale; routing the soil tensor through a dedicated 1D-CNN branch appears to materially improve optimisation stability. Per-baseline scatter plots, architecture details, and best-hyperparameter values are catalogued in Appendix C. The relationship of these ML/DL results to a non-ML persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) is examined separately in Section Feature-Level Interpretability and the Persistence Baseline. Ablation study: deconstructing model performance To validate each component of the proposed architecture, we conducted a series of ablation studies. Table 3 summarizes the performance of eight model variants, with results averaged over five random seeds.Table 3Ablation-study results on the test set, sorted by RMSE. Δ RMSE is relative to the full model. Best values per column in bold.Model variantMAE (kg/ha)RMSE (kg/ha)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\Delta }$$\end{document} RMSE vs. full model (kg/ha)CNN-BiLSTM-Attn273.171369.9630.795-CNN-BiLSTM278.466382.7320.78012.769CNN-LSTM-Attn279.451384.7020.77814.739BiLSTM289.254390.7890.77120.826CNN-LSTM296.464392.1110.76922.148BiLSTM-Attn291.905396.4260.76426.463LSTM-Attn304.105408.9620.74938.999LSTM304.873414.0750.74144.112 Table 3Ablation-study results on the test set, sorted by RMSE. Δ RMSE is relative to the full model. Best values per column in bold.Model variantMAE (kg/ha)RMSE (kg/ha)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\Delta }$$\end{document} RMSE vs. full model (kg/ha)CNN-BiLSTM-Attn273.171369.9630.795-CNN-BiLSTM278.466382.7320.78012.769CNN-LSTM-Attn279.451384.7020.77814.739BiLSTM289.254390.7890.77120.826CNN-LSTM296.464392.1110.76922.148BiLSTM-Attn291.905396.4260.76426.463LSTM-Attn304.105408.9620.74938.999LSTM304.873414.0750.74144.112 Ablation-study results on the test set, sorted by RMSE. Δ RMSE is relative to the full model. Best values per column in bold. The ablation study results (Table 3) confirm that the synergistic combination of all architectural components is essential for optimal performance. The full CNN-BiLSTM-Attention model achieves the lowest RMSE, and removing any component degrades performance, validating our design. The analysis reveals each module’ s distinct contribution. Removing the CNN branch yields the most significant performance drop among primary components, increasing RMSE by 27 kg/ha. This underscores the importance of soil features for accurate yield prediction. The value of capturing full temporal context is evidenced by comparing bidirectional and unidirectional processing since replacing the BiLSTM with a unidirectional LSTM increases RMSE by 15 kg/ha. Finally, the attention mechanism proves refines the model’ s temporal focus, with its removal increasing error by 13 kg/ha. Collectively, these results demonstrate that the proposed architecture’ s superior performance stems not from any single innovation, but from the integration of complementary components, each addressing a specific aspect of the multi-modal yield prediction challenge. Final model evaluation and limitations We evaluate the best-performing seed of the CNN-BiLSTM-Attention configuration on the held-out test set; the optimal hyperparameters are detailed in Appendix B. On the 2022–2023 test set the best seed achieves an MAE of 265.43 kg/ha, RMSE of 358.57 kg/ha, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} of 0.807 and Pearson r = 0.899. These best-seed values bracket the upper end of the seed distribution and are reported here to support the residual and spatial diagnostics that follow. Figure 5 plots predicted versus actual yield (observed yield) for each data point in the test set. To assess spatial prediction patterns, Figure 6 presents three maps visualizing the model’ s performance across India’ s wheat belt for the 2023 sowing season (harvested April–May 2024) with the actual yields observed on the left, predicted yields in the center, and the residuals plotted on the right. The residual map shows that the model underpredicts yields in highly productive regions like Punjab (red) while overpredicting in lower-performing districts (blue), suggesting regression toward the mean.Fig. 5Predicted vs. actual yield on the test set. The dashed line represents the 1:1 identity line.Fig. 6Spatial assessment of model performance for the 2023 sowing year (winter wheat harvested April–May 2024). The maps compare actual yields (left) against predicted yields (center) in kg/ha, while the residuals map (right) visualizes the error distribution (actual − predicted), where red indicates underprediction and blue indicates overprediction. Predicted vs. actual yield on the test set. The dashed line represents the 1:1 identity line. Spatial assessment of model performance for the 2023 sowing year (winter wheat harvested April–May 2024). The maps compare actual yields (left) against predicted yields (center) in kg/ha, while the residuals map (right) visualizes the error distribution (actual − predicted), where red indicates underprediction and blue indicates overprediction. To quantify this pattern, we carried out a state-wise residual analysis. Key metrics for each state are displayed in Table 4. The analysis confirms that high-yield, intensively managed states such as Punjab and Haryana are systematically underpredicted, with mean residuals of approximately 440 kg/ha and 110 kg/ha, respectively, despite mean observed yields above 4.5–4.8 t/ha. In contrast, more water-limited states including Rajasthan and Madhya Pradesh are consistently overpredicted by around 190 kg/ha and 235 kg/ha, while Uttar Pradesh, Gujarat, and Bihar show comparatively small mean residuals (between approximately −60 and +40 kg/ha). The distribution of residuals for each state and the model’s tendency to overpredict yield for some states and underpredict for others is also demonstrated through box plots of prediction residuals shown in Figure 7.Table 4State-wise performance metrics for yield prediction on the test set (sowing years 2022–2023). Residuals are computed as observed − predicted yield; positive values indicate underprediction, negative values overprediction. States are ordered by mean observed yield (highest to lowest).StateMean observed (kg/ha)Mean predicted (kg/ha)Mean residual (kg/ha)Punjab4815.34375.9439.4Haryana4551.84443.2108.6Uttar Pradesh3743.83706.037.8Rajasthan3361.33551.5−190.3Madhya Pradesh3205.43439.2−233.9Gujarat3099.43160.9−61.4Bihar3023.42993.629.8 Table 4State-wise performance metrics for yield prediction on the test set (sowing years 2022–2023). Residuals are computed as observed − predicted yield; positive values indicate underprediction, negative values overprediction. States are ordered by mean observed yield (highest to lowest).StateMean observed (kg/ha)Mean predicted (kg/ha)Mean residual (kg/ha)Punjab4815.34375.9439.4Haryana4551.84443.2108.6Uttar Pradesh3743.83706.037.8Rajasthan3361.33551.5−190.3Madhya Pradesh3205.43439.2−233.9Gujarat3099.43160.9−61.4Bihar3023.42993.629.8 State-wise performance metrics for yield prediction on the test set (sowing years 2022–2023). Residuals are computed as observed − predicted yield; positive values indicate underprediction, negative values overprediction. States are ordered by mean observed yield (highest to lowest). Fig. 7Box plots of prediction residuals (actual − predicted yield, kg/ha) across the states in the test dataset. The dashed red line at zero indicates perfect prediction; values below zero indicate underprediction, values above indicate overprediction. States are ordered by mean actual yield (highest to lowest). Box plots of prediction residuals (actual − predicted yield, kg/ha) across the states in the test dataset. The dashed red line at zero indicates perfect prediction; values below zero indicate underprediction, values above indicate overprediction. States are ordered by mean actual yield (highest to lowest). These residual patterns reflect both the available input features and the structure of the Indian wheat production system. Punjab and Haryana are characterized by very high irrigation intensity (often >90% irrigated wheat), high input use, and relatively stable yields9,61; however, the model does not include explicit irrigation or field-level management variables, and instead relies on district-average weather, satellite vegetation indices, soil properties, and a lagged yield feature. Because both irrigated and rain-fed wheat can exhibit similar peak NDVI/EVI under favorable conditions, the model can only partially distinguish high-potential irrigated systems from more marginal rain-fed districts, tending to pull exceptionally high yields downward toward the historical mean. At the same time, the yield-bin distribution in Table 5 shows that the training data are dominated by medium-yield observations, while the validation and test sets are skewed toward high-yield cases with very few low-yield examples. Together, the absence of explicit irrigation and cropland-fraction information, the class imbalance across yield bins, and the strong predictive influence of lagged yield naturally lead to the observed underprediction in highly productive, irrigated regions and overprediction in low-yield, stress-prone districts.Table 5Distribution of yield classes across train, validation, and test splits. Yield bins are defined as low (<2000 kg/ha), medium (2000–3500 kg/ha), and high (>3500 kg/ha). The test set is heavily skewed toward high-yield observations, while the training set is more balanced. This class imbalance contributes to the model’s regression-to-the-mean behaviour, as it is trained primarily on medium-yield samples but evaluated predominantly on high-yield cases. Within-split percentages in parentheses.SplitLowMediumHighTotalTrain815 (18.7%)2287 (52.4%)1266 (29.0%)4368Validation23 (2.9%)332 (41.3%)448 (55.8%)803Test10 (1.8%)247 (45.3%)288 (52.8%)545 Table 5Distribution of yield classes across train, validation, and test splits. Yield bins are defined as low (<2000 kg/ha), medium (2000–3500 kg/ha), and high (>3500 kg/ha). The test set is heavily skewed toward high-yield observations, while the training set is more balanced. This class imbalance contributes to the model’s regression-to-the-mean behaviour, as it is trained primarily on medium-yield samples but evaluated predominantly on high-yield cases. Within-split percentages in parentheses.SplitLowMediumHighTotalTrain815 (18.7%)2287 (52.4%)1266 (29.0%)4368Validation23 (2.9%)332 (41.3%)448 (55.8%)803Test10 (1.8%)247 (45.3%)288 (52.8%)545 Distribution of yield classes across train, validation, and test splits. Yield bins are defined as low (<2000 kg/ha), medium (2000–3500 kg/ha), and high (>3500 kg/ha). The test set is heavily skewed toward high-yield observations, while the training set is more balanced. This class imbalance contributes to the model’s regression-to-the-mean behaviour, as it is trained primarily on medium-yield samples but evaluated predominantly on high-yield cases. Within-split percentages in parentheses. Interpreting model behavior: insights from self-attention To move beyond performance metrics and understand the model’ s decision-making process, we analyzed the learned self-attention weights. Figure 8 visualizes the average attention scores assigned to each month across all test samples, revealing which periods the model emphasizes during yield prediction.Fig. 8Aggregated temporal attention weights averaged across all test samples. Bars show the mean attention weight per month; background shading indicates Indian winter wheat phenological stages. Aggregated temporal attention weights averaged across all test samples. Bars show the mean attention weight per month; background shading indicates Indian winter wheat phenological stages. The analysis yields three insights that align with agronomic knowledge. First, a dominant attention peak occurs in November, corresponding to the sowing and establishment phase for winter wheat in India. This suggests the model has learned that conditions during this critical window are highly predictive of final yield, consistent with established agronomic understanding that early-season establishment strongly influences crop outcomes62. Second, a smaller but distinct peak appears in June. June marks the monsoon onset, which determines pre-sowing soil moisture, a crucial initial condition for winter wheat planted five months later63. The model’ s attention to this period indicates it has learned to connect pre-seasonal moisture availability to eventual yield. Third, the near-zero attention weights assigned to March and April reveal an interesting interaction between the BiLSTM and attention mechanism. Since the BiLSTM’ s final hidden state already encodes recent temporal information through its recurrent processing, the attention mechanism learns to emphasize earlier, more distinctive periods like November rather than redundantly focusing on immediately preceding months. This suggests the architecture effectively divides labor: the BiLSTM captures sequential context while attention highlights critical distant time steps. Feature-level interpretability and the persistence baseline To complement the temporal-attention analysis of Section Interpreting Model Behavior: Insights from Self-Attention and to test the natural concern that a model with access to a lagged yield input might collapse to a persistence forecaster, we conducted a SHAP attribution analysis on the CNN-BiLSTM-Attention model and compared its predictions against an explicit persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}). SHAP values were computed using a GradientExplainer with 200 background samples and evaluated on all 545 test observations. Figure 9 summarises the per-feature mean absolute SHAP magnitudes. At face value the lagged yield scalar dominates, but a like-for-like comparison must aggregate over the high-dimensional temporal and soil tensors. Summed within input groups, the lagged yield accounts for approximately 690 units of mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\text {SHAP}|$$\end{document}, the full 11-month × 13-feature temporal block accounts for ≈ 218 units, and the 8-property × 6-depth soil block accounts for ≈ 93 units, yielding a lagged yield : temporal : soil ratio of roughly 3 : 1 : 0.4 rather than the ∼100:1 ratio implied by the raw bar chart.Fig. 9Mean absolute SHAP values per input feature on the test set (n = 545, GradientExplainer with 200 background samples). The lagged yield scalar (lagged_yield) dominates the per-feature ranking; aggregated within input groups, however, the temporal and soil tensors account for substantial attribution mass (see text). Mean absolute SHAP values per input feature on the test set (n = 545, GradientExplainer with 200 background samples). The lagged yield scalar (lagged_yield) dominates the per-feature ranking; aggregated within input groups, however, the temporal and soil tensors account for substantial attribution mass (see text). To directly test whether the architecture is doing more than reproducing the lagged yield input, we evaluated the persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) on the same held-out test set (Table 6). Aggregated over both test years, persistence achieves MAE 274.9 kg/ha and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} 0.769, essentially tying the five-seed mean of the proposed model on MAE (273.2 kg/ha) and trailing it by 0.026 in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}. The best-seed configuration improves on persistence by 3.4% in MAE and 8.6% in RMSE, indicating that the architecture preferentially reduces large errors rather than uniformly shifting the error distribution.Table 6Persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) versus the proposed CNN-BiLSTM-Attention model on the held-out test set (2022–2023). Persistence essentially ties the five-seed mean on MAE, motivating the per-year decomposition in Table 7 and the SHAP analysis that follows.PredictorMAE (kg/ha)RMSE (kg/ha)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}Persistence (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document})274.90392.420.769CNN-BiLSTM-Attn (5-seed mean)273.17369.960.795CNN-BiLSTM-Attn (best seed)265.43358.570.807 Table 6Persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) versus the proposed CNN-BiLSTM-Attention model on the held-out test set (2022–2023). Persistence essentially ties the five-seed mean on MAE, motivating the per-year decomposition in Table 7 and the SHAP analysis that follows.PredictorMAE (kg/ha)RMSE (kg/ha)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}Persistence (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document})274.90392.420.769CNN-BiLSTM-Attn (5-seed mean)273.17369.960.795CNN-BiLSTM-Attn (best seed)265.43358.570.807 Persistence baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) versus the proposed CNN-BiLSTM-Attention model on the held-out test set (2022–2023). Persistence essentially ties the five-seed mean on MAE, motivating the per-year decomposition in Table 7 and the SHAP analysis that follows. Stratifying the comparison by sowing year sharpens the picture (Table 7). In the climatically normal 2022 season, persistence and the model are statistically indistinguishable on MAE (222.7 vs. 227.4 kg/ha); the model improves RMSE by only 4.0%. In the dry 2023 sowing season, however, the model improves MAE by 7.2% (303.6 vs. 327.3 kg/ha), RMSE by 11.2% (398.9 vs. 449.0 kg/ha), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document} by 0.061. The proposed model therefore behaves as a persistence forecaster in stable years and as a stress-aware predictor in anomalous years; the lagged yield carries the trend, while the temporal and soil features carry the anomaly signal.Table 7Per-year persistence vs. CNN-BiLSTM-Attention (best seed) on the test set. Δ columns report the percentage MAE/RMSE reduction (positive: model improves over persistence) and the absolute change in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}. The architectural value-add concentrates in the dry 2023 sowing season.PeriodnPersistenceCNN-BiLSTM-AttnΔ (model − persistence)MAERMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}MAERMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}MAE%RMSE%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\Delta R^{2}}$$\end{document}All545274.9392.40.769265.4358.60.807+3.4+8.6+0.0382022273222.7326.40.832227.4313.30.845−2.1+4.0+0.0132023272327.3449.00.711303.6398.90.772+7.2+11.2+0.061 Table 7Per-year persistence vs. CNN-BiLSTM-Attention (best seed) on the test set. Δ columns report the percentage MAE/RMSE reduction (positive: model improves over persistence) and the absolute change in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}. The architectural value-add concentrates in the dry 2023 sowing season.PeriodnPersistenceCNN-BiLSTM-AttnΔ (model − persistence)MAERMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}MAERMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{R^{2}}$$\end{document}MAE%RMSE%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\Delta R^{2}}$$\end{document}All545274.9392.40.769265.4358.60.807+3.4+8.6+0.0382022273222.7326.40.832227.4313.30.845−2.1+4.0+0.0132023272327.3449.00.711303.6398.90.772+7.2+11.2+0.061 Per-year persistence vs. CNN-BiLSTM-Attention (best seed) on the test set. Δ columns report the percentage MAE/RMSE reduction (positive: model improves over persistence) and the absolute change in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}$$\end{document}. The architectural value-add concentrates in the dry 2023 sowing season. The mechanistic basis of the anomaly-year gain is clarified by the SHAP feature × month heatmap (Figure 10). The dominant temporal attribution mass concentrates in February–March — the grain-filling and maturation window for north-Indian winter wheat — and is led by EVI and NDVI, precisely the months in which moisture deficits and warm anomalies can suppress yield. A secondary February–March ridge in surface pressure (PS) is consistent with synoptic-regime mediation of the dry-season anomaly affecting this window.Fig. 10SHAP feature × month heatmap. Cells report mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\text {SHAP}|$$\end{document} for each input feature in each calendar month. The dominant attribution ridge in February–March, led by EVI and NDVI, coincides with the grain-filling and maturation window for north-Indian winter wheat. SHAP feature × month heatmap. Cells report mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\text {SHAP}|$$\end{document} for each input feature in each calendar month. The dominant attribution ridge in February–March, led by EVI and NDVI, coincides with the grain-filling and maturation window for north-Indian winter wheat. Finally, the depth-importance decomposition (Figure 11) shows that the soil-branch attribution increases monotonically with depth, from ≈ 0.2 in the 0–5 cm layer to ≈ 41.6 in the 100–200 cm layer (a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 200\times$$\end{document} ratio). Deep-profile water and nutrient reserves drive the soil branch far more than the topsoil layers, defending the soil pathway against the alternative hypothesis that it contributes negligibly to model behaviour.Fig. 11Mean absolute SHAP per soil depth layer, aggregated across the eight soil properties. Importance increases monotonically with depth, indicating that deep-profile water and nutrient reserves dominate the soil-branch contribution. Mean absolute SHAP per soil depth layer, aggregated across the eight soil properties. Importance increases monotonically with depth, indicating that deep-profile water and nutrient reserves dominate the soil-branch contribution. Conclusion This study presents a controlled benchmark of nine model classes for district-level winter-wheat yield prediction across India’ s wheat belt — Random Forest, XGBoost, LightGBM, a 1D-CNN, an LSTM, a BiLSTM, a single-stream Transformer encoder, the recently proposed Parallel CNN-LSTM-Attention design, and our hybrid CNN-BiLSTM-Attention with modality-specific routing — under identical feature access, an identical chronological split, an identical five-seed evaluation, and an identical Optuna hyperparameter-optimisation budget per model. The proposed CNN-BiLSTM-Attention model is the best of the nine entries, explaining 79.5% of the variance in district-level winter wheat yield (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2} = 0.795$$\end{document}, mean over five seeds). Ablation studies further demonstrate that the CNN soil-routing branch, the bidirectional temporal pathway, and the self-attention mechanism each contribute essential and non-redundant predictive capacity to the hybrid; the model’s performance derives from the synergistic integration of complementary modules rather than from any single component. A persistence forecast (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y} = \text {lagged\_yield}$$\end{document}) achieves an MAE of 274.9 kg/ha on the same held-out test set, essentially tying the proposed model on average. The architectural value-add concentrates in anomalous years: in the dry 2023 sowing season the model improves MAE by 7.2% and RMSE by 11.2% over persistence, and SHAP attribution localises the temporal contribution to the February–March grain-filling and maturation window led by EVI and NDVI signal. We read this as evidence that the lagged yield input carries the inter-annual trend, while the temporal and soil features carry the anomaly signal that becomes decisive precisely when persistence breaks down. Future work should extend the benchmark to additional crops and geographies (in particular kharif rice and post-monsoon rabi pulses, where year-on-year autocorrelation is weaker), incorporate explicit irrigation and field-level management variables, and evaluate longer prediction horizons. Methodologically, the modality-specific routing paradigm and the persistence-aware evaluation protocol introduced here can serve as a template for benchmarks of deep learning architectures in other data-scarce, smallholder-dominated agricultural settings.