Vellore Institute of Technology (Chennai, India) |
Taylor's University (Subang Jaya, Malaysia) |
An Explainable Multimodal Data Fusion Framework for Compatibility Reasoning in Heterogeneous Decision Systems
🌾 AgriCompatNet: Knowledge-Guided Machine Learning & Explainable Decision Support for Intercropping Systems
Published in Elsevier Array (Volume 32, Article 101267, 2026) — Special Issue on Explainable Decision Systems (EDS)
[Official ScienceDirect Article] •
[Download Open-Access PDF] •
[Executable Jupyter Notebook] •
[Cite This Work]
Important
Headline Empirical Findings:
AgriCompatNet achieves a state-of-the-art predictive performance of
Figure 1: High-level architectural workflow of the AgriCompatNet framework, illustrating multimodal agro-environmental ingestion, domain-informed indicator construction, ensemble fusion, and post-hoc SHAP explainability.
- Overview & Core Motivation
- International Collaboration & Authors
- Six-Stage Pipeline Architecture
- Mathematical Formulation & Agronomic Indicators
- Experimental Benchmarks & Validation
- Explainable AI (XAI) & Interpretability Suite
- Visual Results & Empirical Gallery
- Repository Structure
- Quick Start & Reproduction
- Citation
- License & Open Access
Sustainable agriculture and agroecological intensification depend heavily on intercropping—cultivating two or more crop varieties simultaneously on the same parcel of land. While intercropping enhances biodiversity, curtails pest proliferation, optimizes soil microbiome health, and augments land equivalent ratio (LER), discovering mutually synergistic crop combinations across varied microclimates has traditionally relied on trial-and-error field trials or opaque, black-box recommendation models.
AgriCompatNet addresses this challenge through a knowledge-guided multimodal data fusion framework:
-
Decoupling First-Order Heuristics from Emergent Interaction Modeling: Rather than relying solely on deterministic formulas or unconstrained regression on raw sensor feeds, AgriCompatNet establishes an intermediate abstraction layer consisting of four domain-grounded agronomic indicators (
$\text{NBI}$ ,$\text{ERS}$ ,$\text{NutComp}$ ,$\text{CP}$ ). - Preventing Information & Target Leakage: Strict partitioning and evaluation safeguards prevent formula memorization, allowing tree ensembles (Random Forest, LightGBM, CatBoost) to learn non-linear co-adaptation boundaries and ecological constraints.
- Intrinsic Explainable AI (XAI): Full integration of Shapley Additive Explanations (SHAP), Partial Dependence Plots (PDP), Individual Conditional Expectation (ICE), and pairwise interaction analysis provides agronomic transparency for domain experts and farm decision-support systems (DSS).
This scientific investigation was conducted under an international bilateral research collaboration between Vellore Institute of Technology (VIT, India) and Taylor's University (Malaysia):
| Author | Institutional Affiliation | Contact & Profiles |
|---|---|---|
| Gokul Ram K. | School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India | gokulram.k2023@vitstudent.ac.in |
| Dr. U. Vignesh (Corresponding Author) |
School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India | vignesh.u@vit.ac.in |
| Dr. Geetha S. | School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India | geetha.s@vit.ac.in |
| Dr. Parvathi R. | Centre for Advanced Data Science, Vellore Institute of Technology, Chennai, India | parvathi.r@vit.ac.in |
| Assoc. Prof. Abdulkareem Sh. Mahdi Al-Obaidi | School of Engineering, Faculty of Innovation and Technology, Taylor’s University, Subang Jaya, Malaysia | abdulkareem.mahdi@taylors.edu.my |
The overall methodology formalized in Algorithm 1 of the published paper is structured into six modular, interconnected stages:
flowchart TD
%% Styling Definitions
classDef inputStyle fill:#e1f5fe,stroke:#0288d1,stroke-width:2px,color:#01579b;
classDef prepStyle fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px,color:#4a148c;
classDef indicatorStyle fill:#e8f5e9,stroke:#388e3c,stroke-width:2px,color:#1b5e20;
classDef modelStyle fill:#fff3e0,stroke:#f57c00,stroke-width:2px,color:#e65100;
classDef fusionStyle fill:#fbe9e7,stroke:#d84315,stroke-width:2px,color:#bf360c;
classDef xaiStyle fill:#ede7f6,stroke:#512da8,stroke-width:2px,color:#311b92;
%% Stage 1
subgraph S1["Stage 1: Raw Agro-Environmental Inputs"]
D1["Soil Chemical Profile<br/>(N, P, K Macronutrients)"]:::inputStyle
D2["Crop Physiology & Edaphic<br/>(Soil pH, Growth Dynamics)"]:::inputStyle
D3["Microclimatic Factors<br/>(Temp, Relative Humidity, Rain)"]:::inputStyle
end
%% Stage 2
subgraph S2["Stage 2: Harmonization & Multivariate Outlier Removal"]
H1["Scale Invariance Normalization<br/>(Min-Max to [0, 1])"]:::prepStyle
H2["Anomaly Filtering & Outlier Capping<br/>(Interquartile Range & Z-scores)"]:::prepStyle
H3["Collinearity & Redundancy Pruning<br/>(Pearson Correlation Screening)"]:::prepStyle
end
%% Stage 3
subgraph S3["Stage 3: Agronomic Indicator Construction"]
I1["Nutrient Balance Index (NBI)<br/>Proportional Stoichiometric Alignment"]:::indicatorStyle
I2["Environmental Resilience Score (ERS)<br/>Exponential Climatic Co-Adaptation"]:::indicatorStyle
I3["Nutrient Complementarity (NutComp)<br/>Strata Resource-Use Synergies"]:::indicatorStyle
I4["Conflict Penalty (CP)<br/>Resource Niche Overlap & Competition"]:::indicatorStyle
end
%% Stage 4
subgraph S4["Stage 4: Ensemble Compatibility Learning"]
M1["Random Forest Regressor<br/>(Bagging Variance Reduction)"]:::modelStyle
M2["LightGBM Regressor<br/>(Histogram Gradient Boosting)"]:::modelStyle
M3["CatBoost Regressor<br/>(Symmetric Oblivious Trees)"]:::modelStyle
end
%% Stage 5
subgraph S5["Stage 5: Weighted Fusion & Logistic Calibration"]
F1["Meta-Learner & Stacking Fusion<br/>(Non-linear Interaction Mapping)"]:::fusionStyle
F2["Calibrated Compatibility Index (CCI)<br/>Continuous Bimodal Scale [0, 1]"]:::fusionStyle
end
%% Stage 6
subgraph S6["Stage 6: Explainability & Knowledge Validation"]
X1["Global & Local SHAP Attribution<br/>(Summary Beeswarm, Waterfall Plots)"]:::xaiStyle
X2["Partial Dependence & ICE Plots<br/>(Marginal Effects & Saturation Profiles)"]:::xaiStyle
X3["Topological Graph & Manifold Embeddings<br/>(NetworkX Graphs, PCA & UMAP Projections)"]:::xaiStyle
end
%% Connections
D1 & D2 & D3 --> H1 --> H2 --> H3
H3 --> I1 & I2
H3 --> I3 & I4
I1 & I2 & I3 & I4 --> M1 & M2 & M3
M1 & M2 & M3 --> F1 --> F2
F2 --> X1 & X2 & X3
AgriCompatNet constructs four domain-grounded composite indicators to model the biophysical and ecological interplay between crop pairs
Quantifies how closely soil macronutrient proportions match agronomic optimal ratios, penalizing both deficiency and luxury consumption:
Where $r_k$ represents the observed soil nutrient ratio (e.g., $N:P$, $N:K$, $P:K$), $r_k^{opt}$ is the empirically established optimal stoichiometric benchmark, and $m$ is the total number of evaluated nutrient pairs.
Measures climatic co-adaptation under shared temperature, relative humidity, and precipitation variability via an exponential decay formulation:
Where $E_{a,t}$ and $E_{b,t}$ represent environmental tolerance requirements along dimension $t$, and $\sigma_{E_t}$ denotes the population-wide variance, ensuring multi-attribute scale invariance.
Measures reciprocal resource-use partitioning, assessing whether one crop’s high demand is balanced by the partner crop’s modest requirement:
Where $U_{a,k}, U_{b,k} \in [0, 1]$ represent min-max normalized uptake intensities for macronutrient $k$ across the total evaluated nutrients $K$.
Captures resource contention and ecological friction resulting from overlapping rhizosphere niches and divergent climatic regimes:
Where $w_k$ denotes the agronomic importance weight of nutrient $k$, $\rho(E_a, E_b)$ is the correlation coefficient across environmental responses, and $\beta$ balances edaphic vs climatic conflict.
The final target metric combines these indicators into a unified score normalized to the unit interval
The framework was comprehensively benchmarked against standard decision models on a curated dataset of 2,200 crop instances across 2.4 million generated crop-pair combinations:
| Model Architecture | Model Paradigm |
|
RMSE (Root Mean Squared Error) | MAE (Mean Absolute Error) | Inference Characteristics |
|---|---|---|---|---|---|
| Random Forest | Bagging Ensemble | 0.91 |
0.049 |
0.031 |
High variance control, tree diversity |
| LightGBM | Histogram Boosting | 0.93 |
0.045 |
0.029 |
Fast gradient leaf-wise splitting |
| CatBoost | Symmetric Oblivious Trees | 0.90 |
0.051 |
0.033 |
Strong regularization, categorical handling |
| AgriCompatNet (Fusion) | Multimodal Knowledge-Guided Ensemble | 0.96 |
0.039 |
0.027 |
State-of-the-art calibration & domain fidelity |
Note
Ablation Insight: Systematic feature exclusion confirms that removing the Conflict Penalty (
| Pair Rank | Crop A | Crop B | CCI Score | Agronomic Interaction Mechanism | Perturbation Stability ( |
|---|---|---|---|---|---|
0 |
Papaya | Papaya | 0.9975 |
Homogeneous control baseline / self-affinity | High baseline stability |
1 |
Mungbean | Mungbean | 0.9942 |
Leguminous atmospheric nitrogen self-sufficiency | Extremely robust |
2 |
Coconut | Coconut | 0.9925 |
Deep vertical rooting system equilibrium | Resilient |
4 |
Mungbean | Blackgram | 0.9902 |
Stratified root zones, complementary |
Stable ( |
7 |
Rice | Jute | 0.9890 |
Shared hydromorphic waterlogged niche, complementary vegetative canopy heights | Stable ( |
Interpretability in AgriCompatNet is treated as an active model-validation tool rather than an afterthought:
-
Global SHAP Feature Importance: Identifies Conflict Penalty (
$\text{CP}$ ) as the primary determinant of compatibility variance, followed by Nutrient Complementarity ($\text{NutComp}$ ) and Nutrient Balance ($\text{NBI}$ ). - Local SHAP Waterfall Explanations: Deconstructs specific pairwise interactions (e.g., Mungbean–Blackgram) to reveal positive contributions from minimal edaphic friction.
-
Partial Dependence & ICE Profiles: Demonstrate monotonic, saturation-governed responses, proving that beyond threshold values of
$\text{ERS}$ , additional resilience yields diminishing returns. -
Second-Order Interaction Modeling: Uncovers strong non-linear cross-talk between
$\text{CP} \times \text{NBI}$ , where favorable nutrient stoichiometry cannot overcome intense competitive conflict.
All figures extracted directly from the verified pipeline outputs:
AgriCompatNet/
├── .gitignore # Python, Jupyter, virtualenv exclusions
├── CITATION.cff # Citation File Format (v1.2.0) metadata
├── LICENSE # Dual License: MIT (Code) & CC BY 4.0 (Assets)
├── README.md # Comprehensive repository documentation
├── requirements.txt # Pinned dependency requirements
├── images/
│ ├── 1_Graphical_Insights.png # High-resolution framework architecture diagram
│ ├── badges/ # Crisp local SVG status badges
│ │ ├── badge_elsevier_array.svg
│ │ ├── badge_doi.svg
│ │ ├── badge_license.svg
│ │ ├── badge_xai.svg
│ │ ├── badge_benchmark.svg
│ │ ├── badge_collaboration.svg
│ │ └── badge_python.svg
│ ├── logos/ # Institutional logos
│ │ ├── vit_chennai_logo.png
│ │ └── taylors_university_logo.png
│ ├── fig02_feature_normalization.png
│ ├── fig03_outlier_detection_boxplots.png
│ ├── fig04_environmental_sensitivity.png
│ ├── fig05_cci_distribution.png
│ ├── fig06_compatibility_heatmap.png
│ ├── fig07_crop_compatibility_network.png
│ ├── fig08_global_shap_importance.png
│ ├── fig09_partial_dependence_plots.png
│ ├── fig10_local_shap_waterfall_pair1.png
│ ├── fig11_shap_interaction_cp_nbi.png
│ ├── fig12_pca_latent_clusters.png
│ ├── fig13_feature_correlation_matrix.png
│ ├── fig14_shap_beeswarm_impact.png
│ └── ice_individual_conditional_expectation.png
├── notebooks/
│ └── agricompatnet_pipeline.ipynb # Complete 39-cell end-to-end research notebook
└── resources/
└── AgriCompatNet_Elsevier_Array_2026.pdf # Full Elsevier published open-access paper
git clone https://github.com/GOKULRAM-K/AgriCompatNet.git
cd AgriCompatNetpython -m venv venv
# On Linux/macOS:
source venv/bin/activate
# On Windows:
venv\Scripts\activatepip install --upgrade pip
pip install -r requirements.txtLaunch the Jupyter Notebook to reproduce all 6 stages, figures, and benchmark metrics:
jupyter notebook notebooks/agricompatnet_pipeline.ipynbIf you use AgriCompatNet, its algorithmic indicators (
BibTeX Format (Click to Expand / Collapse)
@article{GOKULRAM2026101267,
title = {An explainable multimodal data fusion framework for compatibility reasoning in heterogeneous decision systems},
journal = {Array},
volume = {32},
pages = {101267},
year = {2026},
issn = {2590-0056},
doi = {10.1016/j.array.2026.101267},
url = {https://doi.org/10.1016/j.array.2026.101267},
author = {Gokul Ram K. and U. Vignesh and Geetha S. and Parvathi R. and Abdulkareem Sh. Mahdi Al-Obaidi},
keywords = {Explainable artificial intelligence, Multimodal data fusion, Decision support systems, Crop compatibility, Ensemble learning, Precision agriculture},
note = {Special Issue: EDS}
}APA 7th Edition Format
Gokul Ram, K., Vignesh, U., Geetha, S., Parvathi, R., & Al-Obaidi, A. S. M. (2026). An explainable multimodal data fusion framework for compatibility reasoning in heterogeneous decision systems. Array, 32, Article 101267. https://doi.org/10.1016/j.array.2026.101267
- Software & Notebook Code: Released under the permissive MIT License.
- Research Assets, Figures, & Indicators: Published open-access by Elsevier Inc. under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
- Corresponding Inquiries: Dr. U. Vignesh (
vignesh.u@vit.ac.in) & Gokul Ram K. (gokulram.k2023@vitstudent.ac.in).













