Researchers have developed ExplainableCropAI, an interpretable deep learning decision-support system designed to diagnose paddy crop diseases while offering inspectable visual and textual explanations. Deep neural networks frequently achieve strong diagnostic accuracy in agriculture, but their black-box nature creates hesitation among farmers and agricultural extension officers who must understand the reasoning behind an assessment before taking corrective action in the field. To resolve this operational barrier, the new framework integrates multiple diagnostic verification channels directly alongside automated image classification.
The pipeline couples a ConvNeXt V2-Tiny image classifier with three complementary explanation mechanisms. Visual interpretability is provided through Grad-CAM++ attribution maps that highlight decisive image regions, while diagnostic reliability is quantified using Monte Carlo Dropout uncertainty paired with temperature-scaled calibration. Additionally, the system generates natural-language explanations using the Phi-3-mini model. To prevent ungrounded outputs, text generation is restricted by a strict grounding constraint that limits the language model to typed fields drawn exclusively from model predictions, visual attribution heatmaps, and a curated agronomic knowledge base.
The research team emphasized that individual architectural components are established tools, placing the primary scientific contribution on their unified integration and rigorous empirical assessment. System performance was evaluated on a leakage-controlled, held-out test split containing 1,540 images drawn from the Paddy Doctor dataset, which comprises 10,407 images spanning nine distinct rice diseases alongside healthy leaf samples. On this benchmark, the underlying ConvNeXt V2-Tiny classifier achieved an overall diagnostic accuracy of 97.5 percent, with a 95 percent confidence interval ranging from 96.7 to 98.3 percent, alongside a macro-F1 score of 96.9 percent.
Beyond raw diagnostic scores, the investigators evaluated explanation fidelity and confidence reliability across multiple standardized criteria. Six distinct feature attribution methods were tested under identical experimental configurations using model-based faithfulness metrics. When assessing decision certainty, the temperature-calibrated confidence score successfully separated correct classifications from erroneous predictions with an area under the receiver operating characteristic curve of 0.93. Exploiting this calibrated score, an automated triage rule that routes the 10 percent least-confident diagnoses to human agronomic experts successfully eliminated 74 percent of all classification errors.
The integrated architecture demonstrated substantial resilience when exposed to environmental and photographic noise. Evaluated across eight synthetic image corruptions tested at five severity tiers, the model sustained an average diagnostic accuracy of 89.7 percent. In contrast, six baseline architectures trained under identical conditions achieved mean accuracies between 72 and 78 percent under the same degradation conditions. Furthermore, the automated expert referral mechanism behaved consistently under stress, progressively increasing its referral frequency as image quality deteriorated.
Despite these laboratory strengths, evaluations revealed clear boundary conditions during cross-dataset domain shifts. When tested in zero-shot settings on two independent rice disease datasets collected from Bangladesh and India, diagnostic accuracy roughly halved, and calibrated model confidence ceased to reliably flag erroneous predictions. However, incorporating fifty local images per disease class effectively adapted the network, recovering diagnostic accuracy above 90 percent. To support reproduction and further research, the authors released their code, data splits, model weights, and evaluation scripts.
Subject of Research: Interpretable deep learning for rice crop disease diagnosis
Article Title: ExplainableCropAI: Interpretable Deep Learning for Agricultural Disease Diagnosis and Decision Support
Article References: Kumar, S., Nandan, D., & Kumar, S. (2026). ExplainableCropAI: Interpretable Deep Learning for Agricultural Disease Diagnosis and Decision Support. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01654-9
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01654-9
Keywords: ExplainableCropAI, paddy disease diagnosis, ConvNeXt V2-Tiny, Grad-CAM++, Monte Carlo Dropout, temperature-scaled calibration, Phi-3-mini, grounded language models, Paddy Doctor dataset, rice leaf diseases, model interpretability, visual attribution, uncertainty estimation, AUROC, expert referral, cross-dataset generalization, image corruption robustness
Cite Scienmag News
Alan Morgan. (November 19, 2026). Interpretable Deep Learning Diagnoses Paddy Crop Diseases and Flags Uncertain Cases. Scienmag. https://scienmag.com/interpretable-deep-learning-diagnoses-paddy-crop-diseases-and-flags-uncertain-cases/
Alan Morgan. "Interpretable Deep Learning Diagnoses Paddy Crop Diseases and Flags Uncertain Cases." Scienmag, 19 November 2026, https://scienmag.com/interpretable-deep-learning-diagnoses-paddy-crop-diseases-and-flags-uncertain-cases/. Accessed 19 November 2026.
Alan Morgan. "Interpretable Deep Learning Diagnoses Paddy Crop Diseases and Flags Uncertain Cases." Scienmag. November 19, 2026. https://scienmag.com/interpretable-deep-learning-diagnoses-paddy-crop-diseases-and-flags-uncertain-cases/








