Scholarlink

<p>Apple farming in India's Himalayan region is increasingly vulnerable to disease outbreaks driven by climate variability. Traditional image-based detection methods often fall short in providing early warnings under these dynamic environmental conditions. This study aims to develop a scalable and interpretable predictive model for apple disease outbreaks using environmental data, with the goal of supporting early intervention and sustainable farming practices. We propose an ensemble learning framework that combines independently trained Long Short-Term Memory (LSTM) networks with attention, Transformer models, and Temporal Fusion Transformer (TFT) architectures. The individual models capture complementary temporal patterns in environmental time-series data, and their predictions are aggregated using an averaging-based ensemble strategy to improve robustness and generalization. The model utilizes localized meteorological variables (temperature and humidity) to predict disease risks without reliance on imagery. Performance was assessed through accuracy, AUROC, ablation studies, and cross-regional transferability tests. The proposed model achieved 84% accuracy and an AUROC of 0.95, surpassing conventional baseline models. Ablation studies confirmed the contribution of each model component, while comparative tests showed high adaptability and interpretability across diverse geographies. These outcomes highlight the model’s robustness and suitability for real-world deployment. This work advances climate-resilient agriculture by enabling responsible AI-driven early warning systems that reduce chemical dependency and support sustainable orchard management. The approach aligns with the Sustainable Development Goals (SDGs) by enhancing farmer empowerment, minimizing environmental impact, and promoting transformation of agri-food systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Responsible scaling of deep learning for sustainable apple disease prediction: an ensemble learning approach using LSTM, transformer, and temporal fusion transformer

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Ishana Attri,
  • Lalit Kumar Awasthi,
  • Teek Parval Sharma

摘要

Apple farming in India's Himalayan region is increasingly vulnerable to disease outbreaks driven by climate variability. Traditional image-based detection methods often fall short in providing early warnings under these dynamic environmental conditions. This study aims to develop a scalable and interpretable predictive model for apple disease outbreaks using environmental data, with the goal of supporting early intervention and sustainable farming practices. We propose an ensemble learning framework that combines independently trained Long Short-Term Memory (LSTM) networks with attention, Transformer models, and Temporal Fusion Transformer (TFT) architectures. The individual models capture complementary temporal patterns in environmental time-series data, and their predictions are aggregated using an averaging-based ensemble strategy to improve robustness and generalization. The model utilizes localized meteorological variables (temperature and humidity) to predict disease risks without reliance on imagery. Performance was assessed through accuracy, AUROC, ablation studies, and cross-regional transferability tests. The proposed model achieved 84% accuracy and an AUROC of 0.95, surpassing conventional baseline models. Ablation studies confirmed the contribution of each model component, while comparative tests showed high adaptability and interpretability across diverse geographies. These outcomes highlight the model’s robustness and suitability for real-world deployment. This work advances climate-resilient agriculture by enabling responsible AI-driven early warning systems that reduce chemical dependency and support sustainable orchard management. The approach aligns with the Sustainable Development Goals (SDGs) by enhancing farmer empowerment, minimizing environmental impact, and promoting transformation of agri-food systems.