Scholarlink

<p>Crop recommendation is a key challenge in precision agriculture, aiming to match soil and environmental parameters with the most suitable crop variety. In this work, I present a lightweight exponential-weighted ensemble framework that combines multiple base classifiers—K-Nearest Neighbor, Decision Tree, Random Forest, Support Vector Machine, Naive Bayes, Logistic Regression, Extreme Gradient Boosting—through a simple yet effective reweighting scheme applied offline. This method assigns exponential penalties based on each model’s validation error and then aggregates their probability outputs in the test phase. Unlike more complex stacking or multi-layer ensembles, the proposed approach does not require a separate meta-learner, enabling straightforward implementation and reduced computational overhead. The system is evaluated on a well-known dataset available on Kaggle, achieving a final accuracy of 99.8%, which meets or exceeds the performance of more elaborate ensembles. Although the current study focuses on an offline learning setting, the exponential-weighting mechanism naturally extends to online or incremental learning scenarios, wherein model weights can be easily updated as new data arrive. This adaptability, coupled with excellent predictive accuracy, makes the approach highly suitable for real-world smart farming applications where conditions and data streams are subject to continuous change.</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

A Lightweight Exponential-Weighted Ensemble for Crop Recommendation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月28日

查看全文
  • Luisa Stracqualursi

摘要

Crop recommendation is a key challenge in precision agriculture, aiming to match soil and environmental parameters with the most suitable crop variety. In this work, I present a lightweight exponential-weighted ensemble framework that combines multiple base classifiers—K-Nearest Neighbor, Decision Tree, Random Forest, Support Vector Machine, Naive Bayes, Logistic Regression, Extreme Gradient Boosting—through a simple yet effective reweighting scheme applied offline. This method assigns exponential penalties based on each model’s validation error and then aggregates their probability outputs in the test phase. Unlike more complex stacking or multi-layer ensembles, the proposed approach does not require a separate meta-learner, enabling straightforward implementation and reduced computational overhead. The system is evaluated on a well-known dataset available on Kaggle, achieving a final accuracy of 99.8%, which meets or exceeds the performance of more elaborate ensembles. Although the current study focuses on an offline learning setting, the exponential-weighting mechanism naturally extends to online or incremental learning scenarios, wherein model weights can be easily updated as new data arrive. This adaptability, coupled with excellent predictive accuracy, makes the approach highly suitable for real-world smart farming applications where conditions and data streams are subject to continuous change.