Scholarlink

<p>The present study aimed to standardize and validate machine learning (ML) models for predicting cumulative growing degree day (GDD) thresholds, corresponding to the appearance and disappearance of apple phenological stages. Three ML algorithms viz; Random Forest Regression, XGBoost, and Long Short-Term Memory (LSTM), were employed to forecast the phenological stages of three apple cultivars, ‘Gala Redlum’, ‘Golden Clone&#xa0;B’, and ‘Granny Smith’, at two locations: Experimental Field of Sher-e-Kashmir University of Agricultural Sciences and Technology of Kashmir (SKUAST-K), Shalimar (L1), and Ambri Apple Research Centre (AARC), Pahnoo, Shopian (L2). The base models were trained using 6&#xa0;years of weather and phenological data (2014, 2016, 2018, 2019, 2022, and 2023) for ‘Gala Redlum’ at Shalimar, with 2024 data used as the test set. Model optimization was performed through GridSearchCV, and the refined models were subsequently applied to predict phenological stages using 2&#xa0;years of data (2023 and 2024) for all three cultivars at both sites. For fine-tuning, 2023 served as the training year and 2024 as the testing year. Among the tested algorithms, XGBoost consistently outperformed other models for ‘Gala Redlum’ and ‘Granny Smith’, whereas Random Forest Regression produced the best results for ‘Golden Clone&#xa0;B’, as reflected by lower mean absolute error (MAE) and root mean squared error (RMSE) values and higher coefficient of determination (R<sup>2</sup>) scores. Shapley additive explanations (SHAP) analysis revealed that accumulated GDD exerted the strongest influence on phenological stage prediction compared to humidity and rainfall. The models were further validated using 2025 field data, confirming their reliability and suitability for real-time application in orchard management and decision-support systems. These predictions can assist in optimizing the timing of critical cultural operations such as thinning, spraying, and irrigation, thereby improving fruit quality and marketability.</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

Machine Learning Framework for Phenological Stage Prediction of Apple: A Case Study of the Kashmir Division

  • 期刊论文

  • 发表日期：

    2025年11月25日

查看全文
  • Biaza Sidiqi,
  • Sankalp Gupta,
  • Mahroosh,
  • Piyush Kumar,
  • Ashiq H. Pandit,
  • Mohammad Khalid Pandit

摘要

The present study aimed to standardize and validate machine learning (ML) models for predicting cumulative growing degree day (GDD) thresholds, corresponding to the appearance and disappearance of apple phenological stages. Three ML algorithms viz; Random Forest Regression, XGBoost, and Long Short-Term Memory (LSTM), were employed to forecast the phenological stages of three apple cultivars, ‘Gala Redlum’, ‘Golden Clone B’, and ‘Granny Smith’, at two locations: Experimental Field of Sher-e-Kashmir University of Agricultural Sciences and Technology of Kashmir (SKUAST-K), Shalimar (L1), and Ambri Apple Research Centre (AARC), Pahnoo, Shopian (L2). The base models were trained using 6 years of weather and phenological data (2014, 2016, 2018, 2019, 2022, and 2023) for ‘Gala Redlum’ at Shalimar, with 2024 data used as the test set. Model optimization was performed through GridSearchCV, and the refined models were subsequently applied to predict phenological stages using 2 years of data (2023 and 2024) for all three cultivars at both sites. For fine-tuning, 2023 served as the training year and 2024 as the testing year. Among the tested algorithms, XGBoost consistently outperformed other models for ‘Gala Redlum’ and ‘Granny Smith’, whereas Random Forest Regression produced the best results for ‘Golden Clone B’, as reflected by lower mean absolute error (MAE) and root mean squared error (RMSE) values and higher coefficient of determination (R2) scores. Shapley additive explanations (SHAP) analysis revealed that accumulated GDD exerted the strongest influence on phenological stage prediction compared to humidity and rainfall. The models were further validated using 2025 field data, confirming their reliability and suitability for real-time application in orchard management and decision-support systems. These predictions can assist in optimizing the timing of critical cultural operations such as thinning, spraying, and irrigation, thereby improving fruit quality and marketability.