Scholarlink

<p>Accurate prediction of multi-layer soil moisture within the root zone is critical for cocoa plantations, where water availability directly influences root development, nutrient uptake, flowering and yield stability. In tropical systems, strong rainfall variability, heterogeneous soils, and delayed subsurface responses make depth-resolved moisture forecasting particularly challenging. This study proposes an improved Quantile Convolutional Neural Networks (CNN)-Long Short-Term Memory (LSTM) framework for robust and interpretable multi-layer soil moisture prediction. The model integrates CNN for localized temporal feature extraction with stacked LSTM long short-term memory networks for sequential dependency modelling, while incorporating quantile regression to provide probabilistic forecasts. Autocorrelation-guided lag optimization identified lag-7 temporal window as optimal. Zone 1 data were used exclusively for training and validation, whereas Zones 2 and 3 were reserved for independent testing to ensure spatial generalization and robustness. The proposed model achieved consistently high predictive accuracy across five soil depths (5–105&#xa0;cm), with an overall average R<sup>2</sup> of 0.948, low RMSE (0.39–0.79 across layers), and MAPE generally below 3%. Independent testing in Zones 2 and 3 demonstrated minimal performance degradation, confirming strong transferability under varying field conditions. For uncertainty quantification, quantile regression produced reliable 50 and 80% prediction intervals. Low mean pinball loss values, Prediction Interval Coverage Probability (PICP) close to nominal levels, moderate Mean Prediction Interval Width (MPIW), and favorable Winkler scores indicate well-calibrated and sharp probabilistic forecasts. Explainability analysis using SHapley Additive exPlanations (SHAP) and Integrated Gradients Attribution (IGA) revealed coherent depth-dependent climatic controls, transitioning from short-term atmospheric drivers in shallow layers to slower-varying seasonal influences in deeper horizons. Overall, this research delivers a robust, uncertainty-aware, and interpretable quantile deep learning framework for multi-layer soil moisture forecasting, supporting smart irrigation and climate-adaptive water management in tropical precision agriculture.</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

Explainable Quantile CNN-LSTM model for uncertainty-aware multi-layer soil moisture prediction in tropical cocoa plantations

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月12日

查看全文
  • Sarowar Morshed Shawon,
  • Mukter Zaman,
  • Shamala Maniam,
  • Marran Al Qwaid,
  • Md Tanjil Sarker,
  • Tee Yei Kheng,
  • H. Y. Wong

摘要

Accurate prediction of multi-layer soil moisture within the root zone is critical for cocoa plantations, where water availability directly influences root development, nutrient uptake, flowering and yield stability. In tropical systems, strong rainfall variability, heterogeneous soils, and delayed subsurface responses make depth-resolved moisture forecasting particularly challenging. This study proposes an improved Quantile Convolutional Neural Networks (CNN)-Long Short-Term Memory (LSTM) framework for robust and interpretable multi-layer soil moisture prediction. The model integrates CNN for localized temporal feature extraction with stacked LSTM long short-term memory networks for sequential dependency modelling, while incorporating quantile regression to provide probabilistic forecasts. Autocorrelation-guided lag optimization identified lag-7 temporal window as optimal. Zone 1 data were used exclusively for training and validation, whereas Zones 2 and 3 were reserved for independent testing to ensure spatial generalization and robustness. The proposed model achieved consistently high predictive accuracy across five soil depths (5–105 cm), with an overall average R2 of 0.948, low RMSE (0.39–0.79 across layers), and MAPE generally below 3%. Independent testing in Zones 2 and 3 demonstrated minimal performance degradation, confirming strong transferability under varying field conditions. For uncertainty quantification, quantile regression produced reliable 50 and 80% prediction intervals. Low mean pinball loss values, Prediction Interval Coverage Probability (PICP) close to nominal levels, moderate Mean Prediction Interval Width (MPIW), and favorable Winkler scores indicate well-calibrated and sharp probabilistic forecasts. Explainability analysis using SHapley Additive exPlanations (SHAP) and Integrated Gradients Attribution (IGA) revealed coherent depth-dependent climatic controls, transitioning from short-term atmospheric drivers in shallow layers to slower-varying seasonal influences in deeper horizons. Overall, this research delivers a robust, uncertainty-aware, and interpretable quantile deep learning framework for multi-layer soil moisture forecasting, supporting smart irrigation and climate-adaptive water management in tropical precision agriculture.