Scholarlink

<p>Accurate soil moisture estimation is crucial in hydrology and agriculture. Numerous studies exist, only few have integrated drought indices (DI) with machine learning (ML) for soil moisture (SM) prediction across various depths. This study evaluates optimal ML models and DI combinations for SM estimation across 37 stations in Nigeria while considering dominant soil properties. Six ML models including Artificial Neural Network (ANN), K-Nearest Neighbours, Support Vector Machines, Gradient Boosting, Generalised Linear Models, and Random Forest, were employed alongside 15 DI combinations derived from PDSI, PHDI, ScPDSI, SPEI, and RAI. ANN performed best for SM1 (0–5&#xa0;cm), with RMSE (0.04–0.13) and R² (0.6–0.92). However, Ferric Luvisols, Dystric Nitosols and Thionic Podzols prevalent in high-rainfall (&gt; 2000&#xa0;mm) waterlogged areas, reduced prediction accuracy for SM2 (0–100&#xa0;cm) and SM3 (0–bedrock). Integrating meteorological variables with ML models enhances SM prediction and supports improved irrigation and water resource management globally.</p> Graphical Abstract <p>According to the graphical summary presented, this study was conducted to identify the optimal combination of drought indices as inputs into machine learning algorithms for soil moisture prediction across various soil layers. The complex interactions between drought indices, prevalent soil types and soil moisture (SM) were analysed across different Nigerian states. Historical climate data, including precipitation, relative humidity, solar radiation, wind speed, and minimum and maximum temperatures alongside historical soil moisture data were collected from multiple stations within the study area. Evapotranspiration was estimated using the Penman–Monteith method. Daily precipitation and evapotranspiration data were used to derive several drought indices (DIs), namely the Rainfall Anomaly Index (RAI), Standardised Precipitation Evapotranspiration Index (SPEI), Palmer Drought Severity Index (PDSI), Self-Calibrating Palmer Drought Severity Index (ScPDSI), and Palmer Hydrological Drought Index (PHDI). Each drought index was used individually and in pairs, resulting in 15 different input combinations for simulating multilayer soil moisture using six machine learning (ML) algorithms: Random Forest (RF), Gradient Boosting (GB), Generalised Linear Models (GLM), Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and Artificial Neural Network (ANN). This yielded 270 distinct simulations at each climate station. Model performance was evaluated using the Pearson correlation coefficient, Root Mean Square Error (RMSE), Nash–Sutcliffe Efficiency (NSE), and other relevant metrics. The evaluation revealed that local soil characteristics significantly influenced the accuracy of SM prediction, particularly for soil depths exceeding 5&#xa0;cm and in areas with waterlogged conditions. Furthermore, ANN, RF, and KNN were identified as the most effective algorithms for predicting soil moisture at a 0–5&#xa0;cm depth in 37.8%, 29.7%, and 18.9% of the stations, respectively. The findings from this study are vital for stakeholders in agriculture, particularly those involved in developing techniques and tools for soil moisture monitoring.</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

Multi-depth Soil Moisture Modelling Based on Drought Indices and Machine Learning Algorithms

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • Blessing Funmbi SASANYA,
  • Oluwafemi Oladipupo KOLAJO

摘要

Accurate soil moisture estimation is crucial in hydrology and agriculture. Numerous studies exist, only few have integrated drought indices (DI) with machine learning (ML) for soil moisture (SM) prediction across various depths. This study evaluates optimal ML models and DI combinations for SM estimation across 37 stations in Nigeria while considering dominant soil properties. Six ML models including Artificial Neural Network (ANN), K-Nearest Neighbours, Support Vector Machines, Gradient Boosting, Generalised Linear Models, and Random Forest, were employed alongside 15 DI combinations derived from PDSI, PHDI, ScPDSI, SPEI, and RAI. ANN performed best for SM1 (0–5 cm), with RMSE (0.04–0.13) and R² (0.6–0.92). However, Ferric Luvisols, Dystric Nitosols and Thionic Podzols prevalent in high-rainfall (> 2000 mm) waterlogged areas, reduced prediction accuracy for SM2 (0–100 cm) and SM3 (0–bedrock). Integrating meteorological variables with ML models enhances SM prediction and supports improved irrigation and water resource management globally.

Graphical Abstract

According to the graphical summary presented, this study was conducted to identify the optimal combination of drought indices as inputs into machine learning algorithms for soil moisture prediction across various soil layers. The complex interactions between drought indices, prevalent soil types and soil moisture (SM) were analysed across different Nigerian states. Historical climate data, including precipitation, relative humidity, solar radiation, wind speed, and minimum and maximum temperatures alongside historical soil moisture data were collected from multiple stations within the study area. Evapotranspiration was estimated using the Penman–Monteith method. Daily precipitation and evapotranspiration data were used to derive several drought indices (DIs), namely the Rainfall Anomaly Index (RAI), Standardised Precipitation Evapotranspiration Index (SPEI), Palmer Drought Severity Index (PDSI), Self-Calibrating Palmer Drought Severity Index (ScPDSI), and Palmer Hydrological Drought Index (PHDI). Each drought index was used individually and in pairs, resulting in 15 different input combinations for simulating multilayer soil moisture using six machine learning (ML) algorithms: Random Forest (RF), Gradient Boosting (GB), Generalised Linear Models (GLM), Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and Artificial Neural Network (ANN). This yielded 270 distinct simulations at each climate station. Model performance was evaluated using the Pearson correlation coefficient, Root Mean Square Error (RMSE), Nash–Sutcliffe Efficiency (NSE), and other relevant metrics. The evaluation revealed that local soil characteristics significantly influenced the accuracy of SM prediction, particularly for soil depths exceeding 5 cm and in areas with waterlogged conditions. Furthermore, ANN, RF, and KNN were identified as the most effective algorithms for predicting soil moisture at a 0–5 cm depth in 37.8%, 29.7%, and 18.9% of the stations, respectively. The findings from this study are vital for stakeholders in agriculture, particularly those involved in developing techniques and tools for soil moisture monitoring.