Smartphone-based image analysis and interpretable machine learning for soil moisture estimation across diverse Indian soils
Roy, D.
Smartphone-based image analysis and interpretable machine learning for soil moisture estimation across diverse Indian soils - Amsterdam (Netherlands) : Elsevier B.V., 2025.
Peer review
Reliable soil moisture estimation is crucial for agricultural water management, yet conventional methods are often invasive, costly, and impractical for frequent field-level use. This study presence a smartphone-based, non-destructive approach for estimating soil moisture content (SMC) estimation across five contrasting Indian soil groups from 14 locations. A total of 238 soil images were analyzed to extract 33 colour-based features, which were then used to train and validate ten machine learning (ML) models. The Random Forest (RF) model exhibited the highest predictive accuracy (R2 = 0.78; RMSE = 5.98 %) during validation. To improve interpretability, SHAP and ALE techniques identified Redness Index (RI), Colour Feature Index (ColFeatInd), red band (R), value (V), and X colour space as key predictors. Boruta selection confirmed the relevance of all features. This study demonstrates the potential of combining smartphone imagery and interpretable ML to scalable, low-cost SMC across diverse soil types.
Text in English
2352-9385 (Online)
https://doi.org/10.1016/j.rsase.2025.101655
Machine learning
Soil
Soil Water Content
Monitoring
Remote sensing
Smartphone-based image analysis and interpretable machine learning for soil moisture estimation across diverse Indian soils - Amsterdam (Netherlands) : Elsevier B.V., 2025.
Peer review
Reliable soil moisture estimation is crucial for agricultural water management, yet conventional methods are often invasive, costly, and impractical for frequent field-level use. This study presence a smartphone-based, non-destructive approach for estimating soil moisture content (SMC) estimation across five contrasting Indian soil groups from 14 locations. A total of 238 soil images were analyzed to extract 33 colour-based features, which were then used to train and validate ten machine learning (ML) models. The Random Forest (RF) model exhibited the highest predictive accuracy (R2 = 0.78; RMSE = 5.98 %) during validation. To improve interpretability, SHAP and ALE techniques identified Redness Index (RI), Colour Feature Index (ColFeatInd), red band (R), value (V), and X colour space as key predictors. Boruta selection confirmed the relevance of all features. This study demonstrates the potential of combining smartphone imagery and interpretable ML to scalable, low-cost SMC across diverse soil types.
Text in English
2352-9385 (Online)
https://doi.org/10.1016/j.rsase.2025.101655
Machine learning
Soil
Soil Water Content
Monitoring
Remote sensing