Smartphone-based soil moisture mapping using low-cost sensors and AI inference

Thesis Abstract

The study addresses the challenge of accurate, scalable soil moisture monitoring in agricultural and environmental systems using ubiquitous smartphones and low-cost sensors, which traditional methods often fail to provide due to limited spatial coverage and high equipment costs. The aim is to develop and validate a smartphone-based soil moisture mapping system that integrates low-cost capacitive moisture sensors, smartphone camera data, and AI inference to produce high-resolution field-scale moisture maps. Specific objectives include (1) to design a modular sensing apparatus leveraging hook-and-loop sensor modules and a standardized sampling protocol across diverse soil textures; (2) to develop machine learning models that infer volumetric water content (VWC) from sensor readings, smartphone imagery, and contextual environmental data; (3) to implement a mobile application and cloud-based analytics pipeline for real-time map generation with uncertainty quantification; (4) to evaluate model transferability across soil types and moisture regimes using cross-site validation; and (5) to assess the agronomic and environmental utility of the moisture maps for irrigation scheduling and drought monitoring. The methodology employs a mixed-methods research design combining quantitative predictive modelling with qualitative usability assessments. The population includes farmers and agronomists in three contrasting climatic regions. A stratified random sample of 180 fields (60 per region) is selected, with 15–20 sampling points per field to capture spatial variability. Data collection instruments comprise (i) a low-cost capacitive soil moisture sensor array (0–100 bar), calibrated against gravimetric soil moisture; (ii) a smartphone-based imaging module for colorimetric cues and texture features; (iii) a multi-parameter environmental sensor suite measuring ambient temperature, relative humidity, and solar irradiance; and (iv) a structured survey and semi-structured interviews to evaluate user experience and decision impact. The data analysis employs regression- and ensemble-based modelling (random forest, gradient boosting, and deep learning regression) to predict VWC from sensor features, with cross-validation and nested cross-site validation to assess generalization. Feature engineering includes spectral and textural features from images, sensor-derived time-series statistics, and soil texture indicators. Model selection is guided by Akaike Information Criterion and out-of-sample RMSE comparison. Uncertainty is quantified using prediction intervals and Monte Carlo dropout for neural models. The study also applies a theoretical lens grounded in the Technology Acceptance Model (TAM) and Diffusion of Innovation theory to interpret adoption and usage patterns. A conceptual framework is developed to integrate data fusion, mobile sensing, and decision-support outputs. Expected findings indicate that integrating smartphone imagery with low-cost sensor data improves VWC prediction accuracy by 15–25% over sensor-alone baselines, with RMSE reductions from ~5.0% to ~3.5% volumetric content across sites. Model performance is anticipated to be robust to moderate soil texture variation but will require site-specific calibration for extreme textures. The AI inference framework is expected to deliver spatially explicit moisture maps with uncertainty bounds useful for site-specific irrigation decisions, achieving near real-time (<30 minutes) map generation post-collection. Qualitative results will reveal favourable usability metrics among farmers, with key facilitators including straightforward calibration workflows, offline capabilities, and clear visualization of irrigation recommendations. The contribution to knowledge lies in demonstrating a scalable, low-cost, ICT-enabled approach to field-scale soil moisture mapping that integrates heterogeneous data streams, validates transferability across agro-ecological zones, and provides actionable decision-support outputs for precision agriculture and water resource management. The study concludes that smartphone-based mapping, when paired with AI inference and user-centered design, can substantially expand access to high-resolution soil moisture data. Recommendations include expanding sensor interoperability, incorporating additional soil health variables, and developing open-access datasets to facilitate broader replication and policy integration.