Root-zone Soil Moisture Nowcasting using Context Aware Machine Learning

Estimation of root-zone soil moisture (SM) is crucial for effective agricultural management and water resource planning. However, current methods for soil moisture estimation exhibit several limitations that hinder their practical application. This study introduces a novel nowcasting model, which integrates in-situ and remote sensing data through a Predictive Error Compensated wavelet Neural NETwork (PECNET), addressing the drawbacks of existing approaches.Existing SM estimation techniques often suffer from limited accuracy, inadequate contextual information, and no real-time monitoring capabilities. Although remote sensing technologies offer promising advantages, such as wide spatial coverage and frequent data acquisition, they are not immune to limitations. Vegetation coverage and density present challenges in accurately estimating root-zone SM using remote sensing techniques. These factors can introduce uncertainties and errors in the estimation process, thereby impacting the reliability of the results.To overcome these limitations and enhance the accuracy of root-zone SM estimation, this study proposes the integration of remote sensing data with in-situ measurements. Specifically, Normalized Difference Vegetation Index (NDVI) calculations from Landsat 7 and Landsat 8 satellites are fused with evapotranspiration and rainfall data obtained from agrometeorological stations. Combining these datasets generates an 8-day time series for the target parcels, leveraging the contextual information provided by NDVI and seasonality to improve the accuracy of root-zone soil moisture estimation.To develop a robust and efficient model, we introduce PECNET, which ensures the orthogonality of input features and facilitates the learning of non-linear relationships between variables. Notably, PECNET addresses the challenge of limited labeled training data, minimizing the risk of overfitting and enabling accurate estimation with fewer labeled samples. In addition, this study employs discrete wavelet transformation coefficients as inputs for the neural networks, demonstrating superior performance compared to direct measurements.Validation experiments were conducted to evaluate the performance of the proposed PECNET model. Comparative analyses with simple regression, Kriging, and feed-forward neural networks reveal the significant advantages of the Predictive Error Compensated wavelet Neural NETwork approach in root-zone soil moisture estimation.In conclusion, this research introduces an advanced nowcasting model for root-zone SM estimation. By integrating in-situ and remote sensing data, harnessing contextual information, and utilizing state-of-the-art machine learning techniques, our approach overcomes the limitations of existing methods. It offers a robust solution for accurate soil moisture estimation, with implications for precision agriculture, water management, and decision-making related to agricultural practices and land resource utilization. However, the limitations imposed by vegetation coverage and density on remote sensing technologies should be considered when interpreting the results.