Integrating Satellite Imagery and Infield Sensors for Daily Spatial Plant Evapotranspiration Prediction: A Machine Learning-Driven Approach

Abstract

Continuous monitoring of crop health, especially plant EvapoTranspiration (ET) helps make efficient irrigation decisions. While various infield and remote sensing technologies provide valuable data, challenges lie in obtaining fine-resolution spatiotemporal observations. Moreover, existing research primarily addresses gap-filling past data, limiting the potential for advanced irrigation practice optimization tethered to historical observation-based management. Our work enables preemptive irrigation management by predicting the spatial ET on a day-ahead basis. We implement a Machine Learning (ML)-driven method, named Data Fusion Using Satellite and Infield Observations for Neural-network-based prediction of spatiotemporal plant EvapoTranspiration (DFUSIONET). This methodology in-cludes an Earth Observation (EO) based spatial ET-generation model, a unique interpolation method named Proportional-offset Interpolation (POI) for temporal gap-filling, and a multi-input Feedforward Neural Network (FNN) for predicting daily spatial ET. Results demonstrate that the POI method outperforms other interpolation techniques with an RMSE of 0.11 mm d-1, and the FNN exhibits a RMSE range of [0.13, 0.31] mm d-1.DFUSIONET integrates satellite imagery, infield sensor data, and meteoro-logical parameters and captures spatiotemporal variations of site-and crop-specific parameters to predict plant ET. This forecasting strategy is beneficial for efficient preemptive irrigation management, contributing to sustainable agricultural practices and improved resource utilization in precision agriculture.