Multi-model ensemble mapping of irrigated areas using remote sensing, machine learning, and ground truth data

Abstract

Reliable information about the extent of irrigated areas is critical for water resources management in agricultural regions facing mounting water scarcity challenges due to climate variability and change, extreme events, and increased competition over limited water resources. To this end, we introduce a multi-model ensemble mapping (MEM) approach to develop high-fidelity, high-resolution (30 m) annual maps of irrigated areas from 2007 to 2022 in the Upper Red River Basin (URRB), U.S., using remote sensing, machine learning (ML), and ground truth data. Our approach combines the outputs of different ML classifiers, including Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Gradient Tree Boost, and Classification and Regression Trees (CART) in Google Earth Engine. ML classifiers were trained using different input variables including vegetation indices acquired from high-resolution (30 m) Landsat imagery, soil, topography, and climate data. Furthermore, we developed a rich ground truth dataset of 910 irrigated fields in 2022 to enhance the predictive performance of ML classifiers and assess model accuracy. While CART and SVM classifiers outperformed other models with higher ground truth accuracies of 71 % and 73 %, respectively, the MEM approach improved the ground truth accuracy to ∼84 %. Results indicate a notable upstream expansion of irrigation in the URRB, particularly near tributaries, where new croplands were frequently irrigated even during droughts when downstream irrigation was halted due to diminished surface water availability. The combination of the expansion of upstream irrigated areas and consistency of irrigation has critical long-term implications for downstream agricultural water availability.

Keywords

• Irrigation expansion
• High-fidelity mapping
• Machine learning classifiers
• Vegetation indices
• Agricultural water availability
• Water scarcity