Generating high-precision farmland irrigation pattern maps using remotely sensed ecological indices and machine learning algorithms

Abstract

Conducting field investigations of farmland irrigation patterns on a large scale is a time-consuming and labor-intensive task. The traditional approach of employing satellite remote sensing for large-scale visual assessments is impractical for identifying irrigation patterns due to interference caused by vegetation cover. To address this, we utilized the Google Earth Engine (GEE) platform, integrating environmental covariates and machine learning algorithms, to generate distribution maps of irrigation patterns (micro-irrigation and surface irrigation) at a 30-meter resolution for the Turpan-Hami Basin. Results demonstrate that the Classification and Regression Tree (CART) model achieved a classification accuracy of 0.81, effectively distinguishing between different irrigation patterns. The analysis of feature importance determined NDVI (i.e., Normalized Difference Vegetation Index), EVI (i.e., Enhanced Vegetation Index), MSI (i.e., Moisture Stress Index), Ec (i.e., Transpiration), and NDWI (i.e., Normalized Difference Water Index) as the key indicators linked to irrigation patterns. Regional mapping findings reveal an increase in the proportion of micro-irrigation from 40.2 % in 2015 to 47.0 % in 2023, underscoring the successful implementation of water-saving practices in the Turpan-Hami Basin. Additionally, we developed a GEE-based interactive interface, which enables users to generate corresponding distribution maps of irrigation patterns by selecting a specific year, offering uesful data support for policymakers and farmers to better manage agricultural water resources.