Modeling the distribution of headwater streams using topoclimatic indices, remote sensing and machine learning.

Abstract

Headwater streams (HWS) are ecologically important components of montane ecosystems. However, they are difficult to map and may not be accurately represented in existing spatial datasets. We used topographically resolved climatic water balance data and satellite indices retrieved from Google Earth Engine to model the occurrence (presence or absence) of HWS across Northwest Montana. A multi-scale feature selection (MSFS) procedure and boosted regression tree models/machine learning algorithms were used to identify variables associated with HWS occurrence. In final model evaluation, models that included climatic water balance deficit were more accurate (83.5% ranging from 82.9% to 83.7%) than using only terrain indices (81.1% ranging from 80.7% to 81.4%) and improved upon estimates of stream extent represented by the National Hydrography Dataset Plus High Resolution (NHDPlus HR) (82.7% ranging from 82.5% to 83.1%). Including topoclimate captured the varying effect of upslope accumulated area across a strong moisture gradient. Multi-scale cross-validation, coupled with a MSFS algorithm allowed us to find a parsimonious model that was not immediately evident using standard cross-validation procedures. More accurate spatial model predictions of HWS have potential for immediate application in land and water resource management, where significant field time can be spent identifying potential stream impacts prior to contracting and planning.