||Remote sensing is a key technology that enables us to scale up our empirical, in situ measurements of carbon uptake made at the site level. In low leaf area index ecosystems typical of
semi-arid regions however, many assumptions of these remote sensing approaches fall short, given the complexities of the heterogeneous landscape and frequent disturbance. Here, we investigated the utility of remote sensing data for predicting gross primary production (GPP) in piñon-juniper woodlands in New Mexico (USA). We developed a simple model hierarchy using climate drivers and satellite vegetation indices (VIs) to predict GPP, which we validated against in situ estimates of
GPP from eddy-covariance. We tested the influence of pixel size on model fit by comparing model performance when using VIs from RapidEye (5 m) and the VIs from Landsat ETM+ (30 m). We also tested the ability of the normalized difference wetness index (NDWI) and normalized difference red edge (NDRE) to improve model fits. The best predictor of GPP at the undisturbed PJ woodland was Landsat ETM+ derived NDVI (normalized difference vegetation index), whereas at the disturbed site, the red-edge VI performed best (R2adj of 0.92 and 0.90 respectively). The RapidEye data did improve model performance, but only after we controlled for the variability in sensor view angle, which had a significant impact on the apparent cover of vegetation in our low fractional cover experimental woodland. At both sites, model performance was best either during non-stressful growth conditions, where NDVI performed best, or during severe ecosystem stress conditions (e.g., during the girdling process), where NDRE and NDWI improved model fit, suggesting the inclusion of red-edge leveraging and moisture sensitive VI in simple, data driven models can constrain GPP estimate uncertainty during periods of high ecosystem stress or disturbance.