The study investigates the potential of a commercially available proximal sensing system – based on a 16-band multispectral sensor – for monitoring mean midday gross ecosystem production (GEP_m) in a subalpine grassland of the Italian Alps equipped with an eddy covariance flux tower. Reflectance observations were collected for 5 consecutive years, characterized by different climatic conditions, together with turbulent carbon dioxide fluxes and their meteorological drivers. Different models based on linear regression (vegetation indices approach) and on multiple regression (reflectance approach) were tested to estimateGEP_m from optical data. The overall performance of this relatively low-cost system was positive. Chlorophyll-related indices including the red-edge part of the spectrum in their formulation (red-edge normalized difference vegetation index, NDVI_red-edge; chlorophyll index, CI_red-edge) were the best predictors of GEP_m, explaining most of its variability during the observation period. The use of the reflectance approach did not lead to considerably improved results in estimating GEP_m: the adjusted R² (adjR²) of the model based on linear regression – including all the 5 years – was 0.74, while the adjR² for the multiple regression model was 0.79. Incorporating mean midday photosynthetically active radiation (PAR_m) into the model resulted in a general decrease in the accuracy of estimates, highlighting the complexity of the GEP_m response to incident radiation. In fact, significantly higher photosynthesis rates were observed under diffuse as regards direct radiation conditions. The models which were observed to perform best were then used to test the potential of optical data for GEP_m gap filling. Artificial gaps of three different lengths (1, 3 and 5 observation days) were introduced in the GEP_m time series. The values of adjR² for the three gap-filling scenarios showed that the accuracy of the gap filling slightly decreased with gap length. However, on average, the GEP_m gaps were filled with an accuracy of 73% with the model fed with NDVI_red-edge, and of 76% with the model using reflectance at 681, 720 and 781 nm and PAR_m data.