**4. Discussion**

Because of the isthmian geographical environment of Costa Rica (with only a Pacific-to-Caribbean coast distance of approx. 400 km), the good correlation between the GPS PWV estimates at AACR and LIBE was an expected result of our analysis, as both sites are located on the Pacific slope and share similar climatological features (e.g., analogous precipitation patterns, Figure 4A). Moreover, we also confirm the good agreemen<sup>t</sup> between the GPS PWV calculations and the radiosonde-based measurements reported by others [6,9,26,27,41]. For example, Spearman's correlation coe fficients for the GPS PWV and the radiosonde-based calculations were 0.913 (*p* < 0.001) and 0.902 (*p* < 0.001) during the dry and wet season, respectively. With respect to the MODIS satellite estimations of PWV, our analysis yielded significant biases depending on the season of the year, which are related to the annual cycle of water vapor, the NE trade winds influence, and the ITCZ activity over the Central American Isthmus. For instance, only the MODIS Terra PWV estimations recorded during the dry season were not significant biased with respect the GPS PWV calculations. However, we also found good correlations between the MODIS Aqua and the GPS-based calculations, with Spearman's correlation coe fficients of 0.735 (*p* < 0.001) and 0.621 (*p* < 0.001) for the dry and wet season, respectively. The dry and wet season MODIS Terra Spearman's correlation coe fficients were 0.591 (*p* < 0.001) and 0.368 (*p* < 0.001), respectively. These correlation values were similar to those reported over di fferent regions of the Iberian Peninsula, including island environments like Mallorca and several coastal sites [9]. Such relationships also allowed a further adjustment of the data to fit the observations by adopting a spatial bias (error) correction method like the one applied to precipitation data [21,42,43]. As mentioned above, due to the location of Costa Rica on the narrow land-bridge of Central America, the MODIS near-infrared water vapor retrieval algorithm could be greatly a ffected and the derived column water vapor values over coastal or water areas may vary significantly due to the lower signal-to-noise ratios of the measured spectra [13]. This e ffect on the MODIS retrieval algorithm was particularly evident in the MODIS Aqua PWV estimation which was somewhat high during both the dry and wet season. MODIS Terra also showed deviations but were related to an underestimation during the wet season which can be related to the so-called shielding e ffect (i.e., clouds are probably occulting water vapor underneath them) [13,44]. The di fferences between the MODIS Aqua and MODIS Terra estimations could be attributed to their di fferent passing times over the Central American Isthmus (MODIS Aqua crosses the equator in the afternoon, whereas MODIS Terra does it in the morning) and to the use of di fferent radiations to estimate the water vapor during the day and night. The MODIS Aqua estimations could be higher than the corresponding MODIS Terra values because the algorithm uses IR radiation during nighttime, which could be a ffected by the presence of clouds with water vapor, leading to overestimations. Overall, our HYSPLIT air mass trajectory analysis is consistent with the prevailing regional moisture transport mechanism during the dry season, the Caribbean Low Level Jet (CLLJ). During the wet season, in turn, there is an intensification in the genesis and development of deep convection systems on the Pacific coast of Costa Rica which is generally is associated with the presence of the "Chorro del Occidente Colombiano" or CHOCO jet [45]. These circulation patterns produced the two rainfall maxima observed on the Pacific slope, one in June and one in September, which were interrupted by a relative minimum between July–August, known as the Midsummer Drought, due to the intensification of trade winds over the Caribbean Sea [46]. The radiosonde data were also useful to validate the atmospheric conditions controlling the GPS PWV estimations. First, the composite temperature profiles calculated using the radiosonde data are in agreemen<sup>t</sup> with the previously reported structure of the upper troposphere and lower stratosphere over Costa Rica [47]. As shown in Figure 2A, the temperatures in both the dry and wet season are roughly the same at 25 km, but below this level (e.g., 15–20 km), the boreal winter (December to April) temperature profile is colder than in boreal summer (from May to October). This finding was previously attributed to the influence of wave-induced vertical motions across strong vertical gradients, the source variability in the air masses arriving at Costa Rica (e.g., tropical western Pacific or midlatitudes) resulting from horizontal transport and changes induced along parcel paths due to physical and/or chemical processes [47–49]. Secondly, despite these di fferences in the thermal structure of the tropical atmosphere of Costa Rica, the T m calculations using the Bevis equation showed small di fferences with respect to the corresponding calculations based on radiosonde data. This finding also agrees with the calculations made in Algeria and Argentina where Namaoui et al. and Fernández et al. estimated the uncertainty of the T m values and found that variations up to 15K produced small di fferences in the final estimation of GPS PWV, which did not exceed 2 mm [27,40]. Thirdly, the poor correlation observed between Ts and T m at MROC deserves further discussion. It is generally considered that the most accurate method to obtain T m is by using both temperature and humidity profiles from radiosonde data [42,50]. Therefore, we have confidence that our T m estimations are good approximations of the temperature profiles over the Central Valley of Costa Rica. A possible explanation for this finding is the mountainous and isthmian characteristics of the Costa Rica territory. The atmospheric sounding site is located on the southwestern area of the Central Valley. From this site, the distance to the Pacific coast is only 55–60 km. Radiosondes typically head in that direction after they are launched. Therefore, it seems that the atmospheric profiles estimated from MROC are representative not only of atmospheric conditions over the Central Valley but also of the Pacific coast of Costa Rica. There is also a limitation regarding the time of day when the sounding is performed. At MROC, atmospheric sounding is only done once a day, typically at 12Z or 7:00 a.m. Central American time. Therefore, T m estimations with respect to Ts represent only the atmospheric conditions prevailing during the morning when constant surface temperatures are observed (approx. 294K ± 1K; Figure 2B). In consonance with these results, it was decided to rely on the Bevis equation to estimate the hourly T m values for the Pacific slope of Costa Rica as this model has been extensively applied to estimate weighted atmospheric temperatures in several regions.

The MLR model estimated for GPS PWV data at AACR clearly matched the seasonal changes correctly, simulating smaller GPS PWV values during the dry season (from December to April) and much greater values during the wet season (from May to November). The best-performing and most parsimonious model included, as expected, near-surface (T and RH, Equation 7) and vertical atmospheric predictor variables (FLUX and P, Equation 6). The GPS PWV values were positively correlated with air temperature (T) and relative humidity (RH), with Spearman's correlation coe fficients of 0.210 and 0.426 (*p* < 0.001), respectively, and were negatively correlated with solar radiation (FLUX) and air pressure (P), with Spearman's correlation coe fficients of −0.360 and −0.175 (*p* < 0.001), respectively. These correlation results can be considered physically meaningful and can explain the overall model performance, although it is worth mentioning that, like the MODIS satellite estimations, it su ffers from seasonal biases, specially during the dry season when the small PWV measured by the ground GPS receivers were not reproduced. This worse performance of the model during the dry season compared to the wet season was also evident after biases and RMSE values were additionally estimated using the sounding PWV measurements.
