Weather or not—Global climate databases: Reliable on tropical mountains?

Further up, however, the agreement between measured and modelled data becomes very poor. C shows at least a moderate correlation, but the mean absolute error is with 702 mm very high. The correlation of C is even negative, showing an opposite trend of precipitation amount with altitude compared to the measured data (r = - 0.54) and an absolute error of 859 mm. This poor performance is also reflected in the amount and elevation of the maximum val- ues: the maximum measured MAP was 3299 mm at 1920 m a.s.l., the maximum modelled MAP of WC was 1893 mm at 1400 m a.s.l. and of C 1508 mm at 2770 m a.s.l. On the northern slope, the correlation was better; however, the maximum of measured MAP was 1582 mm at 2010 m a.s.l., while the maximum of modelled MAP of WC was 1313 mm at 3200 m a.s.l. and was 1242 mm at 2360 m a.s.l. in case of C (Table 1). Discussion Both databases perform quite well in low-lying areas where a comparatively dense network of weather stations already existed. However, within the ecologically most important vegetation zones (in terms of biodiversity and ecosystem services, especially rainwater input), perfor- mance is poor and drastic discrepancies exist. Not only is the maximum MAP severely under- estimated (e.g. on the southern slope of Kilimanjaro by about 50% by both models), but the rainfall trends were also completely unrealistic with deviations in the elevation of the rainfall maximum by up to 850 elevation metres. Comparing WC and C, C performs slightly better in

most cases. But also the deviations of C are partly drastic, e.g. the mean absolute error of 700 mm along the higher part of the southern slope of Kilimanjaro and the difference between measured and modelled maximum MAP of 1791 mm. Obviously, there is a slight trend towards greater deviations with increasing height (and thus area) of the mountains studied, which could be related to interpolation errors due to the larger area without station data. Lower correlation with observed precipitation over complex (tropical) mountainous regions was already found by [7] for CW and by [19] for satellite-based precipitation estimates in two areas in Tanzania. [2] also found the lowest performance of downscaled grid-based rainfall data in the tropics, attributing this to the poor quality of the station data. Although Table 2. Metrics for a comparison between WorldClim (WC) and CHELSA (C) data and the measured data of mean annual precipitation (MAP) based on 125 ground stations on Kilimanjaro. Lower southern slope Higher southern slope Northern slope mae WC (mm) 125 855 243 pbias WC (%) 12 46 26 mae C (mm) 202 693 210 pbias C (%) 25 35 26 Pearson’s r WC 0.94*** -0.56*** 0.64* Pearson’s r C 0.81*** 0.62*** 0.85** max. MAP measured (mm) - 3.299 1582 max. MAP WC (mm) - 1.893 1313 max. MAP C (mm) - 1.508 1242 Elevation max. MAP (m a.s.l.) - 1920 2010 Elevation max. MAP WC (m a.s.l.) - 1400 3200 Elevation max. MAP C (m a.s.l.) - 2770 2360

mae = mean absolute error, pbias = percentage bias, r = Pearson correlation coefficient, *** p<0.001, ** p<0.01, *p<0.05. As the maximum MAP only occurs in the higher mountain areas, no value is given for the lower southern slope. https://doi.org/10.1371/journal.pone.0299363.t002 PLOS ONE Global climate datasets on tropical mountains PLOS ONE | https://doi.org/10.1371/journal.pone.0299363 March 13, 2024 8 / 12