Advancing early warning capabilities with CHIRPS-compatible NCEP GEFS precipitation forecasts

Adjusting GEFS forecasts to local climatology

What amount of correction is required for GEFS forecasts to align with CHIRPS local climatology? The amount of correction varies widely across the globe and throughout the year. Figure 1a shows annual mean bias for GEFS reforecast 15-day totals. In this figure, wetter-than-CHIRPS climatology and systematic over-prediction of 15-day totals by GEFS is indicated by positive mean bias values, while the opposite is indicated by negative values. GEFS forecast mean bias was calculated for each month and then averaged across rainy season months, to focus aggregate results on the rainfall seasons, when precipitation forecasts are relevant. Monthly dry masks excluded locations with a monthly average of less than 10 mm, according to CHIRPS climatology. In general, one consistent result from Fig. 1a is a tendency to increase precipitation in many mountainous tropical and subtropical regions. By design, orographic precipitation enhancements in such regions are represented fairly well in CHIRPS, and these are carried through to CHIRPS-GEFS precipitation forecasts. The CHIRPS-GEFS bias-correction process reduces systematic errors (Fig. 1b), with the overall mean absolute bias error going from 24.1 mm for GEFS to 19.7 mm for CHIRPS-GEFS, an ~18% reduction.

Figure 1a through Fig. 5 are based on GEFS reforecast, CHIRPS, and CHIRPS-GEFS data for the 5-day or 15-day periods beginning on the 1^st, 6^th, 11^th, and 16^th day of the month. All these exclude dry season months. Figure 1b shows the corresponding global distribution of annual average error for the GEFS reforecast and CHIRPS-GEFS, and is discussed later.

GEFS has a large annual average positive bias of higher-than 40 mm in some areas of the globe, including in central Mexico, Central America, northern South America, the Andes and Himalayan Mountain ranges, and in southern China, Papua New Guinea, and localized areas of central Africa, the Ethiopian Highlands, and the western montane United States (Fig. 1a). GEFS has positive bias, by more than 5 mm for the annual average 15-day period, across the northern United States including in the Midwest, from Mexico’s northern mountains through most of Central America, in northern South America, the Andes range, eastern Brazil, in parts of central Europe, central and northern Asia, in the area from southern China to Myanmar and Thailand, and in northeastern and western India. GEFS has positive bias in portions of East Africa (Rwanda, Burundi, Tanzania, western Ethiopia), West Africa (Cameroon, Gabon), and Southern Africa (Zambia, central Angola, northern Zimbabwe, eastern South Africa). GEFS has negative bias, by more than 5 mm on average, in parts of central and northern Africa, Senegal, northern Australia, central South America, western India, the Yucatan peninsula, and the United States Gulf Coast.

GEFS’ systematic bias changes throughout the year, as shown by the monthly mean bias in January, April, July, and October (Fig. 2). This is unsurprising, given that drivers of weather change too, but higher bias in particular months can be problematic for forecast users. In Ethiopia, for example, GEFS overestimates by large amounts during the Kirempt season (e.g., in July) and in October in the southwest. In central Brazil, the bias changes markedly by season, from a high negative bias in October to an expansive wet bias in April. In the Midwestern and northern United States, GEFS also shows a more expansive wet bias in April than in January, July, or October. In some areas, like in southern China and the Andes mountains, GEFS means are higher than CHIRPS means throughout the year.

The CHIRPS-GEFS downscaling procedure corrects for systematic errors in GEFS forecasts that vary spatially and temporally. To assess the efficacy of the CHIRPS-GEFS approach, we began by calculating the per-pixel difference between GEFS and CHIRPS, and CHIRPS-GEFS and CHIRPS for 15-day periods. These were calculated for each month, for in-season pixels, and then averaged across the year. We then looked at the histogram of the resulting differences (Fig. 1b), to identify the distribution of annual average errors in these two products. CHIRPS-GEFS errors are shown as gray bars and GEFS errors are overlaid as hollow red bars. A desirable pattern is more small errors (higher bars close to 0 mm) and fewer large magnitude errors (lower bars at larger precipitation values). As shown in Fig. 1b, the bias-correction procedure has this effect, and results in CHIRPS-GEFS having overall lower errors for global rainy seasons compared to GEFS. GEFS 15-day errors more commonly involve over prediction of observed amounts than under prediction, as shown by the higher proportion of positive versus negative moderate to large positive errors. Part of this is due to the lower limit of under prediction being zero precipitation, while over prediction can range from marginal precipitation amounts to very high amounts. As shown in Fig. 1b, the CHIRPS-GEFS bias correction particularly reduces GEFS forecast errors for moderate-to-high rainfall amounts, and it results in a global 15-day error distribution that has a higher proportion of small errors, e.g., errors within −10 mm to 10 mm of CHIRPS values (51% for CHIRPS-GEFS and 43% for GEFS). Errors in categories ranging from 10 mm to 40 mm occur less often in CHIRPS-GEFS, globally, with probabilities in those categories reduced by around 15 and 25 percent at 10 mm to 20 mm and 20 mm to 30 mm, respectively, and by around 30 percent to 40 percent for errors that are higher than 40 mm.

Next, we show performance of the 5-day and 15-day CHIRPS-GEFS precipitation forecasts by correlations and mean absolute errors for the historical record, compared to CHIRPS data for these periods. As described in Data Records, multiple outlets use forecast amounts for these periods. In the Usage Notes section, probability of detection scores for 15-day CHIRPS-GEFS in Africa are presented while describing an operational application of the CHIRPS-GEFS for seasonal monitoring. In that discussion we also examine the performance of 5-day forecasts during the 2020–2021 season in key regions of Kenya, Angola, Zambia, Zimbabwe, and Madagascar.

Pearson correlation coefficients for 5-day and 15-day CHIRPS-GEFS, compared to CHIRPS (Fig. 3), indicate the ability of forecasts to predict deviations from average. It should be noted that correlations are nearly entirely driven by the information coming from the GEFS forecasts. The conversion to CHIRPS-GEFS adjusts the GEFS values to make them more “CHIRPS-like,” while also approximating the historical context of the GEFS forecast. Wet extremes forecasted by GEFS translate into wet extremes in CHIRPS-GEFS. Areas with very low correlations (R < 0.3) are where there should be low confidence in the forecasts, such as in parts of Central and West Africa and the Amazon region of South America. In many other areas, correlations for January, April, July, and October indicate moderate to good skill in forecast 5-day precipitation totals. In many of these areas, 15-day forecasts have lower but still identifiable skill. Some of the regions with primarily moderate (R 0.5 to 0.7) and high correlations (R > 0.7) are the United States, Western Europe, and Eastern Europe, southeastern South America, southern Central Asia, eastern China, parts of East and Southern Africa, and Australia. Globally, correlations are higher in January, April, and October than in July, which indicates generally higher forecast accuracy in those months. Exceptions are in eastern China, southern Brazil, eastern Mexico, northeastern Ethiopia, and central and southern Australia, where July correlations are not substantially lower. 15-day forecasts also have high correlations in some areas, including in the Western and Midwestern United States in January, in central and northern Australia in April, and in eastern Brazil in January and October.

In Africa, a region where CHIRPS data is actively used by the Famine Early Warning System Network (FEWS NET) and other organizations for seasonal monitoring and drought early warning, forecast correlations indicate moderate to good 5-day and 15-day forecast performance in areas of East Africa, Southern Africa, and western North Africa during rainy season months. Some of the highest 15-day correlations in Africa are during important rainy season months, for example, in northeastern Ethiopia in July and April, in Kenya in April, in Zimbabwe and southern Mozambique in January, and in the Sudanian zone of West Africa in October. Very low correlations indicate low forecast skill in the Sahel, coastal West Africa, and in Central Africa in the DRC, Republic of the Congo, and Gabon.

Mean absolute error of the bias-corrected GEFS forecasts highlight the areas where forecast amounts have historically been less reliable (Fig. 4). These indicate non-systematic errors associated with rains not materializing in the forecast location in the forecast period, which can be from GEFS model deficiencies and the inherent challenges of weather forecasting. Extreme precipitation events and warm season, deep moist convection-driven precipitation are notorious challenges for numerical weather prediction systems^48,49, and CHIRPS-GEFS data are not immune to this problem. Remotely sensed data, including CHIRPS, also struggle with estimating extreme high rainfall amounts^13,50, though since we are comparing CHIRPS-GEFS to CHIRPS, the main source of the large errors shown here would be the GEFS reforecast.

As shown in Fig. 4, the magnitude of errors follows climatology, with 5-day errors typically under 10 mm for drier rainy season months. In wetter months and locations errors are typically between 10 mm and 20 mm. With higher rainfall magnitude there is greater potential for larger errors. The 15-day forecast errors exhibit a similar spatial pattern to the 5-day errors, and error magnitudes correspond to the three-times larger accumulation interval as well as expected lower skill at longer lead time. Figure 4 shows especially large 15-day mean absolute errors in January near northern Mozambique and Madagascar, in July and October in parts of Central America, in April in central Kenya and southwestern Tanzania, in July in India’s Western Ghats Mountains and in the Himalayas, and in the Maritime Continent. In southeast China, while the 15-day correlations indicated decent skill at forecasting the sign of precipitation anomalies, large 15-day errors indicate the influence of poorly forecast large storms, which unbiasing cannot correct for. In the Amazon rainforest, many areas with low correlations also have high forecast errors, underscoring poor forecast performance there.

Source: Resources - nature.com