Incorporating hydrology into climate suitability models changes projections of malaria transmission in Africa

Continental-scale models of malaria climate suitability typically couple well-established temperature-response models with basic estimates of vector habitat availability using rainfall as a proxy. Here we show that across continental Africa, the estimated geographic range of climatic suitability for malaria transmission is more sensitive to the precipitation threshold than the thermal response curve applied. To address this problem we use downscaled daily climate predictions from seven GCMs to run a continental-scale hydrological model for a process-based representation of mosquito breeding habitat availability. A more complex pattern of malaria suitability emerges as water is routed through drainage networks and river corridors serve as year-round transmission foci. The estimated hydro-climatically suitable area for stable malaria transmission is smaller than previous models suggest and shows only a very small increase in state-of-the-art future climate scenarios. However, bigger geographical shifts are observed than with most rainfall threshold models and the pattern of that shift is very different when using a hydrological model to estimate surface water availability for vector breeding.

observed in the 3-month categories of season length. The 1-3 month and 9-12 month categories were most sensitive within the 50% range.
Maps of hydro-climatic suitability estimates for selected runoff thresholds covering the range explored within the sensitivity analysis are presented in Supplementary Figure 15. The northern extent of transmission suitability does not change much within a wide range of thresholds; however, most noticeable is the change in year-round malaria suitability. At lower runoff thresholds, the LIS-MAL model identifies a much larger area of year-round suitability, extending to the fringes of transmission. The extent is more sensitive to runoff threshold in southern and eastern Africa; however, within a 50% range (represented by the 20, 30 and 40 m 3 s -1 thresholds in Supplementary Figure 15) there is only a small degree of variability.
Within central Africa, lower thresholds predict more widespread year-round malaria suitability, which may be more appropriate for these regions (see Supplementary Note 2).
For comparison, a similar sensitivity analysis was conducted for the rainfall threshold approach to climatic modelling of malaria transmission suitability. The 60 mm monthly rainfall threshold was taken as the base value in this case. The rainfall threshold was initially adjusted at 5 mm increments between 5 mm and 120 mm. Although this range matches the percentage variability investigated for the runoff threshold, it did not incorporate the range of rainfall thresholds reported in the literature (Supplementary Table 1). Thus, the upper range was extended to match the upper threshold in Supplementary Table 1. Figure 16 indicates greater sensitivity at the extreme rainfall threshold values and for the monthly categories, the overall pattern is similar to that in Supplementary Figure 14. Decreasing the rainfall threshold by 50% increased the suitable area by 9%, while raising the threshold by 50% decreased the suitable area by 8%. Estimates of year-round malaria are sensitive to the rainfall threshold, even within the 50% range. Considering the full range of rainfall thresholds reported in the literature (reflecting the multiple and spatially variable hydrological processes aggregated into this threshold), there is considerable sensitivity to the rainfall threshold. It should be noted that the choice of the 60 mm rainfall threshold as a base value here is somewhat arbitrary and conservative; selecting a higher value from Supplementary Table   1 would yield more sensitive suitability estimates. For a comparison of the effect of threshold variability with Supplementary Figure 15, see Figure 1 in the main text.

While Supplementary
Supplementary Note 2.

Supplementary Note 2. Model validation against observations
Estimations of hydro-climatic suitability for malaria transmission are challenging to validate quantitatively for a number of reasons. First, the goal is to identify an envelope within which malaria transmission may occur; yet, numerous interventions have been undertaken to reduce the distribution of transmission over the last century 1 . For this reason, we focus our validation efforts on the pre-intervention map of Lysenko and Semashko 2 (~1900, reproduced in Supplementary Figure 17a), but recognise that substantial hydroclimatic changes will have taken place in the last 120 years and the exact locations of the boundaries in the pre-intervention map are likely subject to considerable uncertainty. Second, hydro-climatic suitability for malaria transmission does not necessarily mean that transmission will occur in those areas. There are numerous non-climatic controls on malaria transmission 3 4 were each converted into a binary layer of presence/absence at the grid cell length of the hydro-climatic suitability model estimates. The mean of the seven hydro-climatic estimates for the historical period  was also converted into a binary layer using a minimum estimate of 0.5 months transmission as a threshold. Five validation metrics were calculated from the resulting confusion matrix: (i) model accuracy (i.e. the proportion of all prediction that are correct); (ii) precision (i.e. the proportion of positive identifications that are correct); (iii) recall (i.e. the proportion of actual positives that are identified correctly); (iv) F1 score (the harmonic mean of precision and recall); (v) the false positive rate (i.e. the proportion of all actual negatives that are incorrectly identified as positive).
Results of the validation are provided in Supplementary Tables 10 and 11. All models performed similarly in terms of quantitative validation performance, especially given the uncertainty in validation sets. At the continental scale considered here this is not surprising given the strong link climatic controls on malaria transmission -there are many ways to model this relationship. In terms of both model accuracy and recall, the lower rainfall thresholds (Kiszewski 7 , Ermert 8 , Martens 9 , Tanser 10 and Craig 11 ) are best-performing, owing to the large areas estimated to be suitable in these models. By the same token, these same models report lower precision values and higher false positive rates. The converse is true regarding the higher rainfall thresholds (Garnham 12 ,Parham & Michael 13 ) and the rainfall to potential evapotranspiration ratio of Lindsay 14  It is also important to examine the performance of models at the edges of transmission suitability, as these are precisely the areas that are most sensitive to changes in transmission potential due to climate change and where rivers may extend hydro-climatic suitability into the semi-arid regions that surround the core tropical malaria transmission area. We compared LIS-MAL and LIS-MAL + IRRA with an indicative rainfallthermal model calculated from the methods of Tanser Figure 9 shows that the area is well within suitable temperature ranges), the south being suitable for 3-6 months of the year and the middle just 1-3 months. Sénégal has a strong latitudinal pattern of malaria incidence in the country 15 and generally low and seasonal transmission, which follows the rainfall-threshold pattern. However, malaria transmission is also observed on the northern border beyond this threshold, driven by water availability in the Sénégal River basin with some vectors observed year-round 16 . Dia et al. 17 also noted year-round Anopheles larval breeding on the northern border with parasitic prevalence in schoolchildren indicative of a hypo to mesoendemic situation in the low lying river valley. Clearly, the spatial distribution of malaria transmission is more complex than national scale maps aggregated into health districts would suggest, thereby highlighting the difficulty of using such data sets to inform validation without further investigation. For instance, Diouf et al. 18 22 specifically reported the effect of the Yame River in supporting transmission during periods of low discharge when rainfall was reduced.

Sahel Region: Nigeria
Models are generally in agreement with the finding that the whole of Nigeria is malarious 23 , though it seems that a focus in the centre-north of the country is not seen in hydro-climatic suitability estimates.
While Tolulope 24 reported variable patterns in malaria morbidity each year, this northern focus was consistent. A further focus in the west of the country is captured in each hydro-climatic model.

Sahel Region: Niger
Again, models agree on the suitability for malaria transmission at the southern border with Nigeria, which matches observations. Labbo et al. 25 note the near year-round presence of vectors at two villages close to the Niger River which is indicated by LIS-MAL but not rainfall threshold models. The more seasonal transmission reported 26 at the central southern border was also recreated well by both approaches.
Sahel Region: The Nile LIS-MAL predicts much more extensive hydro-climatic suitability for malaria transmission in Sudan than rainfall-based models. Malik and Khalafalla 27 note the presence of malaria throughout Sudan, including hypo-endemic areas on the far north attributed to association with rivers. Ageep et al. 28 confirm the yearround presence of anopheles larvae close to the Nile. In Gezira state year-round malaria cases were observed 29 (2007-2016), with an increase between July and October. Further north, malaria has been wellknown in Egypt since ancient times 30 , with proven vectors present as far north as the Nile delta 31 . Shousha 32 reported year-round occurrence of invading An. gambiae in 1944 prior to a successful eradication campaign in 1945. Thus, while there are currently no local cases of malaria, it is certainly hydro-climatically suitable for malaria and areas close to water bodies are thought to be of high risk of future malaria outbreaks 33 .
Democratic Republic of the Congo (DRC) All hydro-climatic representations of malaria transmission indicate that the Democratic Republic of the Congo (DRC) is situated within the most widespread focus of suitability (Supplementary Figure 19). While quantitative validation suggest that all models perform equally well here, it is nonetheless worthwhile to examine the differences in the model estimates.
A comparison of LIS-MAL and Tanser model estimates reveals three key differences: (1) The LIS-MAL estimates produce more complex spatial patterns; (2) The LIS-MAL estimates indicate that a much higher area of the country is suitable for malaria transmission year-round, with river corridors serving as foci for this sustained transmission suitability; (3) The LIS-MAL model estimates that substantial areas are only hydro-climatically suitable for <3 months of the year as the threshold discharge is not met in these areas Unfortunately, there is a high level of uncertainty regarding the actual spatial distribution of malaria in the DRC, with the fewest data points per land area of any country 34 . The inverse distance weighted spatially interpolated model of Messina et al. 34 based on Demographic Health Survey data highlights clusters of transmission that was observed to be more associated with social factors (e.g. wealth, community use of bed nets). Janko et al. 35 also note a positive association between agriculture and malaria risk in the DRC.
Nevertheless, the generally higher prevalence of malaria parasitaemia in the north was not observed in either model; indeed, the LIS-MAL model suggests <3 months suitable for transmission in some of these areas. Malaria transmission in the DRC is more widespread than suggested by the LIS-MAL model as seen in recent studies 36

Supplementary Figures
Supplementary Figure 1 Supplementary Figure 11. Drivers of projected changes. Hydrological and thermal drivers of projected changes in malaria hydro-climatic suitability for both the LIS-MAL approach and Tanser rainfall thresholdbased estimates.
Supplementary Figure 13. Country-by-country comparison of population-months. Comparison is presented for each period according to estimates from both LIS-MAL and the Tanser rainfall threshold (i.e. population exposed multiplied by the number of additional months of exposure). The 1:1 line is shown in red. Individual countries are highlighted when a >50% difference in predictions is observed. Supplementary