Two accessions of the new species were collected by the second and last author in May 2009. The material was collected in northern Cameroon (Adamawa region). Descriptive terminology follows Grey-Wilson (1980) except for simple symmetrical plane shapes (SACDBT 1962).

In order to study the convergent evolution of I. smetsiana and I. erecticornis R.Wilczek & G.M.Schulze, we analysed the overall morphology of those two species with the closest allies of I. smetsiana using a Multiple Factor Analysis for mixed data (continuous and discrete) as implemented in the FactoMineR R package (Lê et al. 2008). The scores obtained for each individual in the first two dimensions were used to interpret the degree of phenotypical convergence between the new species and its morphological doppelgänger, I. erecticornis. Morphological traits used in the analysis are provided in Table 1.

In order to determine the phylogenetic position of I. smetsiana among the African lineages, we added sequences of the new species to a combined nuclear ImpDEF1/ImpDEF2 and nuclear ribosomal ITS data matrix. Based on earlier phylogenetic studies (Volkmar et al. 2014), I. rubrostriata Hook.f. and I. conchibracteata Y.L.Chen & Y.Q.Lu serve as outgroup species. Total genomic DNA was isolated from silica-dried leaf material using a modified CTAB protocol (Doyle and Doyle 1987), which is optimized for Impatiens by Janssens et al. (2006, 2009b). Primers and temperature profiles used for the amplification of ImpDEF1/ImpDEF2, and ITS follow Janssens et al. (2007, 2008) and White et al. (1990), respectively. Amplification reactions were carried out on a GeneAmp PCR system 9700 (Applied Biosystems). Purified amplification products were sent to Macrogen, Inc. (Seoul, South Korea) for sequencing. Sequences obtained in this study were submitted to GenBank (Supplementary file 1).

Contiguous sequences were assembled using Geneious v.7.0.6 (Biomatters, New Zealand). Automatic alignments were carried out with MAFFT (Katoh et al. 2002) under an E-INS-i algorithm, a 100PAM/k=2 scoring matrix, a gap open penalty of 1.3 and an offset value of 0.123. Subsequent manual fine-tuning of the aligned dataset was done in Geneious v.7.0.6. Congruency between the different datasets was inferred using different methods. First, a series of incongruence length difference tests (ILD; Farris et al. 1995) were carried out with PAUP* v.4.0b10 (Swofford 2003) using the following parameters: simple taxon addition, TBR branch swapping, and heuristic searches of 1000 repartitions of the data. Despite the well-known sensitivity of the ILD test (Barker and Lutzoni 2002), the results of this test were compared in light of the resolution and support values of the obtained nuclear and nuclear ribosomal topologies. As a result, possible conflict between data matrices was visually inspected, searching for conflicting relationships within each topology that are strongly supported (hard vs soft incongruence; Johnson and Soltis 1998).

The best-fit nucleotide substitution model for each plastid and nuclear dataset was selected using jModelTest v.2.1.4. (Posada 2008) out of 88 possible models under the Akaike information criterion (AIC). The GTR+I+G model was found as best fit for ImpDEF1, whereas the GTR+G model was calculated as best substitution model for ImpDEF2 and ITS. As a result, we used a mixed-model approach to apply different evolutionary models on each DNA region of the combined dataset (Ronquist and Huelsenbeck 2003). Bayesian inference analyses were conducted with MrBayes v.3.1 (Huelsenbeck and Ronquist 2001) on four individual data partitions and a combined data matrix. Each analysis was run two times for 20 million generations. Trees were sampled every 5,000^th generation. Chain convergence and ESS parameters were inspected with TRACER v.1.4 (Rambaut and Drummond 2007). Only nodes with Bayesian posterior probabilities (BPP) above 0.95 are considered as well supported (Suzuki et al. 2002). Maximum Likelihood analyses were conducted using the RAxML search algorithm (Stamatakis et al. 2005) as implemented in RAxML v.7.2.8 (Stamatakis et al. 2008) under the GTRGAMMA approximation of rate heterogeneity for each gene (Stamatakis 2006). Five hundred bootstrap trees were inferred using the RAxML Rapid bootstrap algorithm (ML-BS) to provide support values for the best-scoring ML tree.

The molecular clock hypothesis was tested using a chi² likelihood ratio test (Felsenstein 1988) and demonstrated that the substitution rates in the combined dataset are not clock-like (p < 0.001 for all markers). BEAUti v.1.8.0 (Drummond and Rambaut 2007) was used to prepare the xml file prior to the dating analysis in BEAST v.1.8.0 (Drummond and Rambaut 2007). The latter software program applies the Bayesian methodology to compute divergence times. However, in order to surpass the zero likelihood issue in BEAST, we used a starting tree that was obtained by carrying out a ML analysis in RAxML v.7.2.8 (Stamatakis et al. 2008) under GTRGAMMA model with the rooted likelihood tree as input tree for a penalized likelihood (PL) analysis in the software program r8s (Sanderson 2003). Due to differing substitution models among the different gene markers used, a partitioned Bayesian MCMC analysis was performed under the Yule speciation model and a relaxed lognormal clock. Partitions were unlinked for the model of evolution. The analysis ran for 30 million generations and was sampled each 5,000^th generation. Convergence of the chains and ESS parameter evaluation (ESS > 200) was performed with TRACER v.1.6 (Rambaut and Drummond 2007). A maximum clade credibility tree using a posterior probability limit of 0.5 was calculated using TreeAnnotator v.1.8.0. (Drummond and Rambaut 2007). The previously computed age estimate for the most recent common ancestor of the Afro-Madagascan and Equatorial West African clades of Janssens et al. (2009a) was used to calibrate the ingroup of the current study. This calibration point was given a normal distribution with a mean value of 7.87 Ma and a standard deviation of 0.5.

The first two components of the multivariate analysis explained 69% of the total variation in continuous traits. Dimension 1 correlated with upper lateral petal shape, leaf apex, flower colour, bract shape, and lateral sepal size. Dimension 2 with dorsal petal shape, lower lateral petal shape, inflorescence, and to a lesser extent leaf margin and petiole length. The multiple factor analysis for mixed data demonstrates that Impatiens smetsiana (Fig. 1A–B) and its morphological doppelgänger I. erecticornis (Fig. 1D) cluster well together (Fig. 2). This clustering is mainly the result of overall similarities in floral morphology between both species and to a lesser extent because of similarities in vegetative morphology. In addition to the morphological resemblance of I. smetsiana and I. erecticornis, also I. kamerunensis subsp. obanensis and I. kamerunensis subsp. kamerunensis aggregate well in the cluster analysis (Fig. 2).

No significant incongruence between ITS and ImpDEF1/ImpDEF2 (p > 0.05) was found using the partition homogeneity test. Visual examination of the two different partitions of the combined dataset corroborates this congruency analysis. ML and BI analysis of the combined dataset generated topologies with moderate to high support values (Fig. 3).

Impatiens smetsiana is part of a small clade consisting only of Equatorial West African species (Fig. 3). Phylogenetic analyses indicate that I. smetsiana is sister to a clade containing I. oumina, I. nzabiana, I. kamerunensis subsp. kamerunensis, and I. kamerunensis subsp. obanensis (ML: 65, BI: 0.91). Impatiens filicornu Hook.f. is sister to this entire clade (ML: 94, BI: 1.00). The clade formed by the novelty, and the above-mentioned Impatiens species [((((I. nzabiana, I. oumina) (I. kamerunensis subsp. kamerunensis, I. kamerunensis subsp. obanensis)) I. smetsiana) I. filicornu)] is sister to the extant species of the Equatorial West African Impatiens clade, except I. sakeriana, which is the most early divergent species of this Equatorial West African lineage (Janssens et al. 2009a).

Within twenty million generations, stationarity among the different chains as well as ESS values over 200 were obtained. The BEAST maximum clade credibility tree analysis is depicted in Fig. 3. Divergence time for Equatorial West African lineage corroborates the study of Janssens et al. (2009a) and is estimated at 3.67 Ma (4.99–2.74 Ma (95% highest priority density (HPD))) (Fig. 3). Additionally, time divergence estimates of the Afro-Madagascan clade to which I. erecticornis belongs are in accordance with earlier dating analyses of Janssens et al. (2009a), having a mean age of 4.43 Ma (5.39–3.46 Ma (95% HPD)). The split between I. erecticornis and its sister species (I. bequaertii) is situated in the Pleistocene (2.32 Ma (3.25–1.24 Ma (95% HPD)), suggesting a Pleistocene origin. The stem node age of I. smetsiana is situated in the a Pliocene at 2.61 Ma (3.64–1.73 Ma (95% HPD)).

Examination of the general morphological characters of I. smetsiana shows that the novelty is morphologically very similar to I. erecticornis in general appearance: spirally arranged leaves with acuminate leaf apex and attenuate leaf base, subumbellate racemes, ovate bracts and lateral sepals, a curved filiform spur, oval-kidney shaped upper lateral petals and asymmetrically obovate lower lateral petals and equally sized lateral petals. Impatiens smetsiana differs from I. erecticornis in leaf shape (elliptic vs ovate), the absence of fimbriae at the base of the leaf margins, the number of secondary veins (5–7 vs 7–10), the number of flowers per inflorescence (> 6 vs 3–6), and the dorsal petal lacking a distinct crest. Despite the morphological resemblance, molecular analyses indicate that I. erecticornis is part of a distinct clade of Equatorial East African species that are only distantly related to the Equatorial West African lineage to which I. smetsiana belongs. The most recent common ancestor of both species is located in Southwest China around 7.87 Ma (Janssens et al. 2009a). To date, I. erecticornis is only found in eastern DR Congo and in western Rwanda (Albertine Rift) (Fischer 1997).

Furthermore, molecular phylogenetic analyses demonstrated the close affiliation of I. smetsiana with the morphologically more distinct species I. filicornu, I. kamerunensis subsp. kamerunensis, I. kamerunensis subsp. obanensis, I. nzabiana, and I. oumina. Although there are still some common morphological characteristics between the novelty and the first four species mentioned here, there is only limited morphological resemblance with I. oumina and I. nzabiana except for a simple or sparsely branched stem and spirally arranged leaves (Janssens et al. 2011). When comparing I. smetsiana with the two I. kamerunensis subspecies, it is clear that I. kamerunensis subsp. obanensis is morphologically more similar to the new species than I. kamerunensis subsp. kamerunensis. Only the morphological character spur length of the new species falls in between that of the two subspecies of I. kamerunensis. In comparison to I. kamerunensis subsp. kamerunensis, the new species is significantly larger in overall size (on average between 50 and 60 cm in height), has spirally arranged leaves instead of an opposite-subopposite leaf arrangement, a larger lower sepal (up to 9 mm long) and larger lateral petals (19–21 mm), four characters it shares with I. kamerunensis subsp. obanensis. In addition, I. smetsiana differs from I. kamerunensis s.l. in having a subumbellate raceme instead of a lax axillary or pseudo-terminal raceme, and a distinct shape of the united lateral petals. Although molecular analyses strongly support a sister group relationship of I. smetsiana with I. nzabiana, I. oumina, and I. kamerunensis s.l., the novelty has also several morphological characters in common with the more early branched off I. filicornu. Impatiens filicornu is similar to I. smetsiana by the presence of a simple or sparsely branched stem with spirally arranged leaves, subumbellate racemes, ovate bracts and lateral sepals, and pink flowers with a whitish filiform spur. However, it can be easily distinguished from the new species by the presence of fimbriae at the leaf base, the smaller size of both plants and flowers, and the narrowly oblong-elliptic upper lateral petals and suborbicular-oblong lower lateral petals.

Unravelling higher level taxonomic relationships in Impatiens based on morphological data has proven difficult due to the huge floral variation present in the genus (Hooker and Thomson 1859; Hooker 1875, 1905; Warburg and Reiche 1895). Recent molecular phylogenetic analyses identified large incongruence between molecular data and morphological characters in which similar flower types are randomly distributed throughout the phylogeny (Yuan et al. 2004; Janssens et al. 2006; Yu et al. 2015). The study of Janssens et al. (2006) demonstrated that the main Impatiens lineages are biogeographically delimited to one of the major hotspots for the genus. Interestingly, in each of these biogeographically delimited clades in Africa, Madagascar, South India, and Southeast Asia, there is a trend of convergent flower evolution with the independent development of the same major flower types: (1) flat flower type with a shallow lower sepal and long filiform spur, (2) funnel flower type with cucullate dorsal petal, deeply navicular lower sepal and short filiform spur, (3) funnel flower type with cucullate dorsal petal, large saccate lower sepal and bucciniform spur. In general, homoplasy of similar floral characters amongst biogeographically delimited lineages within Impatiens is fairly obvious to recognize, as the development of each of these flower types in each delineated clade is rather distinctive. The present finding is remarkable as it implies a rather pronounced case of convergent evolution in which two distantly related taxa (their MRCA was probably a funnel shaped flower with a saccate spurred sepal; Janssens et al. 2008) independently converged into practically the same flower morphology. It is generally accepted that the convergent evolution of complex phenotypic traits such as the complete turnover of the flower morphology is the result of stochastic processes (Wood et al. 2005). Moreover, the majority of cases in which morphological homoplasy is clearly present appears to be the result of natural selection. As for I. smetsiana and I. erecticornis, the convergent evolution of their flowers is undoubtedly closely correlated with an adaptation to a similar pollination syndrome. The typical floral morphology of the two convergent Impatiens species investigated is exclusively psychophilous (butterfly-pollinated). The shape of the flower causes the pollinator to position itself on the often large lower lateral petals, directly given access to the long nectar spur through a narrow opening between the two pairs of lateral petals underneath the stamens. With its long proboscis, the butterfly can easily access the nectar, which is situated near the tip of the spur. Morphological homoplasy is often suggested to be the outcome of unique genetic changes. In such an extreme case of convergent evolution in which complex floral morphological traits evolved several times, one could assume that only a few major quantitative trait loci (QTLs) are involved. This would cause the number of pathways from phenotype ‘A’ to phenotype ‘B’ to be significantly limited and thus increases the possibility of complex morphological homoplasy to occur under a common genetic basis. However, because the dynamics on how genes regulate the eventual flower morphology in Impatiens is still unclear, it is difficult to say whether convergent evolution of complex morphological traits, as is the case for I. smetsiana and I. erecticornis, have a common genetic background.

Molecular data demonstrates that I. smetsiana is only very distantly related to its morphological doppelgänger I. erecticornis endemic to the Albertine Rift, but is in fact closely related to the Impatiens species of the Equatorial West African clade containing I. kamerunensis subsp. kamerunensis, I. kamerunensis subsp. obanensis, I. nzabiana, and I. oumina. Because I. nzabiana and I. oumina are strict endemics of the Massif du Chaillu (South Gabon) and their divergence time is estimated between 0.61 and 0.017 Ma, both species had most likely no influence on the divergence of I. smetsiana which happened approximately two million years earlier. As a result, we further refer to the Equatorial West African species I. kamerunensis subsp. kamerunensis and I. kamerunensis subsp. obanensis when discussing the biogeography and evolution of I. smetsiana. In contrast to I. smetsiana, which is putatively endemic to the Tchabal Mbabo Mountains in Central North Cameroon (Adamawa Region), I. kamerunensis subsp. kamerunensis and I. kamerunensis subsp. obanensis are more widely distributed and occur in Southwest Cameroon (Northwest, Southwest, West, and Littoral Region), southern Nigeria, Ghana, and Togo. Interestingly, the known distribution area of I. smetsiana does not overlap with that of I. kamerunensis subsp. kamerunensis and I. kamerunensis subsp. obanensis. The montane forests of Tchabal Mbabo to which I. smetsiana is endemic, are part of the Afromontane regional centre of endemism. According to Thomas and Thomas (1996), the Tchabal Mbabo forests are of a rare, dry type, which is completely surrounded by deciduous woodland or shrubland. Impatiens smetsiana was found in a small creek on the north escarpment edge of these montane forests, close to the gallery forests that dissect the plateau grasslands, with a population of I. mackeyana subsp. zenkeri thriving downstream of the novelty. Most likely, the steep slopes of the Tchabal Mbabo montane escarpment forest created an orographic effect resulting in a suitable microclimate for Impatiens populations to grow in that area (Chapman 2004).