After thinking about it for some days I came up with a proof sketch (which is hopefully correct) that it is possible to use Gouraud shading on disconnected triangles, if they share an edge with the same normals as in the picture. The simple idea of this proof is to check that there is no discontinuity between the intensity of the triangles by approaching from each side to the edge. So given two triangles $T_1: (P_1,P_2,P_3)$ and $T_2: (P_1', P_2', P_3')$ which share an edge and let's assume w.l.o.g. that this edges are constructed by $(P_2,P_3)$ and $(P_2', P_3')$. Because Gouraud shading only computes the light intensity at the vertex positions and interpolate in between them, it's save to assume that the intensity at $P_2$ is the same as on $P_2'$, same goes for $P_3$ and $P_3'$. Because it's also save to assume that the values inside the triangles are getting computed correctly so it's only necessary to show that there is no discontinuity between the triangles. To interpolate inside the triangle it's rational to use Barycentric coordinates so that the intensity at any point inside the triangle is given by
\begin{align*}
I &= aI_2 + bI_3 + (1 - (a+b)) I_1\\
I' &= a'I_2' + b'I_3' + (1 - (a'+b')) I_1'
\end{align*}
By now going from $I_1$ respectively $I_1'$ to the edge, a sequence can be constructed to archive this, by setting $a_n = \frac{c}{n+1}, b_n=\frac{d}{n+1}, c + d = 1$ and analogously for $a_n', b_n'$. Then the intensity is
\begin{align*}
I_n &= \frac{c}{n+1} I_2 + \frac{d}{n+1} I_3 + (1 - \frac{c+d}{n+1}) I_1\\
I_n' &= \frac{c'}{n+1}I_2' + \frac{d'}{n+1}I_3' + (1 - \frac{c'+d'}{n+1}) I_1'
\end{align*}
If one now approaches the edge with:
\begin{align*}
&\lim\limits_{n\rightarrow 0^+} I_n = cI_2 + dI_3\\
&\lim\limits_{n\rightarrow 0^+} I_n' = c'I_2 + d'I_3'
\end{align*}
and using the requirement that $I_n$ and $I_n'$ share the same position $P_n$ if they are on an edge, which can be constructed by
\begin{align*}
P_n &= \frac{c}{n+1} P_2 + \frac{d}{n+1} P_3 + (1 - \frac{c+d}{n+1}) P_1\\
P_n &= \frac{c'}{n+1}P_2' + \frac{d'}{n+1}P_3' + (1 - \frac{c'+d'}{n+1}) P_1'
\end{align*}
with the fact that $P_2 =P_2'$ and $P_3 = P_3'$ it follows:
\begin{align*}
&\lim\limits_{n\rightarrow 0^+} P_n = cP_2 + dP_3\\
&\lim\limits_{n\rightarrow 0^+} P_n' = c'P_2 + d'P_3
\end{align*}
and $d = 1 - c$
\begin{align*}
&\lim\limits_{n\rightarrow 0^+} P_n = cP_2 + (1-c)P_3\\
&\lim\limits_{n\rightarrow 0^+} P_n' = c'P_2 + (1-c')P_3
\end{align*}
this results in $c = c'$ and therefore this gives the final result:
\begin{align*}
\lim\limits_{n\rightarrow 0^+} I_n = cI_2 + (1-c)I_3 = c'I_2' + (1-c')I_3' = \lim\limits_{n\rightarrow 0^+}I_n'
\end{align*}
Thereby is possible to use Gouraud shading over disconnected triangles as in my case, because the intensity on the edge is the same on both triangles.
What actually happened in my case was that the normals differs slightly (thank you @Rahul for the tip) which therefore results in this strange look. I fixed them by averaging the normals over a small radius to ensure
the proper set up of the normals.
glShadeModelor interpolation qualifier). $\endgroup$ – user106 May 23 '17 at 2:06