This paper investigates the problem of accuracy of localization with GNSS in constraint environments. The ultimate goal is to provide a first confidence index on the accuracy of the position given by the GNSS. In this paper, we propose to use the complementarity between the GNSS signals and the development in image processing to count satellites with direct reception state. It consists to use a vehicle equipped with a GPS-RTK and a camera oriented upwards to capture images and count after repositioning, the satellites with direct signals (resp. with blocked/reflected signals) i.e. located in the sky region of the image (resp. located in the not-sky region). The proposed approach is based on an optimal clustering applied on simplified images. More preciously, the acquired image is simplified using a geodesic reconstruction with an optimal contrast parameter. Then, a clustering step is made in order to classify the regions into two classes (sky and not-sky). For that, a set of unsupervised (KMlocal, Fuzzy C-means, Fisher and Statistical region Merging) and supervised (Bayes, K-Nearest Neighbor and Support Vector Machine) clustering algorithms are compared in order to define the best classifier in terms of good classification rate and processing time. Experimental results are shown for hundred images taken in different conditions of acquisition (illumination changes, clouds, sun, tunnels, etc).