Introducing spatial information in k-means algorithm for clouds detection in optical satellite images

Due to restricted visibility time of remote sensing polar platforms from earth reception station, only a limited number of images can be transmitted. In the case of optical images, an in-board cloud cover detection module will allow to transmit only useful (i.e. weakly cloudy) images. In order to derive such a module, we propose a method to detect cloudy areas from subsampled images. For a pixel ground surface of about 100 × 100 m², cloudy areas appear as the highest radiometric value homogeneous areas. The algorithm presented in this paper is based on the k-means method. Its main originality is to improve classical results by introducing isotropic spatial information. Input data are the sorted components of a vector composed of radiometric values for each pixel and its neighbours (4-connexity). Then a classical k-means method with constraints on the cloudy class gravity center is used on these vectors. We tested the method on a set of 206 subsampled SPOT XS and 138 SPOT P images and their manmade interpretation masks. To evaluate the quality of our results, we used the probability of false alarm (PFA) depending on the number of pixels which have been wrongly declared cloudy, and the probability of non detection (PND) depending on the number of pixels which have been wrongly declared non cloudy. We obtained rather good PFA(< 1%) and PND(< 30%), and compared these values with results obtained with other methods.