Feature extraction techniques for ground-based cloud type classification

The appearance of each cloud type can tell the different weather conditions. Clouds may tell the coming of storms, hails, or even lightning strikes. Therefore, cloud type classification can help to reduce preventable losses. This paper studies the classification of cloud types using ground-based images. Seven sky conditions are considered, namely, cirrus, cirro and altocumulus, stratocumulus, cumulus, cumulonimbus, stratus, and clear sky image. We present an algorithm that computes a matrix of feature vectors for cloud classification with five alternative ways of extracting cloud features. The five feature extraction techniques include textures, moments of two-dimensional functions, abs-FFT, log-FFT, and the new technique called Fast Fourier Transform Projection on the x-axis (k-FFTPX). We propose the k-FFTPX algorithm that extracts features by projecting the values of logarithmic magnitude of FFT images on the x-axis of the frequency domain before selecting k sampling values of the data as k dimensions of a feature vector. To the best of our knowledge, there is no research on ground-based cloud type classification using such technique before. Then, a comparison of the techniques is made through a series of five experiments and the accuracies are ranged between 80.76% and 90.40%. Our new method provides the highest accuracy. The advantages are that we can now classify more cloud types than the existing methods with further improved in accuracy, and our method requires no expensive tools, only a digital camera is used to obtain ground-based images. This suggests a variety of practical solutions in combination with other meteorological sensors to report weather conditions inexpensively.