Regional scale crop mapping using multi-temporal satellite imagery

One of the problems in dealing with optical images for large territories (more than 10,000 sq. km) is the presence of clouds and shadows that result in having missing values in data sets. In this paper, a new approach to classification of multi-temporal optical satellite imagery with missing data due to clouds and shadows is proposed. First, self-organizing Kohonen maps (SOMs) are used to restore missing pixel values in a time series of satellite imagery. SOMs are trained for each spectral band separately using nonmissing values. Missing values are restored through a special procedure that substitutes input sample's missing components with neuron's weight coefficients. After missing data restoration, a supervised classification is performed for multi-temporal satellite images. An ensemble of neural networks, in particular multilayer perceptrons (MLPs), is proposed. Ensembling of neural networks is done by the technique of average committee, i.e. to calculate the average class probability over classifiers and select the class with the highest average posterior probability for the given input sample. The proposed approach is applied for regional scale crop classification using multi temporal Landsat-8 images for the JECAM test site in Ukraine in 2013. It is shown that ensemble of MLPs provides better performance than a single neural network in terms of overall classification accuracy, kappa coefficient, and producer's and user's accuracies for separate classes. The overall accuracy more than 85% is achieved. The obtained classification map is also validated through estimated crop areas and comparison to official statistics.