Cloud detection from multilayer perceptron neural networks on Landsat images

Abstract

In the following article the methodology comes to detect clouds using Landsat satellite images, which complicates the use of data in the optical domain of the satellites, as influencing their analysis, causing inaccurate atmospheric correction, skewing values index normalized difference vegetation (NDVI), misclassification and land cover changes confusion vegetable coverages. The resulting model becomes a support tool for further studies in multiple disciplines to facilitate the process of analyzing different phenomena or studies being developed. For this is taken as an example, the area between the Central and Occidental mountain ranges near the town of Puerto Berrío, Antioquia, where there is information relating to 12 Landsat 7 satellite images. Cloud detection network analysis is used as a powerful tool to classify different elements, making more efficient computational process, integrating information from multiple sources to incorporate new features and also dispense with the use of statistical models unlike other approaches. The process allowed train the network with EMC 0.0339 with 7 neurons in the hidden layer, with less than 100 iterations for each of the images used. The method obtained an overall accuracy of 91% better than the results achieved with the method of thresholds developed in previous studies, the accuracy was 87.37%.