Evaluating the Deep Learning accuracy in data extraction from synthetic image sequences

Abstract

We have investigating the use of Deep Learning (DL) to process sequences of images captured by satellites aiming to improve the quality of river flow forecasting methods, because current ones are still not accurate enough to support efficient management of large national electrical systems. Towards this goal, we are assessing the accuracy of DL networks in extracting information from image sequences by means of classification processes. We have set a test environment composed by an image sequence generator, some generating models, the Nvidia DIGITS tool, two DL preset networks, and the needed hardware. Each model produced one image sequence and one data series corresponding to a selected measure in the images. We have trained the DL networks and evaluated its accuracy in extracting the measured data. In this paper, we show that the performance of DL is extremely sensible to the image type, the measure taken into account, and the DL network applied. Our process presented better performance recognizing coverage area rates in images that resemble clouds and linear distances, but had poor accuracy with angles.