Remote Sensing Image Lossy Compression Based on JPEG with Controlled Visual Quality

Remote sensing image compression plays an important role in Earth observation applications. Lossy compression techniques are efficient means to reduce the size of images, although the distortions are unavoidable. Consequently, the image quality should be controlled carefully. This paper discusses this topic with its application to JPEG. Using visual quality metrics, we show that JPEG is able to provide compression performance not worse than many advanced coders for invisible distortions and/or if images have complex content. We show that mean square error (MSE) can be provided by a two-step approach and predicted with high accuracy using a simple neural network. Such a neural network employs QF as one of the inputs and several parameters characterizing image complexity as other inputs. Different sets of input parameters are analyzed, and we show that even four input parameters produce a quite accurate prediction of MSE and, respectively, PSNR, with errors less than 1.5 dB with low computation load. This is shown using images used in training and for verification and validation image sets. Then, it is easy to determine a needed QF to provide a desired quality.