Research Projects

Advancements in satellite technology witnessed a booming growth in the size of remote sensing (RS) image data archives. Storage of these growing data is one of the big challenges faced by the RS research community. Traditional RS image satellite mission used to use compression algorithms such as Differential Pulse Code Modulation (DPCM), Joint Photographic Experts Group (JPEG) or JPEG 2000 to store its data. Efforts were made to perform scene classification/retrieval tasks for several application based on the features extraction from the coefficients obtained from these transformation based encoding-decoding approach used in JPEG/JPEG-2000. However, in this big data era, manually obtaining key features from these individual images based on these transformation based coefficients is time-demanding and operational. Recently, deep learning based methods have demonstrated a remarkable improvements in compression performance due to its ability to handle massive amounts of data and learn key intrinsic features relevant to a particular application. Numerous efforts were made to develop several image compression algoorithms based on these deep learning approaches. In addition, recent nanosatellite missions such as Cube Sat have intergrated an online intelligent processing module using convolutional neural network (CNN) to allow real time processing of data before sending the image to the ground station. The proposed LHFIC-Earth project proposes to consider relevant areas of interest and adaptive assign bit rates to those pixels to achieve a higher compression ratio. The project aims to study the possibility to i) allocate minimal number of bit rate to pixels that comes within the area of interest and vice versa; ii) obtain relevant features within the compressed domain to perform scene classification/retrieval from the compressed image archives; iii) to study the trade-off between the computational complexity and performance of the proposed model with resepect to the relevant applications. Existing works only consider the same bit rate being allocated to the pixels. But the proposed works claims to attain even better compression ratio considering the various characteristics of the RS images. The proposed work initially requires a thorough outline on the various compression algorithms used for RS data. The main challenge is to accurately obtain the mask for the area of interst for compression. The strength of this proposed work is that the proposed methods i) can be integrated for onboard processing such that the compressed images can be send to the ground station and only images with relevant key features can be decompressed and ii) can be used in other research communities to minimize the size of the stored data by changing the preprocessing approaches used. The work is expected to have a high impact over the space science community considering the trend of using nanosatellites that basically works with intelligent processing modules.

Artificial Intelligence, onboard processing, remote sensing, image compression, scene classification.