Transmission of images over a digital channel in the transform domain can lead to reduced bandwidth requirements. This is a consequence of redundancy reduction as the linear transformation compacts the image energy into a small region. By transmitting the transform components that have the most energy/information, images can be reconstructed at the receiver with negligible degradation in the subjective picture quality and with reduced bit requirements. The criterion for selecting these components is generally based on geometrical zone or magnitude or variance all in the transform domain. Magnitude sampling although adaptive, requires additional bits as their location needs to be specified. Variance criterion, on the other hand, is in general adapted to the average picture statistics, and hence may not be an optimal selection for the specific image being processed. As a compromise between these two, hybrid sampling which considers both magnitude and variance is proposed. This technique is applied to GIRL and MOONSCAPE images which are quantized uniformly to 256 gray levels. Processing is carried out on (16 x16) pixel subimages using discrete transforms such as Haar, Walsh-Hadamard, Hadamard-Haar and discrete cosine. Mean square error (mse) for various data compression ratios utilizing hybrid selection between the original and reconstructed images is computed and is compared with those for the magnitude and for the variance selections. The mse for the hybrid approaches that for the magnitude which shows that the former is an attractive scheme for data compression with significant bit reduction and negligible increase in mse. Various ratios for magnitude-variance selection are being adopted. This may lead to an optimal ratio in terms of bit rate, mse and image quality.