Limited downlink rate is a bottleneck while transmitting large datasets acquired by imagers onboard spacecraft to Earth. Compressive Sensing (CS) provides an effective sensing and reconstruction methodology that is suitable under limited onboard computational resource environment in addition to the low downlink rate. However, we observe that random subsampling of a scene by CS leads to higher reconstruction error at regions in an image which consist of high frequency components. These high frequency components are often regions of interest (ROI). Very low reconstruction error is essential at the ROIs for further analysis by human experts. We propose a blocked ROI aware CS approach where the image is divided into blocks and a modified sensing matrix acquires higher number of samples at blocks overlapping with ROIs. This varying sampling density across an image is determined by leveraging the information from past observations of the scene, which automates the generation of the ROI based non-uniform sampling pattern. This technique is also suitable for ROIs that have varying spatial extent in a scene between successive observations. We develop an algorithm to vary the sensing matrix based on changes in the ROI. Comparison of image quality of CS with ROI aware CS shows significant improvement in reconstruction error at ROI at the same compression ratio by retaining the scientific quality of the image. We also observe the same image quality to be reached at a much higher compression ratio by the blocked ROI aware CS than conventional CS showing a potential improvement by a factor of four. Additionally, reduction in memory requirement due to compression implies an increased potential of the imager to acquire higher number of observations overall.