In RX anomaly detection algorithm, when backgroud being contaminated from anomaly pixels, the local backgroud covariance matrix estimation can not reflect backgroud distribution accurately, which results in low detection capacity. To overcome this problem, a new method based on the robust background subspace was proposed. Utilizing the spatial rank depth, the position of every sample relative to the distribution space of whole background samples could be measured. Samples which locating at the edge of the distribution space were regarded as anomaly, and being mapped into the distribution space. In this way, the local background covariance matrix was estimated, and the principal component analysis as background space was obtained which can characterize background more accurately. An anomaly detection model was constructed on this subspace using mahalanobis distance. The effectiveness of the proposed method is validated by experimental results from simulated and real data.