Aiming to reduce the limitation in local background covariance matrix estimation of RX algorithm，an improved RX (I-RX) algorithm is proposed for anomaly detection in hyperspectral imagery.Based on a singular value decomposition (SVD)，We firstly project the hyperspectral imagery onto the background orthogonal subspace to obtain the remaining imagery which only consists of noisy background and anomaly.On this basis，by calculating the spatial rank depth value of every sample，the remaining imagery can be divided into two sample sets：noise background set and potential anomaly set.Using the noise background set to estimate the background covariance matrix of the whole imagery and the potential anomaly set as test examples to be detected whether has anomaly or not.Numerical experiments are performed on simulated data and real hyperspectral data.The ROC curves demonstrate that the detection probability of I-RX algorithm is about 2 times than RX algorithm at the same false alarm rates.