Abstract: Detecting oil depot has been a hot research field in object detection from remote sensing images. As the images usually are storage by JPEG2000, a oil depot detection algorithm was proposed. With only part data decompression, the object could be detected directly by the wavelet coefficients. According to the different frequency of wavelet coefficients, Hough transform was used to extract circle feature in low-frequency subband and stacked denoising autoencoders could be used to represent the feature of oil depot in high-frequency subband. At last, support vector machine was utilized to make feature fusion and object classification. Many experiments indicate that the proposed algorithm can detect oil depot with high accuracy as well as fast processing speed.