Lossy compression has been widely used in various applications due to its variable compression ratio. However, distortions are introduced unavoidably, and this decreases the image quality. Therefore, it is often required to control the quality of the compressed images. A two-step method has been proposed recently to provide the desired visual quality. The average rate-distortion curve was used to determine the proper parameter value that controls compression. However, its performance for the wavelet-based coder Set Partitioning in Hierarchical Trees (SPIHT) is insufficient because there are very wide limits of visual quality variation for different images for a given value of the compression control parameter (CCP). Additionally, previous work has demonstrated that the level of errors, which is the subject of our study relates to texture features of an image to be compressed, where texture presence is an inherent property of remote sensing images. In this paper, our goal is to develop an adaptive two-step method for SPIHT to improve accuracy. The following tasks were solved. First, a prediction of visual quality for a particular parameter value is conducted. The prediction scheme is based on the information extraction from a certain number of image blocks to perform a visual quality calculation of the image compressed for a given CCP value. A threshold is adopted as the complexity grouping; in this paper, images are divided into two groups: simple and complex images. Second, the results of the grouping determine the adaptive curve model adopted. Finally, a two-step compression method is applied according to this curve. The classical metric Peak signal-to-noise ratio (PSNR) is employed to evaluate the image quality. The research method is based on a validation experiment that is conducted for an image set covering different image complexity and texture features. The comparison results of four typical desired values prove that the accuracy has been generally improved, the variances of both the first and second steps have been reduced sufficiently, and the mean absolute error has also been improved. Conclusion: the improvement effects are significant, particularly in the low desired visual quality. A remote sensing image is taken as an example to analyze in detail; the quality of the decompressed images meets the user’s visual requirement, and the errors are acceptable.