A predictive tree structure is presented for classifying the wavelet coefficients, and a new scheme is proposed to construct the trees based on rate distortion function, including both the optimal hierarchical quadtree construction and the predictive spatial orientation tree development. The full search quadtree optimization is applied first to the highest level of high frequency subbands, exploiting the intrasubband correlation of wavelet coefficients. The generated optimal quadtree serves as a predictor to construct the trees for other lower level subbands in which only the leaf nodes are to be analyzed in terms of the associated Lagrange costs for further expansion, taking advantage of the self-similarity across subbands. Constraining the full search quadtree optimization within the highest level subband reduces the computational complexity significantly. Simulation results indicate the proposed scheme is efficient and the performance of the system is comparable to some of the popular image compression techniques.