This chapter presents a robust training sample selection method for space-time adaptive processing (STAP) based on the similarity of clutter reconstructed using atomic norm minimization (ANM). Traditional sample selection methods struggle in heterogeneous environments due to the difficulty in acquiring independent and identically distributed (IID) training samples. The proposed method first reconstructs the clutter covariance matrix (CCM) using ANM, addressing the off-grid problem. A clutter similarity metric is then introduced, leveraging the reconstructed CCM to assess the similarity between training samples and the cell under test (CUT) without extracting the clutter subspace. Outliers are eliminated using the generalized inner product (GIP) method, and samples with clutter distributions similar to the CUT are selected. Numerical experiments show that the proposed method enhances STAP performance by effectively selecting representative training samples, demonstrating its robustness and efficiency in heterogeneous environments.