In this paper, an adaptive sparse pattern learning aided joint activity detection and channel estimation scheme is proposed for grant-free random access with affine frequency division multiplexing (AFDM) in LEO satellite communication systems. Specifically, the channel parameters and active index will be detected in the context of unknown sparsity. First, a sparse prior is preset to perform active detection and channel estimation. Subsequently, the prior is refined by a graph network architecture, where the delay-Doppler grid is modeled as the graph nodes delivering the message. Furthermore, a novel graph aggregation module based on self-attention mechanism is designed to capture the correlation between nodes to infer the unknown sparsity patterns inherent in dynamic terrestrial-satellite links (TSL). Simulation results demonstrate that, compared with conventional nearest neighbor sparsity pattern learning (NNSPL), the proposed scheme achieves a 2 dB improvement in normalized mean-square error (NMSE) and reduces pilot overhead by approximately 25%.