Seed shape (SS) of soybean (Glycine max (L.) Merr.) is an important morphological trait that significantly affects the quality of marketable seed. Study of the genetic architecture of SS is important and basic to soybean molecular breeding. In the present study, a natural soybean population of 202 diverse accessions mainly from China was used to analyse the genetic basis of SS via genome-wide association analysis (GWAS), which was based on single-nucleotide polymorphisms (SNP) generated by specific-locus amplified fragment sequencing method. In total, 27 335 SNPs were finally identified with minor allele frequencies >5%. By using GWAS, 14 quantitative trait nucleotides (QTNs) were identified to be associated with seed length, 13 with seed width and 21 with seed thickness in four tested environments. Among these QTNs, 21 QTNs overlapped or were located in the linked genomic regions of the reported quantitative trait loci related to SS or seed weight; and the other 27 QTNs were novel loci for SS. Ten QTNs showed environmental stability and were detected under at least two environments. In total, 83 genes were predicted in the 200-kbp flanking region of six stable QTNs that could be detected under >three environments. Gene-based association analysis was performed by using 38 accessions of diverse SS; 778 SNPs were found in the 83 genes based on 38 accessions, and 270 SNPs from 41 genes were found significantly associated with SS. Twenty-eight genes were environmentally stable and/or pleiotropic in controlling two or more SS-related traits at the same time. The identified loci along with the candidate genes could be of great value for studying the molecular mechanisms underlying SS and improving the potential seed yield of soybean.