Functional magnetic resonance imaging (fMRI) has demonstrated significant potential in the early diagnosis and study of pathological mechanisms of Alzheimer's disease (AD). To fit subtle cross-spatiotemporal interactions and learn pathological features from fMRI, we propose a fine-grained spatiotemporal graph neural network with self-supervised learning (SSL) for diagnosis and biomarker extraction of early AD. First, considering the spatiotemporal interaction of the brain, we design two masks that leverage the spatial correlation and temporal repeatability of fMRI. Afterwards, temporal gated inception convolution and graph scalable inception convolution are proposed for the spatiotemporal autoencoder to enhance subtle cross-spatiotemporal variation and learn noise-suppressed signals. Furthermore, a spatiotemporal scalable cosine error with high selectivity for signal reconstruction is designed in SSL to guide the autoencoder to fit the fine-grained pathological features in an unsupervised manner. A total of 5,687 samples from four cross-population cohorts are involved. The accuracy of our model was 5.1% higher than the state-of-the-art models, which included four AD diagnostic models, four SSL strategies, and three multivariate time series models. The neuroimaging biomarkers were precisely localized to the abnormal brain regions, and correlated significantly with the cognitive scale and biomarkers (P< 0.001). Moreover, the AD progression was reflected through the mask reconstruction error of our SSL strategy. The results demonstrate that our model can effectively capture spatiotemporal and pathological features, and providing a novel and relevant framework for the early diagnosis of AD based on fMRI.