Spatial interactions between multiple cities are important to the temporal and spatial evolution of urban expansion, and even significant to the carbon cycle. In this paper, an ELM-CA model was proposed by introducing extreme learning machine (ELM) into cellular automata (CA) to obtain the CA's conversion rules. Taking Guangdong Province as an example, the effects of urban expansion on net primary productivity (NPP) were investigated by coupling Biome-BGC with the ELM-CA model. To represent the close interconnections between different cities, their spatial interactions were explicitly embedded in the ELM-CA model. Our results indicated that: the ELM-CA model could simulate the urban expansions in Guangdong Province at a high accuracy. In addition, the urban expansions exhibited crucial impacts on the NPP in Guangdong, which reduced the vegetation NPP evidently. According to the inertial trends of the urban expansion from 2000 to 2005, we found that the urban land development in 2020 may cause a reduction in NPP, which had taken up about 1.79% of the total provincial NPP of Guangdong. In summary, a reasonable guidance on the planning of future urban expansion is critical for the maintenance of carbon balance and climate change.