Spatial data are being collected at unprecedented scales: the volume of Earth observation data is expected to grow to hundreds of petabytes by 2025, and the number of GPS receivers has surpassed 6 billions. With more revolutions to come, today’s satellites can already create a high-resolution scan of the entire Earth’s surface on a daily basis. Such data provide critical information across sectors including agriculture, water, health, climate, infrastructure, etc. Direct applications of machine learning on spatial data often fall short due to their unique characteristics. This talk will start with the fairness aspect, discussing its impact and formulation in the spatial context, as well as associated challenges and learning frameworks. It will also show knowledge-guided fairness adaptation methods to generalize fair models to unseen regions and future timestamps. Next, the talk will introduce spatial variability, a cause of the spatial bias, and a spatially-explicit framework to automatically recognize and separate regions with different distributions. Finally, the talk will present examples of social good applications of spatial AI in climate change (e.g., carbon budget), crop monitoring, poverty mapping, etc.