This manuscript presents an overview of my work in the field of geospatial machine learning, a rapidly growing interdisciplinary field that poses many methodological challenges and has a wide range of impactful applications. Throughout my research, I have focused on developing bespoke approaches that leverage the unique properties of geospatial data to create more efficient, precise, and parsimonious models. This manuscript is divided into four main chapters, each covering a different property of geospatial data structures that can be leveraged algorithmically. The first chapter presents a versatile mathematical framework formalizing the concept of spatial regularity with graphs. We propose an efficient algorithm that tackles a broad family of spatial problems and provides novel convergence guarantees and significant speed-ups compared to generic approaches. The second chapter introduces a deep learning method that extends the idea of exploiting graph regularity to the case of massive 3D point clouds. We simplify the task of large-scale semantic segmentation by formulating it as as a small graph labelling problem. Our compact models reach high precision at a fraction of the computational cost of other approaches. In the third chapter, we present a collection of methods designed to take advantage of the data structure inherited from 3D sensors. By considering the sensors’ structure, we develop powerful networks with state-of-the-art accuracy, latency, and robustness for various applications and data types. The last chapter dives into the real-life challenge of automated satellite time series analysis for crop mapping. Recognizing the difference between such data and standard formats used in computer vision, we propose novel and streamlined architectures that achieve unprecedented precision while remaining efficient and economical in memory and preprocessing. We also introduce the task of panoptic segmentation for satellite time series and an efficient architecture to solve this problem at scale. In summary, this manuscript argues that geospatial problems represent a challenging and impactful venue for evaluating the newest machine learning and vision methods and a fertile source of inspiration for designing novel approaches.