This work considers how large-scale datasets and novel machine learning methods can be applied to challenges in climate and sustainability, with a particular focus on agriculture. Effectively leveraging these advancements for sustainable development research requires answering two questions: first, how can complex data be translated into useful and accurate information? And second, under what circumstances does this information offer real insight into an important problem? In answer to the second of these questions, the research in the three chapters of this dissertation falls broadly into one of two categories: problems for which high spatial- or temporal-resolution data is necessary but infeasible to collect at scale (Chapters 1 and 3); and problems for which the structure of relationships between features and outcomes is complex, with important non-linearities, interactions, or other nuances that may be overlooked by traditional approaches (Chapters 1 and 2). Both such categories of problem are common in the domain of agriculture, an industry which is critical for food security and economic well-being, but highly susceptible to fluctuations in weather and climate. In Chapter 1, I introduce and validate a method for creating high-resolution estimates of planting and harvest dates for United States crops with satellite imagery. This data is an important input for many research applications, but is only tracked at the state level. The resulting dataset is then used to generate more accurate measures of the weather conditions crops are exposed to during their growing season, and thus more precise estimates of how these conditions impact yields. These estimates suggest a 17% larger impact of extreme heat (>29C) on crop yields than previously documented, with substantial variation in heat sensitivity over the course of the growing season. However, the overall impact of increased temperatures is partially offset by a reduced estimate of growing season duration and a 276% increase in the estimated benefits of warm (10-29C) temperatures. Finally, I present novel evidence that farmers use early planting as a form of adaptation to warming, with planting dates shifting earlier by 0.13 days for each additional 30C day during the growing season. Chapter 2 presents an even more flexible formulation for estimating US crop yields. I introduce a deep learning model that predicts yields directly from daily weather data, and show that it reduces out-of-sample error by 10.7% relative to standard linear modeling approaches. Using interpretable machine learning techniques, I demonstrate that this model learns a number of nuanced patterns consistent with expectations from agronomic theory, including spatial and geographic variation, interactions between weather features, and nonlinearity over weather feature values. Over several simulations, these models estimate future impacts of warming that are two to three times less severe than prior modeling approaches would suggest. However, the complexities of causal identification with highly flexible models mean that these results must be interpreted with caution; primarily, they suggest that estimates of climate impacts may be highly sensitive to feature selection, and to precise trends in warming over the course of the growing season. Finally, Chapter 3 turns to smallholder farms in Kenya, as part of research done with support from Atlas AI. A collection of approaches for real-time yield monitoring at the field level are introduced and tested, using satellite-based assessment of vegetation health. I discuss a remotely-sensed proxy for crop yields for use in environments where reliable ground truth data is unavailable, and present a model that can capture 73.5% of variation in this yield proxy by roughly 6 weeks post-planting. A range of approaches are evaluated for incorporating location- and crop-specific features, handling low volumes of training data, and adjusting for variable timing of satellite imagery collection. Taken together, these chapters demonstrate the value of remote sensing and machine learning for understanding the impacts of climate on crops and identifying strategies for adaptation. They also emphasize the complementarity between novel machine learning approaches and traditional statistical and economic methods: in Chapter 1, for example, satellite imagery is used to generate a novel dataset for analysis with more standard models; and in Chapter 2, I present a non-parametric approach to feature discovery for future causal inference work. Finally, these chapters demonstrate that estimates of climate impacts can be highly sensitive to what features are used and how they are encoded; this underscores the importance of careful consideration in constructing accurate feature inputs, and caution in interpreting the results of any one model.