Crop yields are susceptible to weather events whose characteristics and frequency may be changing with global warming.  Annual changes in yields affect the stability of crop supply and prices.  Hence, analysis of year-to-year variations in agricultural output is crucial but is often constrained by the limited data, particularly in lower-income countries. At the same time, multisector dynamic models like the Global Change Analysis Model (GCAM), require global coverage in the data sets they ingest for experiments. Machine Learning (ML) offers a range of tools that can help. In this work, we investigate whether ML algorithms can be utilized in data-rich regions to understand the complexity of variables relating weather to crop yields and then transfer these relationships to data-scarce regions. With the goal of capturing interannual variation in crop yield in data-limited regions, we propose a novel transfer learning application that aids in accurately predicting the effects of weather variations on crop yields in areas with a dearth of data. We utilize Long Short-Term Memory (LSTM) to develop a transfer learning model. As a proof-of-concept, the model is trained using historical county-level corn yield data in the USA to predict district-level corn yields in India. The transfer learning approach results in an average RMSE of 0.30 bu/acre between observed and predicted yield fluctuations in India. This is comparable to the benchmark model (RMSE 0.23 bu/acre) that is trained and tested in India. The proposed approach can be used in diverse contexts and has the potential to advance the understanding of key socioeconomic and environmental challenges in human-Earth system, especially in areas with sparse data and high shock susceptibility.