Presented at the 9th European Seminar on COmputing (ESCO): Land surface models are a critical tool in understanding and predicting the Earth's response to climate change and environmental shifts. Accurate and efficient models are vital for assessing water resources, forecasting natural disasters, and informing sustainable land management practices to mitigate climate impacts. This talk introduces two machine learning(ML) techniques that promise significant strides in developing the next generation accurate and scalable land surface models. First, we introduce a novel diffusion-based uncertainty quantification method for efficient model calibration. The approach is a score-based diffusion model that leverages Monte Carlo simulation to estimate the score function and evaluates a simple neural network to quickly generate samples for approximating parameter posterior distributions. Secondly, we introduce a multimodal-data-driven ML model to improve land surface model prediction. This method uses vision language model to extract dynamic land surface characteristics from remote sensing data and then combines these characteristics with hydrometeorological sequences and other land surface static attributes to simulate land surface responses. We demonstrated these two methods in modeling carbon flux and streamflow and obtained promising results.