Soils, particularly near the earth surface, are composed of rich nutrients from organic plant materials and animal matter. It is believed that soils, especially agricultural soils, can adsorb carbon dioxide over a billion tons each year. Therefore, soil-based carbon dioxide adsorption may be considered as a negative emission technology that naturally removes CO2 from the atmosphere. To determine how much soil-based carbon dioxide adsorption can help fight climate change, one needs to model carbon dioxide storage in soils. To advance our knowledge of soil-based CO2 sorption, we analyze soil images available through the Molecular Observation Network (MONet) program. We numerically simulate soil CO2 adsorption and desorption under ambient conditions using the Tahoma supercomputer at EMSL. We also comprehensively analyze soil images and determine some other important soil characteristics (e.g., specific surface area, pore connectivity, tortuosity, as well as backbone and dead-end fractions) that are not reported through the MONet program but affect the CO2 sorption process in soils, thus completing MONet efforts as well. To predict CO2 adsorption and desorption in soils, we develop deep learning models using two different sets of input parameters. First only soil images are used to train and test the model. Next, in addition to images other data collected on soil samples (e.g., pore connectivity, specific surface area and backbone and dead-end fractions), are used to train and test the model. To further assess the deep learning model, we evaluate the developed deep learning models using unseen data and soil images.