We used deep learning networks to establish a relationship model among MODIS daily surface reflectance product (MOD09GA) and Arctic melt ponds fraction (MPF), ice fraction (IF), and open water fraction (OWF). We applied this model to MODIS 8-day surface reflectance (MOD09A1) to derive Arctic 8-day MPF and SIF (SIF as the sum of IF and MPF). The results demonstrate that our model improved MPF estimation accuracy to an RMSE of 3.7%, compared with previous models. The characteristics of MPF spatiotemporal changes seen in early summer (May-July) indicate that MPF increased first from May-June, reaching its peak around early July, and then decreased. In addition, early summer MPF was significantly negatively correlated with sea ice extent (SIE) in September. We also found that early summer MPF caused sea ice in the Beaufort Sea, the Chukchi Sea, and the East Siberian Sea to move to warm water. Moreover, the movement of sea ice from the marginal sea to the center of the Arctic was shown to be conducive to the reduction of SIE in September. Early summer MPF was also related to Arctic oscillation (AO) during June to July, and significantly positively related to air temperature in the East Siberian and Chukchi Seas in September. As a consequence, these areas produced more open water and absorbed more heat, reducing the extent of sea ice in September, while increasing their air temperatures. The results also show that early summer MPF has a high negative correlation with air temperature in northern China, and MPF can be used to predict air temperature in northern China. These new findings should be investigated in future studies with additional data collection and field observations.