The　global　food　supply　system　is　under　increasing　pressure　due　to　population　growth　and　more　extreme　climate　events.　Developing　forecast　models　for　accurate　prediction　of　crop　yields　is　helpful　for　early　warning　of　food　crises.　Amid　the　different　environmental　predictors,　soil　moisture　(SM)　is　an　important　agricultural　drought　indicator.　However,　current　operational　microwave　SM　products　have　generally　low　spatial　resolution,　challenging　the　effective　characterization　of　SM　spatial　heterogeneity.　In　this　study,　empowered　by　the　hourly　land　surface　temperature　(LST)　observations　from　geostationary　operational　environmental　satellites　(GOES),　we　first　spatially-downscale　SM　using　machine　learning　(ML)　algorithms.　Then,　by　designing　three　sets　of　experiment　respectively　using　downscaled　SM,　coarse-resolution　SM,　and　precipitation　observation,　we　assess　the　comparative　performance　of　downscaled　SM　among　its　counterparts　in　estimating　crop　yield　variability,　based　on　three　mainstream　ML　algorithms　and　two　traditional　regression　algorithms.　Our　research　shows　that　downscaled　SM　based　on　high　temporal　resolution　GOES-LST　demonstrates　outstanding　performance　in　characterizing　the　spatial　variation　of　SM.　With　respect　to　yield　estimation,　downscaled　high-resolution　SM　out　performs　coarse-resolution　SM　and　precipitation　products,　with　the　average　R-2　between　the　crop　yield　estimates　and　the　yield　records　being　0.814,　0.809,　and　0.805,　respectively.　In　addition,　we　find　that　among　the　five　algorithms,　the　nonlinear　ML　algorithms　exceed　the　linear　algorithms　in　crop　yield　estimation,　with　the　average　R-2　being　0.827　and　0.783,　respectively.　Our　research　demonstrates　the　great　potential　of　infusing　different　satellite　information　to　improve　the　monitoring　of　crop　growing　status　and　yield　prediction.