Subdural electrode arrays provide stable, less invasive electrocorticogram (ECoG) recordings of neural signals than multichannel needle electrodes. Accurate reconstruction of intracortical local field potentials (LFPs) from ECoG signals would provide a critical step for the development of a less invasive, high-performance brain-machine interface; however, neural signals from individual ECoG channels are generally coarse and have limitations in estimating deep layer LFPs. Here, we developed a high-density, 32-channel, micro-ECoG array and applied a sparse linear regression algorithm to reconstruct the LFPs at various depths of primary motor cortex (M1) in a monkey performing a reach-and-grasp task. At 0.2 mm beneath the cortical surface, the real and estimated LFPs were significantly correlated (correlation coefficient (r); 0.66 ± 0.11), and the r at 3.2 mm was still as high as 0.55 ± 0.04. A time-frequency analysis of the reconstructed LFP showed clear transition between resting and movements by the monkey. These methods would be a powerful tool with wide-ranging applicability in neuroscience studies.