Abstract: ADDRESSING THE ISSUES OF COMPLEX IMAGING MECHANISMS AND LIMITATIONS DUE TO SPECIFIC PRIOR INFORMATION IN TRADITIONAL COMPRESSIVE SENSING ALGORITHMS WHEN PROCESSING SPARSE DATA, A SPARSE IMAGE SEGMENTATION METHOD FOR GROUND-PENETRATING SAR BASED ON CONVOLUTIONAL NEURAL NETWORK U-NET IS PROPOSED.TRANSFORMING THE RECONSTRUCTION PROBLEM OF SPARSE DATA INTO A SEGMENTATION PROBLEM TO SOLVE IT, AVOIDING THE COMPLEX DESIGN AND OPTIMIZATION OF TRADITIONAL RECONSTRUCTION, ACHIEVING END-TO-END TARGET EXTRACTION.THE METHOD ADOPTS A THREE-LAYER LIGHTWEIGHT U-NET ARCHITECTURE, AND THE INPUT OF THE NETWORK IS REAL, IMAGINARY, AND AMPLITUDE THREE-CHANNEL DATA AND EMPIOYS A HYBRID LOSS FUNCTION THAT COMBINES BCE LOSS AND THE DICE COEFFICIENT SIMULATIONS SHOW THAT THE RECOVERY EFFECT OF THIS METHOD IS BETTER, AND WHEN THE MIXED WEIGHT RATIO IS 6: 4, THE NETWORK ACHIEVES OPTIMAL PERFORMANCE.