As the question is not very specific, I can only submit some considerations:

• Use a better sensor
• Apply better signal evaluation (there exist techniques to e.g. "wring" 12-Bit valid data out of a 10-Bit ADC)
• Implement more sophisticated signal evaluation (e.g. self-calibration)