Hi, the simple answer is - you can't. It's a broadband greeness index. You can receive the same values for algae species as for a forest. For tons per hectar you should incorporate 3D information as for instance from LIDAR or spaceborne SAR measurements. Those kind of inventories are quite complex. Maybe you can find more on this topic here: https://www.uni-trier.de/index.php?id=11282
I agree with Christian RogaB. NDVI comes from TM data which is optical. That means it is like taking a picture from a camera - thick broad-leaf vegetated area and the green rice field may have the same NDVI - so no way to extract in weight or volume. Sorry!
You can develop a empirical 'local' yield model wherein you may take NDVI as a surrogate. Field based measurements will be key input and the results may not be that accurate. Moreover, this yield model can't be generalized over other areas.
If you have filed sampled yield then you can correlate with NDVI and can develop regression model. That model you can apply to NDVI image to observe yield in the entire area.