I assume your dependent/response variable is onion yield and your independent/predictor variable is onion variety.

I would suggest that, if possible, you group the similar varieties together.

There seem to be many many varieties:

http://agropedia.iitk.ac.in/content/onion-varieties

Option 1: red vs. white onions.

Option 2: 3 kinds:

The three basic types of onions grown in gardens are short-day, long-day and day-neutral. https://www.gardeningknowhow.com/edible/vegetables/onion/onion-plant-varieties.htm

Option 4: 6 types of varieties

https://www.tasteofhome.com/collection/types-of-onions/

Option 2: 11 varieties

https://www.seriouseats.com/2014/06/differences-between-onions-yellow-red-vidalia-what-are-ramps-shallots-how-to-cook-with-onions-guide.html

ANOVA cannot be used when there is only one sample in some categories. However, there are some remedies:

1) Merge groups which are in the neighborhood. In the end, there should be no group with only one sample.

2) Remove groups with one sample.

Note:

The assumptions of normality and constant error variance should hold. However, The one-way ANOVA with the equal sample is robust to departures of normality and constant error variance.

Use exact non-parametric test if the departure of normality and constant error variance is serious.

Technically there is no problem. Unless stated differently, ANOVA assumes variance homogeneity, and k groups with k+1 values in total are sufficient to estimate the variance.

Robustness of conclusions from such small samples is douptful - but that's a different topic.

Instead of removing/ignoring data or use groups that don't fit to the scientific ideas of how the groups should be defined, I suggest to make your assumptions and priors explicit and go for a proper Bayesian analysis.