Yes, binning is a problem. Even without binning, TS and hurricanes are relatively rare events that don't provide sufficient numbers to do statsitics on annual numbers. Perhaps grouping years would help with that. But as you can see on the lower charts where I plot the detection trends, there are lots of dots at zero, 50% and 100%. That means that in many years, with just one or two or three storms, it's likely that zero percent or 100% of storms went near land or stayed away from land out of happenstance. Lots of happenstance can affect the trend. I removed all the year-bins with zero storms in that bin. Also those are not in the trend calculation. Python makes some of those things very easy to do.
But yes, bins are discrete, their thresholds are arbitrary, "near land" is a binary choice, etc. Bins could be gaussian with overlap. Near land could be graded. There are ways to help deal with the small numbers problems that come with binning.