Extract a significant sample of data from a non-normal distribution before modelling a fitting curve Background: In my master thesis I am measuring how many areas(in square meter) are suitable for greening at different building height in an urban area. So I have two variable: Suitable Areas and Height. In the following plot is it shown how the amount of Suitable Area changes at different heights and I plot a curve that join those points (an automatic function of ggplot2 in R) 
Now my supervisor required to model a fitting curve (and to find its formula) that describes the data(the distribution of suitables area at different height), but before doing so he suggest me to consider reducing my data. Therefore the question is at which height the 95 % of the total Suitable area are located and consequentially remove the rest from my distribution. I guess this would mean to cut the tail on the left of my plot.
I thought that this 95% of the measurement comes from the "68-95-99.7" rule (three-sigma rule) based on standard deviation. But the three-sigma rule works better for normal distribution, while my distribution is non-normal, and I fear that it would neglect values in the first meters, considering them as outliers.


*

*Is a good practice at all to select a significant sample of my
distribution before modelling a fitting curve?

*Which statistical methods are given to find significant sample of
my data?  

*Since I consider the measurements in the first
meters(highest frequency of suitable area) very important in my
distribution should I use a method not based on standard deviation?


Sorry if my question is unprecise but I have poor mathematical knowledge.
 A: I extracted data from the scatterplot for analysis. Except for the first two data points, I see two separate models - one model for height below 16 meters and one model for height above 16 meters. Combining these two models would not require data reduction. I suspect that there might be some dominant effect above and below approximately 16 meters.
For height below 16 meters a hyperbolic type equation "y = (a + (b * x)) / (c + x)" seems to fit well, with parameters a = -3.6248957321627102E+04, b =  1.6715762703817916E+04, and c = -2.5094538863935254E+00 yielding RMSE = 149.6 and R-squared = 0.9908:

For height above 16 meters a Lorentzian peak type equation, " y = a/ (b + pow(x-c, d))" seems to be a possible candidate, with parameters a = 6.7383016504505267E+06, b = 4.0349701664549445E+02, c = 1.6059999999999995E+01, and d = 3.2099267050815645E+00 yielding RMSE = 374.0 and R-squared =  0.9957:

I did not model any overlap near 16 meters, adding one or two points of overlap between the two models would be a good idea. If you use the "two models" route, the values I have provided should be good initial parameter estimates for your own non-linear fitting with the actual data.
