Temperature sensor accuracy

If they are waterproof you can stick them in a slush of ice and water to get a zero degrees reference. If they aren't waterproof you can put them in a plastic bag in the ice/water mix.

If they can take 100 degrees, you can put them in a plastic bag in boiling water for a second reference point.

This will tell you which sensor is more accurate and allows 2-point (linear curve fit) calibration.

Thanks Doug!

I think the ds18b20 is water proof, I’ll give that a shot.

I didn’t realize these things needed the ole physics class tricks to get them set up!

And then do I just add or subtract from the readings they give?

If a subtraction fixes both zero and 100 than that is all that is needed. If it doesn't fix both readings, you need to add or subtract a number and multiply by a scaling factor so that the result matches both at 0 and at 100. If you publish your readings we can explain how to calibrate them. 

Comparing data sheets it looks like the accuracy of the DHT11 is ±2°C

whereas the DS18B20 is ±0.5°C

It will be interesting to see what your testing shows.

During the ice-water test, you have to be careful not to touch the ice-cubes. They are still colder than zero degC. I learnt this during a thermocouple roadtest. 

I’m pretty sure mine are from “questionable sources”, so I wonder if they might be defects that didn’t fall within the correct parameters  

Given what you’ve just pointed out though, it seems that a close enough 80-20 solution is to adjust the dht values to match the ds18b20 numbers.

Hi Nico,

If required, these types of boards are quite neat to act as a sort of semi-reference. The accuracy is claimed to be +- 0.1 degrees C within a certain range. I have an older model that claims +- 0.3 degrees C I think.

They can be accessed via Bluetooth on mobile, and can show historical data as well. Link: https://www.newark.com/sensirion/sht4x-smartgadget/ref-design-board-temp-humidity/dp/74AJ2003 

Thanks for pointing that out, that's really accurate!
That would solve the problem of which sensor to trust as a baseline.

A man with a watch knows what time it is. A man with two watches is never sure

I've added it to my wishlist

Good point, thanks for the heads-up on that. I would have been tempted to do exactly the wrong thing!

I've also observed ~ 2° differences when comparing temperature sensors,  and in most cases that's their actual stated accuracy.

As others have told you, you can measure known temperatures and see what the sensors read at those points, and you can do a simple fit regression over them.
For two points is a lot simpler.
Let's say, for example, that you measure the ice melting point (which should be 0°C) and the thermometer says it's 2°C. Then you measure the water boiling point (that should be 100°C) and the sensor reads 99°C That means that near 0 degrees the sensor is measuring 2 degrees higher, and near 100 is 1°C too low.

You want to create a linear function that takes the sensor reading and converts it into a "closer to reality" reading, this is: F(what_the_sensor_says) = what_it_should_be
So F(-2) = 0 and F(99)  = 100.

Since we assume the relationship is linear, the function would have the form F(x) = m*x + b

We can compute the slope (m) with our two points, as [F(x2) - F(x1)] / (x2 - x1), In my example that would be  (100 - 0) / (99 - [-2]) = 100 / 101 = 0.9901

Our function is looking like this now:

F(x) = 0.9901*x + b

We only need to compute b. Replacing any of the two values we know, should allow us to get it. Let's use our reading at 100C:

100 = 0.9901*(99) + b
100 = 98.0199 + b
b = 100 - 98.0199

b = 1.9801

So the final function would be F(x) = 0.9901*x + 1.9801 and you should be able to plug your sensor reading in place of x, and get an adjusted reading. This is of course, for my made-up example. You need to make your own measurements and follow this procedure. It doesn't need to be at 100 and 0, but those are good because you can compare against an actual physical phenomenon.