Makes more sense now that I’ve a better idea of the target application.
If I understand correctly, what you’re really after here is to be able to resolve small temp deltas in multiple, contemporaneous measurements that occur over minute-scale time frames. Not absolute accuracy or long term stability or even stability with regard to temp change so much, since your measurements are only concerned with what happens in the next minute or so, rather than measuring changes between now and a year from now. That actually offers some significant flexibility.
RTDs offer better linearity, thermistors more sensitivity. The latter might let you get by with a cheaper ADC, though probably would yield less consistent system behavior as a function of baseline temperature. Littelfuse and others are already in the business of making temperature probes, and can do customs if the opportunity is right. That’d be more of an off-line conversation if you’d care to have it.
I’d suggest an arrangement like that napkin-sketched below. The concept would be to use a 2+ input, high-res ADC referenced to the excitation supply and measuring the 2 intermediate points in a 3-element voltage divider created by two temperature sensors with the lowest-tempco resistor you’ve got budget for in the middle. That resistor functions as the reference for the system, which is measuring the proportion of the excitation/supply voltage appearing at the inputs, rather than the actual magnitudes of such. Assuming 1K all around, 16 bits should allow resolving to about 0.06°C with the 3850PPM/°C tempco common among platinum RTDs, if I ran the numbers right.
Assuming a 1K reference resistor value, there are OK, good, and really good options available. I like the LTC2493 as a higher-end ADC option that seems to offer hope of achieving good measurement resolution with the lower sensitivity of RTDs, though the Nuvoton part mentioned might suffice also–some filtering might help with the noise issue.
You mentioned a surface mount RTD; how was that electrically insulated? Sap would likely be conductive enough to mess with the readings if it wasn’t.