I think you'll want to go down the path busz suggests. Search for the concept "diode noise". The PN junctions in diodes and transistors can produce close to perfect Gaussian white noise. Sampling that should be a source of entropy that's better than any environmental source.
The problem with most environmental/ambient data is the values just don't change that much over time: temperature, humidity, light and sound all have less than an order-of-magnitude of variability with really strong modes. An accelerometer to measure motion might be a good source of variability if mounted on a person, but you'd likely have to do a bit of signal processing to remove the normal modes of oscillation that are present in how humans move. An ambient light & sound source might have some pretty high variability if placed in an high-density urban space, but again I think there would be a lot of repetition. I still think the best source of entropy would be going down towards fundamental physical properties of materials like diode noise than going up in scale and looking at environmental factors capable of being read by a microcontroller.
Most micros support a low-power 32.768 kHz watch crystal oscillator with some kind of prescaler and timer interrupt. Set the prescaler so the timer is counting slowly and the interrupt happens at the period you desire.
Some micros also have a built-in low-power RC timer if exact timing isn't critical.
The datasheet for any low-power micro will list the power with 32.768 oscillator (and nothing else) running. It's pretty close to zero. You can do the math to see if this is acceptable, and compare it to the current drawn by the watchdog.
OK, for example on the msp430f2013, let's look at power in the datasheet.
0.5 μA is almost zero, although it is five times the true OFF mode.
For more detail, we can look inside the datasheet.
Going from LPM4 (everything off) to LPM3 (running the oscillator) is the difference between 0.5 μA and 1 μA.
Suppose the battery is CR2032 with 225 mAh capacity. Then standby in LPM4 is about 50 years and in LPM3 is about 25 years. 25 years is long enough for many applications, because the ON-current (during the measurement itself) dominates the consumption.
Best Answer
How bad do you to use the XMega? If the crypto and random number generation are a big part of your project, Atmel's SecureAVR series has a hardware random number built in, and is designed for cryptographic applications.
Regardless, I doubt that you'll find a random seed source that has a good distribution. You'll want to run it through a pseudo random number generator a few times As long as you start with a different seed every time, this will give you a nice set of random numbers. An LGC is a quick and easy pseudo random generator: