For RNAseq, the differences in the read depth can simply arise because of expression (that's trivial). Highly expressed regions will have higher number of reads.
However the read depth non-uniformity also arises because of the nucleotide composition of the regions. This is true for all kinds of nucleic acid sequencing such as RNAseq, whole genome sequencing and ChIP seq. This is called sequence bias. This also depends on the sequencing methodology. A variety of factors play a role such as how well can the region be sheared, how well can it be amplified etc.
For details have a look at this article by Ross et al (2013).
Sequencing technologies are vulnerable to multiple sources of bias.
Methods based on bacterial cloning and Sanger-chemistry sequencing 
were subject to many coverage-reducing biases, notably at GC extremes,
palindromes, inverted repeats, and sequences toxic to the bacterial
host [9–17]. Illumina sequencing  has been shown to lose coverage
in regions of high or low GC [19–22], a phenomenon also seen in other
'next-generation' technologies [3, 6]. PCR amplification during
library construction is a known source of undercoverage of GC-extreme
regions [20, 21] and similar biases may also be introduced during
bridge PCR for cluster amplification on the Illumina flowcell .
Illumina strand-specific errors can lead to coverage biases by
impairing aligner performance . Ion Torrent , like 454 ,
utilizes a terminator-free chemistry that may limit its ability to
accurately sequence long homopolymers [4, 27, 28], and may also be
sensitive to coverage biases introduced by emulsion PCR in library
construction. Complete Genomics  also uses amplification along
with a complex library construction process. The Pacific Biosciences
 process is amplification-free; therefore, one might expect it to
exhibit lower levels of coverage bias than the other technologies.
In addition to sources in the wet lab, bias can be introduced by any
of the computational steps in the sequencing pipeline.
Signal-processing and base calling limitations could result in
under-representation or increased error rates in some locations, as
can inaccurate alignment. An inaccurate reference or sample-reference
differences can cause coverage or accuracy variations that may be
misdiagnosed as sequencing bias. Therefore, detecting bias is only the
first step and must be followed by more detailed experiments to assign
responsibility to the library preparation, sequencing, or
Some analysis software try to correct for these biases while estimating the read abundance.
Ross, Michael G., et al. "Characterizing and measuring bias in sequence data." Genome Biol 14.5 (2013): R51.