Radio Tomographic Imaging (RTI) is a promising wireless imaging technology to reconstruct the attenuation map in an area covered by a dense radio frequency (RF) wireless network. The attenuation map is also referred to as the spatial loss field (SLF), which quantifies the attenuation rate of the RF waves at each location in the network. Any object on a signal propagation path will cause signal attenuation on received signal strength (RSS) measurements between wireless transceiver pairs. Therefore, SLF can be recovered from the RSS measurements. With the reconstructed SLF images, RTI can be utilized to realize passive object localization/detection, environment monitoring, and even through-wall imaging in security systems. The objective of the thesis is to explore deep learning techniques for RTI with both one-shot and online designs, with the goal of achieving high-resolution SLF recovery while considering diverse wireless environments. In the context of one-shot estimation, deep learning methods use a single snapshot of RSS measurements to reconstruct SLF. In the online estimation scenario, deep learning approaches consider a time sequence of RSS measurements and leverage the correlation between current and past observations to estimate SLF. Moreover, we investigate the problem of system model mismatch between the mathematical approximate RTI system model and the actual diverse environments. Efficient transfer learning schemes are developed to adjust the deep-learning-based RTI techniques in different environments with mismatched models.