Space-time adaptive processing (STAP) for bistatic airborne radar offers several advantages, such as the higher possibility of detecting stealth targets. However, in a bistatic environment, the usual impediment and possible clutter in-homogeneity is further complicated by the rangedependent nature of the clutter ridge in the angle-Doppler plane induced by the physical geometry of the two aircrafts. This complicates the clutter suppression problem and leads to signi cant degradation in performance. The major objective of this thesis is to develop training methods for bistatic radar operation in a dense environment of ground-moving targets. The work is directed towards what may be called `small STAP', where the number of spatial channels is small and the array is non-uniform. The work is motivated by a desire to minimise the amount of navigational data associated with both the transmitter and receiver. Furthermore, it is directed towards environments where all range gates may contain targets. This thesis presents several novel STAP approaches, which can be classi ed into two main categories, to address the range dependency problem within a bistatic airborne radar framework. The rst category is on training strategies for joint-domain localised (JDL)-STAP in a bistatic environment. The JDL algorithm is originally proposed to reduce the computational complexity for monostatic radar by using a two-dimensional discrete Fourier transformation to transform the data from the space-time domain into the angle-Doppler domain. However, it has restrictions that essentially assume the receiving antenna to be an equi-spaced linear array of ideal, isotropic, point sensors. Two novel algorithms are proposed to overcome these two restrictions and they incorporate angle and Doppler compensation into the JDL processor to mitigate the bistatic clutter Doppler range dependency problem. In addition, a novel JDL in-the-gate processing approach is proposed, which forgoes the training data requirement and operates solely on the test data set. This single data set detection approach alleviates the high target density or heterogeneity problems associated with the training data requirement of conventional STAP algorithms. It is particularly applicable to heterogeneous environments where the clutter homogeneity assumption does not hold or independent training data is not readily available. The second category is on bistatic STAP training without navigation data. A novel technique is proposed to predict the range-dependent inverse covariance matrix, which is used to compute the STAP lter weights, by utilising linear prediction theory. The proposed technique provides mitigation against additional clutter notches resulting from range and Doppler ambiguities. It also allows for detection in other range gates under test without having to re-compute the prediction weights. Another novel technique is proposed to obtain an estimate of the rangedependent inverse covariance matrix by using an eigen-analysis based method. This technique involves applying eigen-decomposition to the covariance matrix in each range gate, sorting the eigenvalues by using maximum inner-product of the eigenvectors of the training range gate with respect to the test range gate and then averaging the resulting sorted eigenvalues. Both of the proposed techniques eliminate the requirement for a uniform linear array and can be applied to arrays of arbitrary con guration. No navigational data or parameter estimation is necessary as only the clutter data is required, thus reducing real-time computational costs.