Lameness in dairy cattle is a welfare concern that negatively impacts animal productivity and farmer profitability. Micro-Doppler radar sensing has been previously suggested as a potential system for automating lameness detection in ruminants. This thesis investigates the refinement of the proposed automated system by analysing and enhancing the repeatability and accuracy of the existing scoring method in cattle mobility scoring, used to provide labels in machine learning. The main aims of the thesis were (1) to quantify the performance of the micro-Doppler radar sensing method for the assessment of mobility, (2) to characterise and validate micro-Doppler radar signatures of dairy cattle with varying degrees of gait impairment, and (3) to develop machine learning algorithms that can infer the mobility status of the animals under test from their radar signatures and support automatic contactless classification. The first study investigated inter-assessor agreement using a 4-level system and modifications to it, as well as the impact of factors such as mobility scoring experience, confidence in scoring decisions, and video characteristics. The results revealed low levels of agreement between assessors' scores, with kappa values ranging from 0.16 to 0.53. However, after transforming and reducing the mobility scoring system levels, an improvement was observed, with kappa values ranging from 0.2 to 0.67. Subsequently, a longitudinal study was conducted using good-agreement scores as ground truth labels in supervised machine-learning models. However, the accuracy of the algorithmic models was found to be insufficient, ranging from 0.57 to 0.63. To address this issue, different labelling systems and data pre-processing techniques were explored in a cross-sectional study. Nonetheless, the inter-assessor agreement remained challenging, with an average kappa value of 0.37 (SD = 0.16), and high-accuracy algorithmic predictions remained elusive, with an average accuracy of 56.1 (SD =16.58). Finally, the algorithms' performance was tested with high-confidence labels, which consisted of only scores 0 and 3 of the AHDB system. This testing resulted in good classification accuracy (0.82), specificity (0.79), and sensitivity (0.85). This led to the proposal of a new approach to producing labels, testing vantage point changes, and improving the performance of machine learning models (average accuracy = 0.7 & SD = 0.17, average sensitivity = 0.68 & SD = 0.27, average specificity = 0.75 & SD = 0.17). The research identified a challenge in creating high-confidence diagnostic labels for supervised machine learning-based algorithms to automate the detection and classification of lameness in dairy cows. As a result, the original goals were partially overridden, with the focus shifted to creating reliable labels that would perform well with radar data and machine learning. This point was considered necessary for smooth system development and process automation. Nevertheless, we managed to quantify the performance of the micro-Doppler radar system, partially develop the supervised machine learning algorithms, compare levels of agreement among multiple assessors, evaluate the assessment tools, assess the mobility evaluation process and gather a valuable data set which can be used as a foundation for subsequent studies. Finally, the thesis suggests changes in the assessment process to improve the prediction accuracy of algorithms based on supervised machine learning with radar data.