Frequent biomass measurement is a key activity for optimal perennial ryegrass (Lolium perenne) management in intensive forage-based dairy operations. Due to the necessary high-frequency (i.e., weekly or monthly) pasture monitoring and continuous trend of larger dairy farms, such activity is perceived as an operational bottleneck. Consequently, substantial effort is directed to the development of accurate and automated technological solutions for biomass assessment. The popularization of unmanned aerial vehicles (UAVs) combined with multispectral cameras should allow for an optimal observational system able to deploy machine learning algorithms for near real-time biomass dry-matter (DM) mapping. For successful operation, these systems should deliver radiometrically accurate orthomosaics and robust models able to generalize across different periods. Nevertheless, the accuracy of radiometric calibration and generalization ability of these models is seldom evaluated. Also, such pipelines should require minimum processing power and allow for fast deployment. This study has established a two-year experiment comparing reflectance measurements between a handheld spectrometer and a commercial multispectral UAV camera. Different algorithms based on regression-tree architecture were contrasted regarding accuracy, speed, and model size. Model performances were validated, providing error-metrics for baseline accuracy and temporal validation. The results have shown that the standard procedure for multispectral imagery radiometric calibration is sub-optimal, requiring further post-processing and presenting low correlation with handheld measurements across spectral bands and dates. Nevertheless, after post-calibration, the use of spectral imagery has presented better baseline error than the point-based sensors, respectively displaying an average of 397.3 and 464.2 kg DM/ha when employed alongside the best performing algorithm (i.e., Cubist). When trained and validated across different years, model performance was largely reduced and deemed unfit for operational purposes. The Cubist/M5 family of algorithms have exhibited advantageous characteristics such as compact model structure, allowing for a higher level of model interpretability, while displaying a smaller size and faster deployment than the Random Forest, Boosted, and Bagged Regression Trees algorithms.