Comparing machine learning, principal component analysis and conventional regression for body weight prediction in Murrah buffaloes (Bubalus bubalis)

Context Body weight guides feeding, breeding, health and marketing decisions in buffaloes, but weighing large buffaloes on a scale is often impractical on smallholder farms, so weight is estimated from body measurements taken with a tape. Principal component analysis (PCA) combined with machine learning has been proposed for this task, though whether it predicts better than the simple methods already in use has not been tested. Aims To benchmark PCA-based machine learning against chest-girth regression, multiple linear regression and published buffalo equations for estimating body weight in Murrah heifers and adults. Methods Eleven morphometric traits and body weight were recorded on 62 heifers and 127 adult Murrah buffaloes. PCA was applied to the eleven traits with body weight excluded. Eight machine learning algorithms, together with multivariate adaptive regression splines, were trained on the principal components and on the raw traits using five-fold cross-validation repeated over resamples, with all transforms fitted inside each training fold and hyperparameters tuned within a nested loop. Model differences were tested with Friedman and Wilcoxon signed-rank tests. Key results Chest girth was most correlated with body weight (r = 0.93 heifers, 0.86 adults). The first principal component explained 76.8% (heifers) and 61.3% (adults) of trait variation. In heifers, multiple linear regression (R2 = 0.94) was not significantly bettered by any model, and chest girth alone performed comparably. In adults, multiple regression reached R2 = 0.76, and after hyperparameter tuning, a tree ensemble and multivariate adaptive regression splines gave a small but consistent gain (R2 ≈ 0.79). Published equations transferred poorly without recalibration. Conclusions Conventional regression is a strong, low-cost predictor of buffalo body weight; machine learning gives no gain in heifers and only a small gain in adults, and only after tuning. Implications For routine on-farm use, a single chest-girth measurement or a simple regression remains practical and accurate; the small adult gain from machine learning was realised only after task-specific tuning and remained modest even then, so heavier analytical methods should be adopted only where they demonstrably improve prediction.