Extracting heart girth and body length from 2D images with age and sex, XGBoost predicts body weight in unseen Surti goats at R²=0.7955

The study proposes and evaluates a non-contact body weight estimation pipeline: Mask R-CNN first separates each goat from the image background, then heart girth and body length are extracted from two-dimensional images and merged with age and sex metadata into a combined feature set. Relative to manual weighing that relies on suitable equipment and often involves animal handling or restraint, this work bases weight estimation on image-derivable morphometric information plus basic metadata, addressing the abstract's stated difficulty of repeated monitoring under farm conditions. Evidence comes from the pipeline described in the abstract: Mask R-CNN segmentation, image-derived morphometric measurements, combination with age and sex metadata, and training of three gradient-boosting/ensemble models on that feature set.

On a dataset partitioned at the Animal ID level, XGBoost achieved the highest predictive performance on both the validation set and the completely unseen test set, with test-set R²=0.7955, RMSE=5.40 kg, MAE=4.01 kg and MAPE=16.38%. The abstract states that dataset partitioning was conducted strictly at the Animal ID level (group-wise split) to prevent data leakage and evaluate true model generalisation on unseen animals, so the test-set result reflects prediction on unseen individuals rather than fitting to the same animals. The validation set contains 37 goats and the test set 37 goats, with R², MSE, RMSE, MAE and MAPE reported; XGBoost outperformed the weighted ensemble on both sets.

A weighted ensemble assigning 0.5, 0.3 and 0.2 to XGBoost, CatBoost and Random Forest respectively recorded R²=0.7713, RMSE=5.71 kg, MAE=4.13 kg and MAPE=16.60% on the unseen test set, below XGBoost alone. This provides a concrete comparison: under this feature set and group-based evaluation setting, the ensemble did not exceed the best single model, offering a comparable baseline for subsequent modeling choices. The ensemble and the single models use the same feature set and the same test set (37 goats), so the metrics are directly comparable.

The study concludes that combining visual and morphometric information under strict group-based evaluation provides reliable, non-contact body weight estimation on unseen livestock. This positions the method's value on two points, non-contact operation and generalisation to unseen animals, pointing to a feasible path for repeated monitoring under farm conditions. The conclusion rests on the validation and test metrics reported in the abstract, with the test set comprising 37 completely unseen goats.

The result targets settings where goats need repeated body weight monitoring under farm conditions and where two-dimensional images plus age and sex records can be obtained; the abstract indicates the evaluation concerns Surti goats, with 37 goats in the validation set and 37 in the test set, the latter completely unseen, so its direct scope is non-contact estimation for this breed and this data collection approach.

The abstract does not describe the imaging device, distance, angle or lighting conditions, nor does it report Mask R-CNN segmentation accuracy, so how extraction error in heart girth and body length affects final weight prediction remains an open question; the abstract also omits total dataset size, age and sex distributions, and comparisons with other body weight estimation methods, which would inform how the pipeline performs across different populations.