估算群體飼養下生長期豬隻之個別飼料採食量

Abstract

飼料採食量在豬隻生產過程中相當重要，提供豬隻生長穩定且良好的營養來源，使能維持穩定且健康的成長，豬場支付的生產成本中飼料費用佔比超過60%，有效控管豬隻採食飼料的多寡可以幫助經營者降低生產成本，因此種豬的飼料效率選拔經常是重點之一。台灣的種豬檢定常以飼養每欄一至二頭，或以一欄多頭的平均採食量來進行飼料採食量量測，然而豬隻的採食量在群體和單頭的飼養方式之下會有一定程度的差異。獨立餵飼站 (Feeding Station) 可以在群養欄中測量每隻豬的個別飼料採食量，卻因其單價高，且對欄位大小、豬隻頭數與周邊設備具有限制，降低台灣豬場的使用意願。本研究即為在低程度變動欄位及給飼設備下，估算群養欄豬隻的個別飼料採食量。 於2025年10月至2026年3月之間，使用54頭Landrace × Yorkshire生長期閹公豬，分成兩批次共四個欄位，各欄位分別飼養13、13、15、13頭，起始與結束體重分別為29.28 ± 4.76、73.17 ± 9.10公斤。每兩週測量一次豬隻的體重 (BW)和腰眼面積 (Loin muscle area)，每十分鐘收集欄位溫度、濕度，並計算溫溼度指標(Temperature-Humidity Index, THI)，從而產生每欄的整日均溫 (mean Temp)、平均白天期間 (6至18時) 均溫 (daytime Temp) 與平均溫溼度指標 (daytime THI)。將商用的飼料桶進行設計改造可自動秤量並記錄每天的欄位總飼料採食量 (pen feed consumption, PFC)。以線性迴歸法估算每隻豬在每試驗日的體重，以二次函數計算每隻豬在每試驗日的腰眼面積，並計算出每日的代謝體重 (BW0.6) 、體增重 (WG) 、腰眼面積 (LMA) 與腰眼面積增加量 (LMA gain)。以複迴歸法將前述所有變因 (Full model) 估算每日每頭之飼料採食量。使用擬牛頓法 (Quasi-Newton Methods) 調整各變因的估算參數，使每日估算的同一欄位豬隻飼料採食量加總與實際欄位總飼料採食量的差值最小，判斷各變因的合適參數。以一天及合併二至五天為五種PFC、WG與LMA gain資料的數據處理方式，各自計算BW0.6、LMA、mean Temp、daytime Temp與daytime THI等資料的平均值，作為數據處理方式的變因資料，並以Full model計算每頭豬在不同天數之數據處理方式的R2，挑選具較高R2的數據處理方式進行後續分析。後續使用複迴歸之backward elimination選出向後消去模型的變數組合，另外根據調整R2 (adj R2) 與赤池資訊量準則 (AIC)挑選出最適化模型的變數組合。將兩模型與Full model進行複迴歸機器學習驗證：挑選任意三欄位數據作為訓練集，剩餘一欄位數據作為驗證集，以驗證各種模型是否有解釋度不足或過度擬合的現象。 資料分析結果顯示，以每四天合併之數據處理方式具最高R2=0.851，故後續分析皆使用四天合併數據。Full model的adj R2=0.829，AIC=256.58；Backward selection決定之自動化模型 (BW0.6、WG與LMA gain) 的adj R2=0.819，AIC=256.25；最適化模型 (BW0.6、WG、mean LMA與mean Temp) 之adj R2=0.823，AIC=256.06。其中，Full model具有最高之adj R2，但其AIC最高，可能有過度擬合問題；最適化模型比起Backward selection決定的向後消去模型有較好的adj R2與AIC，較能準確估算個別豬隻飼料採食量。複迴歸機器學習驗證的結果顯示，Full model在訓練集的adj R2 介於0.584~0.661，驗證集的adj R2 竟落在-0.637~-0.267之間；向後消去模型在訓練集的adj R2 則為0.771~0.873，其驗證集的adj R2廣布於-0.236~0.836之間；最適化模型訓練集的adj R2介於0.788~0.862，在驗證集的adj R2則在0.135~0.776。結果顯示挑選最適化模型有較高的解釋度及較低的過度擬合，因此相對其他兩個模型，此最適化模型具有較佳的個別豬隻飼料採食量之預測能力，代表在進行複迴歸模型選擇時，有人為協助選擇模型能夠增加模型的解釋度及降低模型因子的雜訊干擾。

Feed intake is an important factor in pig production because it provides the nutrients and energy required to support stable growth and health. Feed costs account for more than 60% of total pig production costs; therefore, effective control of feed intake can help producers reduce production expenses. Consequently, feed efficiency is commonly included as an important selection trait in breeding programs. In Taiwan, feed intake measurement in pig performance testing is either by housing one or two pigs per pen or by the averaging feed intake from multiple pigs housed in the same pen. However, feed intake may differ between individually housed and group-housed pigs. Automated feeding stations can measure individual pig’s feed intake in group-housed pigs, but cost, requirements for pen size, stocking density, and supporting equipment have limited their adoption by Taiwanese pig farms. Therefore, this study aimed to estimate individual feed intake in group-housed pigs with minimal modifications to existing pens and feeding equipment. From October 2025 to March 2026, 54 Landrace × Yorkshire growing barrows were assigned to four batches and housed in two pens containing 13, 13, 15, and 13 pigs, respectively. Initial and final body weights were 29.28 ± 4.76 and 73.17 ± 9.10 kg, respectively. Body weight (BW) and loin muscle area (LMA) were measured every 2 weeks. Pen temperature and relative humidity were recorded every 10 min, and the temperature-humidity index (THI) was calculated. Daily mean temperature (mean Temp), daytime mean temperature from 0600 to 1800 h (daytime Temp), and daytime THI were calculated for each pen. A commercial feeder was modified to automatically weigh and record daily pen feed consumption (PFC).Linear regression was used to estimate BW on each experimental day for each pig, whereas a quadratic function was used to estimate daily LMA. Daily metabolic body weight (BW0.6), weight gain (WG), LMA, and LMA gain were calculated. A multiple regression full model including all variables was used to estimate daily individual feed intake. A quasi-Newton optimization method was applied to adjust the parameter estimates so that the difference between the summed estimated individual feed intake within each pen and the observed PFC was minimized. Five data-processing methods were evaluated, including daily data and data aggregated over 2, 3, 4, or 5 days. For each aggregation interval, BW0.6, LMA, mean Temp, daytime Temp, and daytime THI were averaged, whereas PFC, WG, and LMA gain were calculated according to the corresponding interval. The coefficient of determination (R²) of the full model was compared among data-processing methods, and the method with the highest R² was selected for subsequent analyses. Backward selection was then used to derive a reduced model. An optimized model was additionally selected based on adjusted R² (adj R²) and Akaike’s information criterion (AIC). The full model, backward-selection model, and optimized model were evaluated using leave-one-pen-out validation, in which data from three pens were used for model training and data from the remaining pen were used for validation. Results showed that aggregation over 4 days produced the highest R² (0.851); therefore, all subsequent analyses were conducted using 4-day aggregated data. The full model had an adj R² of 0.829 and an AIC of 256.58. The backward-selection model, including BW^0.6, WG, and LMA gain, had an adj R² of 0.819 and an AIC of 256.25. The optimized model, including BW0.6, WG, mean LMA, and mean Temp, had an adj R² of 0.823 and an AIC of 256.06. Although the full model had the highest adj R², it also had the highest AIC, suggesting possible overfitting. Compared with the backward-selection model, the optimized model showed a higher adj R² and lower AIC, indicating better accuracy in estimating individual feed intake. During validation, the adj R² of the full model ranged from 0.584 to 0.661 in the training sets but declined to between −0.637 and −0.267 in the validation sets. For the backward-selection model, training adj R² values ranged from 0.771 to 0.873, whereas validation adj R² values ranged from −0.236 to 0.836. The optimized model showed training adj R² values ranging from 0.788 to 0.862 and validation adj R² values ranging from 0.135 to 0.776. These results indicate that the optimized model achieved greater explanatory power and lower overfitting than the other models. Thus, the optimized model showed superior predictive ability for estimating individual feed intake in group-housed pigs. The findings also suggest that biologically informed model selection can improve model performance and reduce interference from unnecessary variables in multiple regression models.