Mahantappa Sangappa Sankanur
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mahantappa Sangappa Sankanur has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Fatimah Khalid
Universiti Putra Malaysia
2 shared publications - Mas Rina Mustaffa
Universiti Putra Malaysia
2 shared publications - Rana Ranjeet Singh
Kamdhenu University
2 shared publications - Timur Rampalsingh Ahlawat
Navsari Agricultural University
2 shared publications - Anuradha Agrawal
National Agricultural Higher Education Project (NAHEP), Krishi Anusandhan Bhawan-II, Indian Council of Agricultural Research (ICAR), New Delhi, India
1 shared publication Fathehah Farid Rozaidi
Universiti Putra Malaysia
1 shared publication
Latest Publications
- 9 September 2026
- Published by Universiti Putra Malaysia in Journal of Intelligent Media Computing
- Vol. 1 (1) , 157
Leveraging Visual Data for Accurate Body Weight Prediction in Surti Goats
Routine weighing of goats requires suitable equipment and often involves animal handling or restraint, making repeated monitoring difficult under farm conditions. This study estimated Surti goat body weight by combining image-derived morphometric measurements (heart girth and body length extracted from two-dimensional images) with age and sex metadata. Mask Region-based Convolutional Neural Network (Mask R-CNN) was used to separate each goat from the image background before the relevant visual measurements were extracted. 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. XGBoost, Random Forest and CatBoost were trained using the combined feature set, and their outputs were integrated through a weighted ensemble with weights of 0.5, 0.3, and 0.2 assigned to XGBoost, CatBoost, and Random Forest, respectively. On the validation set (37 goats), XGBoost achieved an 𝑅2 of 0.8514 (RMSE = 4.26 kg, MAE = 3.47 kg, MAPE = 16.27%). On the completely unseen test set (37 goats), XGBoost achieved the highest predictive performance with an 𝑅2 of 0.7955 (MSE = 29.15 kg², RMSE = 5.40 kg, MAE = 4.01 kg, and MAPE = 16.38%), while the weighted ensemble recorded an 𝑅2 of 0.7713 (MSE = 32.60 kg², RMSE = 5.71 kg, MAE = 4.13 kg, and MAPE = 16.60%). The findings show that combining visual and morphometric information under a strict group-based evaluation provides reliable, non-contact body weight estimation on unseen livestock.
- 9 September 2026
- Published by Universiti Putra Malaysia in Journal of Intelligent Media Computing
- Vol. 1 (1) , 157
Leveraging Visual Data for Accurate Body Weight Prediction in Surti Goats
Routine weighing of goats requires suitable equipment and often involves animal handling or restraint, making repeated monitoring difficult under farm conditions. This study estimated Surti goat body weight by combining image-derived morphometric measurements (heart girth and body length extracted from two-dimensional images) with age and sex metadata. Mask Region-based Convolutional Neural Network (Mask R-CNN) was used to separate each goat from the image background before the relevant visual measurements were extracted. 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. XGBoost, Random Forest and CatBoost were trained using the combined feature set, and their outputs were integrated through a weighted ensemble with weights of 0.5, 0.3, and 0.2 assigned to XGBoost, CatBoost, and Random Forest, respectively. On the validation set (37 goats), XGBoost achieved an 𝑅2 of 0.8514 (RMSE = 4.26 kg, MAE = 3.47 kg, MAPE = 16.27%). On the completely unseen test set (37 goats), XGBoost achieved the highest predictive performance with an 𝑅2 of 0.7955 (MSE = 29.15 kg², RMSE = 5.40 kg, MAE = 4.01 kg, and MAPE = 16.38%), while the weighted ensemble recorded an 𝑅2 of 0.7713 (MSE = 32.60 kg², RMSE = 5.71 kg, MAE = 4.13 kg, and MAPE = 16.60%). The findings show that combining visual and morphometric information under a strict group-based evaluation provides reliable, non-contact body weight estimation on unseen livestock.
- 9 September 2026
- Published by Universiti Putra Malaysia in Journal of Intelligent Media Computing
- Vol. 1 (1) , 42
Comparative Evaluation of Image Classification Architectures for Surti Buffalo Muzzle Recognition With CAM-Based Explainability Analysis
Advancements in image classification architecture have greatly improved image-based animal identification. However, model comparisons are frequently restricted to classification accuracy. By comparing multiple pre-trained CNN-based and Vision Transformer architectures for muzzle image categorisation utilising predicted performance, computational efficiency, and visual interpretability, this work expands the evaluation. The accuracy, storage capacity, training time, and interpretability of pre-trained CNN and Vision Transformer architectures for muzzle image classification are evaluated in this work using four CAM variants: Grad-CAM, Grad-CAM++, Score-CAM, and XGrad-CAM. According to the results, MobileNetV2 strikes a fair balance, obtaining 96.63% accuracy with just 30.08 MB of storage and 3.80 hours of training time per fold, while DenseNet121 obtains the maximum accuracy at 97.02% with smallest variance (±0.17%). Grad-CAM++ produced the most instructive activation maps because it highlighted relevant muzzle regions more precisely than the other CAM approaches. Consistent activation maps were found to be associated with stable classification performance. However, Vision Transformer models had significant volatility in explanation quality and prediction accuracy (±2.28%). Overall, explainability is important for improving trust in these architectures because it provides insight into the reasons behind model predictions. Therefore, evaluation of image classification architectures should consider not only accuracy and efficiency but also how reliably the model's decisions can be interpreted.
- 9 September 2026
- Published by Universiti Putra Malaysia in Journal of Intelligent Media Computing
- Vol. 1 (1) , 42
Comparative Evaluation of Image Classification Architectures for Surti Buffalo Muzzle Recognition With CAM-Based Explainability Analysis
Advancements in image classification architecture have greatly improved image-based animal identification. However, model comparisons are frequently restricted to classification accuracy. By comparing multiple pre-trained CNN-based and Vision Transformer architectures for muzzle image categorisation utilising predicted performance, computational efficiency, and visual interpretability, this work expands the evaluation. The accuracy, storage capacity, training time, and interpretability of pre-trained CNN and Vision Transformer architectures for muzzle image classification are evaluated in this work using four CAM variants: Grad-CAM, Grad-CAM++, Score-CAM, and XGrad-CAM. According to the results, MobileNetV2 strikes a fair balance, obtaining 96.63% accuracy with just 30.08 MB of storage and 3.80 hours of training time per fold, while DenseNet121 obtains the maximum accuracy at 97.02% with smallest variance (±0.17%). Grad-CAM++ produced the most instructive activation maps because it highlighted relevant muzzle regions more precisely than the other CAM approaches. Consistent activation maps were found to be associated with stable classification performance. However, Vision Transformer models had significant volatility in explanation quality and prediction accuracy (±2.28%). Overall, explainability is important for improving trust in these architectures because it provides insight into the reasons behind model predictions. Therefore, evaluation of image classification architectures should consider not only accuracy and efficiency but also how reliably the model's decisions can be interpreted.