Work

HEART-Net: An ultra-lightweight multimodal neural network for on-device cattle live-weight estimation and precision phytogenic dosing

Year: 2026

Type: article

Abstract: Accurate live-weight estimation is fundamental for precision livestock informatics, yet mechanical weighing remains logistically complex and stressful for animals in decentralized farming systems. However, current deep-learning models rely on heavy architectures that exceed the memory and latency constraints of field-level mobile hardware. We hypothesized that a lightweight multimodal network could match the accuracy of heavy baselines while satisfying strict edge constraints. To test this, we introduce HEART-Net (Hybrid Edge-Adapted Regression Translator Network), an ultra-lightweight framework optimized for on-device phenotyping. Using 60,000 lateral RGB images from 30 indigenous Pabna cattle, we applied an animal-based stratified split (21 training, 4 validation, 5 testing animals) to prevent data leakage. Although EfficientNetV2B0 achieved the lowest mean absolute error (5.58 kg), HEART-Net operated with 0.35 MB of storage and 51.48 ± 1.77 ms inference latency while maintaining statistically equivalent accuracy to MobileNetV3-Small (MAE = 7.52 kg, p = 0.090). The model showed low systematic bias (+0.15 kg) and strong concordance across the 213 to 412 kg weight range. Deployed in the Smart-Herbo Progressive Web Application with metabolic scaling (W 0.75 ) for automated dosing, the system achieved a System Usability Scale score of 82.25.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2026.102587

Authors: Ashif Ahmed Shuvo, Sakibul Hasan, Abir Anjum, Asaduzzaman Shahed, Mohammad Al-Mamun

Institution: Bangladesh Agricultural University

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $1,770