Early Pregnancy Detection in Sheep Using
Thermal Imaging and Explainable Artificial Intelligence:
A Joint Industry–Academic Research Initiative

This proposal seeks grant funding for a three-year collaborative research program to develop and validate a novel, non-invasive system for the early detection of pregnancy in sheep using thermal imaging cameras and Explainable Artificial Intelligence (XAI). The project brings together the AI innovation capability of an industry partner specialising in Generative AI, Explainable AI, and multi-agent systems — with the world-class livestock research infrastructure and decades of longitudinal animal datasets held by SARDI Livestock Sciences (South Australian Research and Development Institute).

Pregnancy detection in sheep remains one of the most labour-intensive and economically critical tasks in Australian livestock management. Current methods — including ultrasound scanning, blood progesterone assays, and visual observation — are expensive, stressful to animals, and often performed too late to optimise reproductive outcomes. This project will deliver a real-time, paddock-deployable system that detects early pregnancy (Days 15–30 post-mating) from thermal infrared signatures correlated with hormonal and physiological changes, interpreted by transparent XAI models that provide veterinary-grade confidence scores and clinical reasoning.

The outcomes will directly support Australia's $4.5 billion sheep meat and $3.2 billion wool industries by reducing reproductive wastage, enabling precision flock management, and accelerating the uptake of non-invasive livestock monitoring technologies across extensive production systems.

2.1 The Australian Sheep Industry Challenge

Australia manages approximately 65 million sheep across extensive grazing systems, predominantly in South Australia, New South Wales, Victoria, and Western Australia. Reproductive efficiency is the single greatest driver of flock profitability: an improvement of 10% in lambing rates across the national flock represents an estimated economic benefit exceeding AUD $280 million annually (Meat & Livestock Australia, 2024). Yet, reproductive failure — including undetected pregnancy loss, non-pregnant ewes maintained through winter, and late identification of reproductive status — remains endemic.

Early and accurate pregnancy diagnosis (before Day 30 of gestation) enables producers to: (i) segregate pregnant and non-pregnant ewes for differential management; (ii) re-join empty ewes in time for the next reproductive cycle; (iii) allocate high-nutrition feed supplements to confirmed pregnant ewes; and (iv) predict lambing dates with sufficient precision for labour scheduling and welfare monitoring.

2.2 Limitations of Current Technologies

Table 1. Comparison of current pregnancy detection methods with the proposed system.

2.3 Thermal Imaging as a Biological Signal

Thermography captures infrared radiation emitted by body surfaces, providing a non-contact, non-invasive proxy for physiological state. In pregnant ewes, progesterone-driven vasodilation and metabolic changes produce measurable thermal asymmetries — particularly in the perineal, mammary, and abdominal regions — as early as 15 days post-conception (Ciliberti et al., 2022; Teke et al., 2014). These thermal signatures are subtle, multivariate, and confounded by environmental variables (ambient temperature, fleece depth, wind), making manual interpretation unreliable and necessitating machine learning approaches.

2.4 The Role of Explainable AI

Standard "black-box" deep learning models are poorly adopted in livestock industries because producers and veterinarians require interpretable confidence scores, not opaque binary outputs. Explainable AI (XAI) — specifically SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-Agnostic Explanations), and attention-based saliency mapping — enables the system to communicate which thermal regions and features drove each prediction, aligned with known physiology. The industry partner's proprietary XAI framework, developed for clinical decision support applications, will be adapted and validated for this livestock application.

This project addresses the following overarching research question:

Primary Aims

4.1 Study Design Overview

The project employs a three-phase longitudinal cohort design across three consecutive mating seasons (2026, 2027, 2028) at SARDI research stations in South Australia. Phase 1 focuses on data collection and model development; Phase 2 on model validation and refinement; Phase 3 on field-scale demonstration and commercial readiness assessment.

4.2 Animal Cohort and Site Description

A total of 1,800 Merino and Merino-cross ewes will be recruited across SARDI's Turretfield Research Centre (Rosedale, SA) and Struan Research Centre (Naracoorte, SA). Animals will be synchronised using the CIDR-eCG protocol to create defined mating cohorts. Pregnancy ground truth will be established by qualified SARDI veterinarians using transabdominal B-mode ultrasound at Day 28 and Day 42 post-mating, with progesterone assay confirmation for ambiguous cases.

4.3 Thermal Imaging Protocol

A FLIR T860 research-grade thermal camera (resolution: 464×348 pixels, sensitivity: <30mK NETD) will be mounted on a portable gantry system enabling standardised capture from dorsal, lateral, and perineal angles during routine yarding. Images will be captured at Days 7, 14, 21, 28, and 35 post-mating, between 06:00–09:00 ACST to minimise solar thermal interference. RFID ear-tag readers will ensure automatic individual animal identification and image-record linkage.

4.4 AI/XAI Model Architecture

Figure 1. XAI model pipeline architecture.

4.5 Statistical Analysis Framework

Model performance will be evaluated using Area Under the ROC Curve (AUC-ROC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at each detection timepoint. Bayesian hierarchical models will be used to partition variance attributable to site, season, breed, and body condition score. Target performance thresholds: AUC-ROC ≥0.90, Sensitivity ≥88%, Specificity ≥85% at Day 21 post-mating.

Prior to this grant application, the industry partner (Tech Adaptive Pandit) has conducted preliminary investigative work to validate the technical feasibility of thermal XAI-based pregnancy detection in sheep. A working proof-of-concept demonstration has been developed and is publicly accessible, showcasing the thermal imaging pipeline, XAI inference engine, and producer-facing output interface.

Key Preliminary Findings

Note: Preliminary work was conducted using simulated and publicly available livestock thermal datasets. Full field validation with SARDI-collected Merino ewe data is the primary objective of this grant.

5.1 Roles and Responsibilities Matrix

●● = Lead responsibility, ● = Contributing role

6.1 Detailed Budget Summary

8.1 Scientific Innovation

This project represents the first systematic application of Explainable AI to thermal pregnancy detection in sheep under Australian extensive production conditions. The environmental correction model (Aim 3) specifically addresses the primary technical barrier that has prevented thermal imaging from transitioning from laboratory proof-of-concept to field-reliable technology. The integration of longitudinal temporal modelling (LSTM) with thermal data represents a methodological advance over existing single-timepoint detection studies.

8.2 National Economic Benefit

8.3 Animal Welfare Benefit

The proposed system is entirely non-invasive and non-contact, eliminating the physiological stress associated with yarding for manual palpation, transrectal ultrasound, or blood sampling. This aligns with Australia's National Animal Welfare Standards and the Australian Government's commitment to the Five Domains of animal welfare. Early identification of non-pregnant ewes also enables prompt re-joining, reducing the proportion of ewes experiencing repeated failed pregnancies.

8.4 Technology Commercialisation Pathway

The industry partner holds existing IP in XAI platforms and will commercialise the sheep pregnancy detection module as a standalone precision livestock product. A licensing agreement with SARDI for validated algorithm use in commercial products will be negotiated during Phase 3. Target markets include Australian livestock enterprises, New Zealand sheep producers, UK upland farming systems, and Argentinian Merino producers, with a projected commercial launch by Q3 2029.