Ovulation-Anchored Evaluation of IMU-Derived Activity and Posture-Related Behavioral Changes Across Natural Estrus Phases in Dairy Cattle

1. Introduction

Accurate estrus detection is a critical component of reproductive management in dairy herds because timely identification of cows in estrus supports appropriate timing of artificial insemination and improves reproductive efficiency [1,2]. Failure to detect estrus at the appropriate time can reduce conception rates, prolong calving intervals, and increase economic losses associated with reduced fertility performance [3,4]. Estrus in dairy cows is characterized by physiological and behavioral changes, including standing to be mounted, increased locomotor activity, mounting-related behavior, restlessness, and altered resting patterns [2,5]. Although visual observation remains widely used, its effectiveness is limited by labor requirements, observation frequency, housing conditions, short estrus duration, and variation in behavioral expression among cows [1,6]. Consequently, automated monitoring technologies have become increasingly important under practical farm conditions because they enable continuous detection of behavioral changes that may be missed by intermittent visual observation [6,7,8].

Precision dairy technologies, including pedometers, collars, ear tags, and wearable sensors, have been widely used to monitor activity and reproductive status in dairy cattle [2,9,10]. However, many estrus-monitoring studies have relied primarily on general activity measures or accelerometer-derived movement intensity to distinguish estrus from non-estrus periods, which may overlook coordinated changes in locomotion, mounting-related activity, social interaction, and resting behavior during [11,12]. Inertial measurement unit (IMU) systems, which combine accelerometer and gyroscope sensors, provide more comprehensive behavioral information by capturing both linear acceleration and rotational movement [13]. Although IMU-based systems have been used to classify cattle behaviors such as lying, standing, walking, feeding, and ruminating [13,14,15], limited information is available on how these sensor-derived variables change across the biological progression of natural estrus [16].

This limitation is important because natural estrus is a dynamic process rather than a single uniform event. Behavioral expression changes progressively across the peri-estrus period, and the relationship between behavioral estrus and ovulation timing varies among cows [2,5,17]. Therefore, evaluating estrus only as “estrus” versus “non-estrus” may obscure biologically meaningful temporal patterns in activity and posture-related behavior, particularly across the transition from baseline behavior to pre-estrus, standing estrus, peri-ovulation, and post-ovulation recovery [1,18]. Accordingly, a phase-based approach can provide a more detailed understanding of how sensor-derived behavioral variables change throughout natural estrus.

Despite increasing interest in sensor-based estrus monitoring, limited evidence is available on phase-specific changes in IMU-derived activity and posture-related variables during natural estrus anchored to ultrasound-confirmed ovulation. Previous studies have mainly relied on general activity measures, accelerometer-based indices, or broad estrus versus non-estrus comparisons, with limited integration of dynamic acceleration, rotational movement, and posture-related behavior within a biologically defined estrus-phase framework [7,19,20]. This represents an important knowledge gap because activity intensity alone does not fully describe estrus-related behavioral organization.

To address this gap, the present study characterized IMU-derived activity features and posture-related behaviors across six biologically defined phases of natural estrus: normal, pre-estrus, standing estrus, late estrus, peri-ovulation, and early post-ovulation. These phases were defined using video-derived behavioral signs together with ultrasound-confirmed ovulation to provide biologically meaningful temporal classification of estrus progression. Therefore, this study aimed to characterize changes in IMU-derived activity features and posture-related behaviors across estrus phases in dairy cows and to determine whether these features differed among phases. The study addressed the following research question: how do IMU-derived activity features and posture-related behaviors differ across biologically defined phases of natural estrus in dairy cows?

2. Materials and Methods

This study was approved by the Institutional Animal Care and Use Committee (IACUC) of the Faculty of Veterinary Science, Chulalongkorn University, Thailand (Protocol No. 2031047), and was conducted in accordance with institutional regulations and the Ethical Principles and Guidelines for the Use of Animals for Scientific Purposes established by the National Research Council of Thailand.

2.1. Study Design and Animal Population

This observational study was conducted to characterize changes in inertial measurement unit (IMU)-derived activity features and posture-related behaviors across biologically defined phases of natural estrus in dairy cows. The study was performed under routine farm management conditions to capture naturally occurring estrus-associated behavioral dynamics without disrupting normal herd management.

Data collection was conducted from August to October 2021 in the loose barn of the Farm Animal Hospital, Faculty of Veterinary Science, Chulalongkorn University, Nakhon Pathom, Thailand. The animals were dairy cows housed in a 30 × 15 m concrete-floored enclosure designed to allow free movement and expression of natural behaviors. Cows were fed a total mixed ration, on a dry matter basis, comprising 80% roughage and 20% concentrate at 2.5% of body weight, with feed provided twice daily at 09:00 and 14:00 h. Clean drinking water was available ad libitum throughout the study period.

After data screening, the final analytical dataset consisted of five dairy cows and eleven natural estrus cycles. Estrus cycles were included when complete sensor records, video-based behavioral observations, and confirmation of ovulation were available. Each cow–cycle combination was considered the biological unit of analysis.

During the monitoring period, neck-mounted inertial measurement unit (IMU) devices continuously recorded movement activity, while video observations and reproductive examinations were used to obtain reference data. All procedures were observational and interfere with routine animal handling, feeding, or reproductive management practices.

2.2. Sensor Configuration and Placement

Cow activity was measured using a custom-built neck-mounted IMU device equipped with a tri-axial accelerometer and tri-axial gyroscope sensor (MPU-6050, InvenSense Inc., California, USA). The accelerometer quantified linear acceleration along three orthogonal axes, whereas the gyroscope captured angular velocity around the same axes. Together, these sensor components provided complementary information about translational and rotational motion.

The sensor device was configured to continuously record movement throughout the monitoring period. Raw acceleration and angular velocity signals were collected along the X-, Y-, and Z-axes and stored with corresponding timestamp information. Sensor data were recorded as mean values over consecutive 10-second intervals, corresponding to an effective sampling frequency of 0.1 Hz.

Each IMU device was enclosed in a protective housing and securely attached to the right side of the neck collar of each cow to acquire continuously head and neck movement data. The sensor axes were oriented relative to the cow’s body as follows: the X-axis represented the forward–backward direction, the Y-axis the vertical direction, and the Z-axis the lateral direction. This placement was selected because it was practical under farm conditions and appropriate for detecting posture and activity patterns associated with resting, standing, walking, and estrus-associated behavioral changes.

Accelerometer and gyroscope signals were stored with corresponding timestamp information, enabling synchronization with video-based behavioral observations and reproductive reference data. The resulting time-aligned dataset was subsequently used for feature extraction and phase-level analysis.

2.3. Conceptual Framework of the Study

The conceptual framework of this study was designed to integrate animal monitoring, reference data generation, sensor data preparation, feature extraction, estrus phase assignment, data summarization, and statistical analysis into a structured analytical workflow. This framework enabled IMU-derived movement features and posture-related behavioral variables to be interpreted in relation to biologically defined phases of natural estrus.

As illustrated in Figure 1, the workflow began with continuous sensor-based monitoring and video observation, followed by the preparation of synchronized sensor, behavioral, and reproductive reference datasets. The processed dataset was subsequently used for feature extraction, estrus phase classification, cow-cycle-phase-level summarization, and phase-level statistical analyses.

2.4. Data Acquisition and Ground-Truth Observation

Reference behavioral data were obtained from continuous closed-circuit television (CCTV) footage. Two trained observers reviewed the footage and classified cow behaviors according to a predefined ethogram. Before formal annotation, both observers were trained using a pilot set of video records to harmonize behavioral definitions and annotation criteria. When ambiguous behaviors or disagreements occurred, the records were rechecked and resolved through discussion and consensus. This procedure was used to reduce observer-related variability in the ground-truth behavioral labels.

The main behavioral categories analyzed were lying, standing, and walking. These categories represented biologically interpretable posture and activity states relevant to estrus-associated changes.

Estrus-related behavioral signs were identified from continuous video records using primary and secondary indicators adapted from the visual estrus scoring system [21]. Standing heat was assigned the highest score of 100 points and was used as the principal indicator of standing estrus. Secondary indicators were scored according to behavioral relevance, including mounting the head side of another cow (45 points), mounting, or attempting to mount another cow (35 points), chin resting (15 points), sniffing the vagina or vulvar region of another cow (10 points), being mounted without standing (10 points), restlessness (5 points), and flehmen response (3 points). These scores were used to classify standing estrus and secondary estrus signs in the present study.

Standing to be mounted was considered the principal indicator of standing estrus and was assigned the highest estrus score. Secondary indicators included flehmen response, restlessness, sniffing the vulva or vaginal region of another cow, chin resting, being mounted without standing, mounting, or attempting to mount another cow, and mounting the head side of another cow[22].

Estrus-related events were first identified from continuous video records and then assigned to 1-min observation windows for alignment with the 10-s IMU records. Standing estrus was classified when standing-to-be-mounted behavior occurred within a 1-min observation window, corresponding to an estrus score ≥100. Secondary indicators were recorded when one or more secondary signs occurred within a 1-min observation window and the corresponding estrus score was ≥15 and <100. To reduce misclassification of isolated or ambiguous events, a secondary indicator was considered valid for pre-estrus identification only when another secondary indicator occurred within the following 2 h. Estrus scores were assigned according to the presence and temporal clustering of these indicators, following the principle that estrus expression should be assessed using both standing behavior and secondary signs rather than activity level alone.

Ovulation was determined by transrectal ovarian ultrasonography. Ultrasonographic monitoring began when estrus-related behavioral signs were first detected from video observation. Ovarian examinations were performed every 6 h, at approximately 05:00, 11:00, 17:00, and 23:00 h, to monitor the dominant follicle until it was no longer visible. For each cycle, the ovulation interval was defined as the period between the last ultrasound detection of the dominant follicle and the first subsequent examination in which it was no longer detected. A six-hour ovulation window was then assigned based on this interval [5,19,23]. Video-derived behavioral signs, estrus scores, and ultrasound-confirmed ovulation served as biological references for estrus phase classification and were aligned with IMU records for phase-specific analysis [5,17].

2.5. Data Preprocessing and Cleaning

Raw IMU data were processed using Python. Data preparation procedures included data importation, timestamp formatting, chronological sorting, missing-value inspection, removal of malformed entries, and assessment of time gaps. Observations containing missing or invalid timestamps, missing sensor values, or biologically implausible values were excluded prior to analysis. Biologically implausible values were defined as values outside the plausible operating range of the device or values inconsistent with valid accelerometer and gyroscope outputs.

Data completeness was evaluated at the cow-cycle level. Completeness was calculated as the proportion of observed IMU records relative to the expected number of records within each analytical window. The expected number of records was estimated from the duration of the analytical window and the expected sensor recording interval. Only cycles with sufficient IMU coverage across the analytical window were retained for analysis, with the target completeness threshold defined as >95% of expected IMU records.

Time gaps were identified based on the time difference between consecutive records. Given the expected sensor recording interval of 10 seconds, gaps exceeding 120 seconds were classified as extended gaps. Extended gaps were identified to minimize potential bias in the estimation of duration-based posture variables, particularly behavioral duration, posture percentage, and lying bout characteristics.

For each cow-cycle combination, the analytical window was defined relative to ovulation and extended from the normal pre-estrus baseline to the early post-ovulation period. Following cleaning, observations were assigned to estrus phases according to the rule-based definitions described in Section 3.6.

2.6. Estrus Phase Definition

Each natural estrus cycle was divided into six biologically meaningful phases using the video-derived behavioral signs, estrus scores, and ultrasound-confirmed ovulation described in Section 3.4. The six-phase classification was study-specific but was set in accordance with previous estrus-behavior and ovulation time frameworks. Previous studies have divided estrus into biologically meaningful subperiods, including pre-standing, standing, and post-standing periods, supporting the concept that behavioral estrus is a temporal process rather than a single uniform event [24]. In addition, the relationship between behavioral estrus and ovulation is well documented, with ovulation generally occurring after estrus onset and after the end of standing estrus, although the interval between estrus expression and ovulation varies among cows [5,17]. Therefore, the present study adapted these concepts by integrating behavioral signs, estrus scores, and ultrasound-confirmed ovulation into six analytical phases: normal, pre-estrus, standing estrus, late estrus, peri-ovulation, and early post-ovulation. The operational definition of each phase is provided in Table 1. This phase-based structure enabled the characterization of sensor-derived activity and posture-related behavioral changes before, during, and after standing estrus in relation to ovulation.

These definitions were applied independently to each cow-cycle combination. Phase duration was calculated in hours, and sensor-derived activity features and posture-related variables were summarized at the cow-cycle-phase level. This structure enabled phase-level evaluation of movement and behavioral changes while preserving the cow-cycle organization of the dataset.

2.7. Feature Extraction

IMU-derived activity features and posture-related variables were extracted to characterize movement and behavioral changes across biologically defined estrus phases. The feature set included absolute movement measures, baseline-adjusted activity features, an exploratory Combined Activity Index, and posture-based outcomes. The primary and secondary outcome variables are summarized in Table 2.

Absolute activity measures were calculated from accelerometer and gyroscope signals and included signal vector magnitude of acceleration, vectorial dynamic body acceleration, and gyroscope magnitude. These measures were selected because they captured complementary aspects of cow movement, including linear acceleration intensity, dynamic body acceleration, and rotational activity. Posture-based outcomes were derived from annotated behavioral labels and included the proportions of time spent lying, standing, and walking, together with lying bout characteristics.

2.7.1. Signal Vector Magnitude of Acceleration

Signal vector magnitude of acceleration (SVM_acc) was calculated from tri-axial accelerometer signals to quantify overall linear movement intensity. For each timestamp, acceleration values from the X, Y, and Z axes were combined using the Euclidean norm:

$$SV{M}_{acc}=\sqrt{\left({AccX}^2{+AccY}^2{+AccZ}^2\right)}$$

where AccX, AccY, and AccZ represent acceleration values along the three orthogonal axes. Higher SVMacc values indicate greater overall linear movement intensity.

2.7.2. Vectorial Dynamic Body Acceleration

Vectorial dynamic body acceleration (VeDBA) was calculated to quantify dynamic movement after reducing the influence of static acceleration. For each accelerometer axis, the static component was estimated using a rolling mean, and the dynamic component was obtained by subtracting the static component from the raw acceleration signal. VeDBA was then calculated as:

$$VeDBA=\sqrt{\left({DBAx}^2{+DBAy}^2{+DBAz}^2\right)}$$

where DBAx, DBAy, and DBAz represent dynamic body acceleration along each axis. VeDBA was used as an activity-related feature because it reflects dynamic body movement and may increase during active estrus-related behaviors.

2.7.3. Gyroscope Magnitude

Gyroscope magnitude (Gyro_mag) was calculated from tri-axial gyroscope signals to quantify rotational movement intensity. For each timestamp, angular velocity values from the X, Y, and Z axes were combined using the Euclidean norm:

$$Gyro\_mag=\sqrt{\left({GyroX}^2{+GyroY}^2{+GyroZ}^2\right)}$$

where GyroX, GyroY, and GyroZ represent angular velocity around the three axes. Gyro_mag was included because rotational movement of the head and neck may increase during restlessness, walking, mounting attempts, or other estrus-related activities.

2.7.4. Baseline-Adjusted Activity Features

Baseline-adjusted activity features were calculated to account for individual differences in baseline activity among cows. The normal phase was used as the cow-specific reference period. For each cow, the mean and standard deviation of each activity feature during the normal phase were calculated and used to generate baseline-standardized and baseline-difference transformations.

Baseline-standardized features were calculated as:

$${Base}_z={\displaystyle \frac{X_t-{\mu }_{baseline}}{{\sigma }_{baseline}}}$$

where Xt represents the activity value at time t, μ baseline represents the cow-specific baseline mean during the normal baseline phase, and σ baseline represents the cow-specific baseline standard deviation during the normal baseline phase. These values indicate how many standard deviations each observation deviated from the cow’s normal baseline activity level. Baseline-difference features were calculated as:

$$Bas{e}_{diff}=X_t-{\mu }_{baseline}$$

This transformation quantified the absolute deviation from baseline while preserving the original feature scale. Baseline-adjusted features were generated for SVM_acc, VeDBA, and Gyro_mag.

2.7.5. Exploratory Combined Activity Index

An exploratory Combined Activity Index was developed to integrate dynamic body movement and rotational activity into a single composite measure. This index was calculated using VeDBA and Gyro_mag because these features represent complementary dimensions of estrus-related movement: VeDBA reflects dynamic body acceleration, whereas Gyro_mag reflects rotational head and neck movement.

Before combination, VeDBA and Gyro_mag were normalized using min–max normalization to place both variables on a comparable 0–100 scale:

$$X_{norm}={\displaystyle \frac{X-X_{min}}{X_{max}-X_{min}}}\times 100$$

where X represents the observed value, Xmin represents the minimum value, and Xmax represents the maximum value of the feature.

The Combined Activity Index was calculated as a weighted composite of normalized VeDBA and normalized Gyro_mag:

$$Combined Activity Index=\left({\omega }_1\times {VeDBA}_{norm}\right)+({\omega }_2\times {Gyro}_{norm})$$

Where ${\omega }_1+{\omega }_2=1$

Several weighting combinations were explored, including VeDBA80:Gyro20, VeDBA70:Gyro30, VeDBA60:Gyro40, VeDBA50:Gyro50, VeDBA40:Gyro60, VeDBA30:Gyro70, and VeDBA20:Gyro80. The VeDBA40:Gyro60 weighting was selected for presentation because it represented the minimum gyroscope contribution that still showed exploratory separation between the standing estrus phase and the two non-standing estrus phases based on phase-level distribution patterns. The final Combined Activity Index was calculated as:

Combined Activity Index = 0.40 × normalized VeDBA + 0.60 × normalized Gyro_mag

Higher values indicated greater combined dynamic and rotational movement. The Combined Activity Index was interpreted as an exploratory composite indicator of estrus-related activity intensity.

2.7.6. Posture Duration

Posture-related behavioral features were derived from video-based behavioral annotations. The primary behavioral categories used to generate posture-related variables were lying, standing, and walking. Although eating was defined in the ethogram, it was not included as a primary posture variable in the phase-level analysis because the main focus of the study was on resting, standing, and locomotor activity during estrus. Behavioral definitions are summarized in Table 3.

The primary posture-related variables were the percentages of time spent lying, standing, and walking within each estrus phase. Proportional variables were used instead of raw duration totals because phase durations differed among cow-cycle combinations. For each cow-cycle-phase unit, the percentage of each behavior was calculated as follows:

$$Percent behavior={\displaystyle \frac{Duration of behavior within phase}{Total phase duration}}\times 100$$

Lying bout characteristics were also calculated to describe resting patterns across estrus phases. A lying bout was defined as a continuous period during which a cow was classified as lying. Lying bout rate was expressed as the number of lying bouts per hour, whereas mean lying bout duration was expressed in minutes per bout. These variables were included to characterize changes in the frequency and duration of resting episodes, rather than total lying time alone.

2.8. Data Summarization and Analytical Unit

To minimize pseudoreplication, the dataset was aggregated at the cow-cycle-phase level prior to inferential testing. Each cow-cycle-phase unit served as the analytical unit rather than treating individual timestamp-level measurements as independent observations.

Activity and posture-related variables were summarized within each cow-cycle-phase unit. Activity summaries included mean SVM_acc, mean VeDBA, mean Gyro_mag, baseline-standardized activity features, and baseline-difference activity features. Posture-related summaries included the percentages of time spent lying, standing, and walking, lying bout rate per hour, and mean lying bout duration expressed in minutes.

The resulting phase-level dataset served as the basis for inferential comparisons among estrus phases. This approach allowed comparisons to reflect biological variation among cows and estrus cycles rather than differences in raw sensor-record counts. It also reduced the risk of artificially inflated sample sizes from repeated within-cow time-series measurements.

2.9. Statistical Analysis

Statistical analyses were conducted in R using the cow-cycle-phase summary dataset described in Section 3.8. Because the analysis was based on cow-cycle-phase summaries, statistical comparisons were interpreted as exploratory phase-level comparisons. Descriptive statistics were calculated for each IMU-derived activity feature and posture-related measure across estrus phases.

Before inferential testing, each phase-level measure was assessed for normality using the Shapiro–Wilk test. When the normality assumption was met, phase-related variation was evaluated using one-way analysis of variance (ANOVA), followed by Tukey’s honestly significant difference post hoc test. When the normality assumption was not met, the Kruskal–Wallis test was applied, followed by Bonferroni-adjusted pairwise Wilcoxon rank-sum tests.

Statistical significance was defined as p < 0.05. Compact letter displays were generated for figures and summary tables to indicate significant phase differences. Phases sharing a letter were considered not significantly different, whereas phases with different letters were considered significantly different.

Graphical outputs included phase-level boxplots with compact letter displays, trajectories of activity and posture-related variables relative to ovulation, and supplementary plots for individual cow–cycle combinations. These visualizations were used to evaluate overall phase-level patterns and individual variability in estrus-associated behavioral dynamics.

3. Results

3.1. Data Structure and Phase Distribution

A total of 285,337 observations across 11 natural estrus cycles from 5 dairy cows were included in the analysis (Table 5). These observations were distributed across six biologically defined estrus phases. The normal phase (Phase 1) represented the largest proportion of the dataset, with 88,100 observations (30.9%). This was followed by the early post-ovulation phase (Phase 6; 22.4%) and the late estrus phase (Phase 4; 19.4%). The pre-estrus phase (Phase 2; 8.4%) and peri-ovulation phase (Phase 5; 6.4%) accounted for the smallest proportions of the dataset. The number of observations varied among cows and estrus cycles, ranging from 7,249 in Cycle 7 (Cow3) to 32,534 in Cycle 11 (Cow5). All estrus phases were represented across the cows and cycles included in the study.

3.2. Estrus Phase Duration Characteristics

Estrus phase durations are summarized in Table A1. The normal phase (Phase 1) had the longest mean duration (26.92 ± 7.41 h), whereas the peri-ovulation phase (Phase 5) had the shortest mean duration (5.92 ± 0.24 h). Duration ranges varied across estrus phases. The pre-estrus phase (Phase 2; 8.74 ± 6.69 h) and late estrus phase (Phase 4; 17.77 ± 8.56 h) exhibited greater variation among estrus cycles. Standing estrus (Phase 3) had a mean duration of 11.41 ± 5.85 h.

3.3. Changes in Absolute Activity Features Across Estrus Phases

Absolute activity features are presented in Figure 2 and Table A3. VeDBA and Gyro_mag showed significant differences across estrus phases, whereas SVM_acc showed no significant differences among phases. VeDBA was highest during standing estrus (Phase 3; 0.17 ± 0.05). Lower values were observed during the remaining phases. Elevated Gyro_mag values were observed during the pre-estrus phase (Phase 2; 20.19 ± 6.56) and standing estrus (Phase 3; 22.10 ± 6.85), whereas lower values occurred during the remaining phases. Overall, VeDBA and Gyro_mag increased from the normal phase to standing estrus and decreased during the late estrus, peri-ovulation, and post-ovulation phases.

3.4. Baseline-Standardized and Baseline-Difference Activity Changes

Baseline-standardized and baseline-difference activity features are shown in Figure 3, Figure 4, and Table A3. Baseline-standardized VeDBA and Gyro_mag showed significant differences across estrus phases, whereas baseline-standardized SVM_acc showed no significant differences among phases. Baseline-standardized SVM_acc distributions overlapped substantially among phases. The highest baseline-standardized VeDBA values were observed during standing estrus (Phase 3). Elevated baseline-standardized Gyro_mag values were observed during pre-estrus (Phase 2) and standing estrus (Phase 3). Lower values for both features were observed during the late estrus, peri-ovulation, and post-ovulation phases, including negative values during peri-ovulation. Baseline-difference VeDBA and Gyro_mag showed similar phase-related patterns, with elevated values during pre-estrus and standing estrus and lower values during peri-ovulation and post-ovulation phases. Baseline-difference SVM_acc showed no significant differences among phases. Overall, both baseline-standardized and baseline-difference activity features increased during pre-estrus and standing estrus and decreased during peri-ovulation and post-ovulation phases.

3.5. Exploratory Combined Activity Index Across Estrus Phases

The exploratory Combined Activity Index (VeDBA40 Gyro60) is shown in Figure 5 and Table A3. The Combined Activity Index showed significant differences across estrus phases. Elevated Combined Activity Index values were observed during pre-estrus (Phase 2; 27.41 ± 7.34) and standing estrus (Phase 3; 30.36 ± 5.33). Lower values were observed during the normal, late estrus, peri-ovulation, and post-ovulation phases. Overall, the Combined Activity Index increased from the normal phase to standing estrus and decreased during the late estrus, peri-ovulation, and post-ovulation phases.

3.6. Changes in Posture-Related Behavior Across Estrus Phases

Posture-related behavior proportions are presented in Figure 6 and Table A3. Lying, standing, and walking proportions showed significant differences across estrus phases. Lying proportion was lowest during standing estrus (Phase 3; 13.89 ± 9.71%) and highest during the peri-ovulation phase (Phase 5; 68.31 ± 24.94%). Standing proportion was highest during standing estrus (Phase 3; 70.16 ± 14.74%), whereas walking proportion also reached its highest value during standing estrus (13.70 ± 6.95%). Lower walking proportions were observed during the remaining phases. Overall, lying behavior decreased from the normal phase to standing estrus, whereas standing and walking behaviors increased. During the peri-ovulation and post-ovulation phases, lying behavior increased while standing and walking decreased.

3.7. Lying Bout Characteristics

Lying bout characteristics are presented in Figure 7 and Table A3. Lying bout rate did not differ significantly across estrus phases, whereas mean lying bout duration showed significant differences among phases. The shortest mean lying bout duration was observed during standing estrus (Phase 3; 22.33 ± 29.77 min). Higher values were observed during the normal, late estrus, peri-ovulation, and post-ovulation phases. Overall, mean lying bout duration decreased from the normal phase to standing estrus and increased during the late estrus, peri-ovulation, and post-ovulation phases.

3.8. Individual Cycle Variability in Estrus-Related Behavior

Individual time-series patterns of activity and posture-related behaviors are presented in Figure A1 to A3. Temporal variations in activity, posture proportions, and mean lying bout duration were observed among cows and estrus cycles. Across cows, VeDBA and Gyro_mag increased during pre-estrus and standing estrus, whereas higher lying proportions and longer mean lying bout durations were observed during peri-ovulation and post-ovulation phases. Cycle No.5 (Cow 2) showed pronounced increases in VeDBA and Gyro_mag together with high standing proportions during standing estrus. Cycle No.10 (Cow4) exhibited the highest standing proportions during pre-estrus and standing estrus, whereas Cycle No.11 (Cow5) showed the highest walking proportion during standing estrus (18.06%). Shorter mean lying bout durations were observed during standing estrus across cows.

4. Discussion

This study demonstrated that IMU-derived activity features and posture-related behaviors differed across biologically defined phases of natural estrus in dairy cows. The findings indicate that behavioral expression during natural estrus is dynamic rather than uniform across the peri-estrus period. The clearest behavioral differentiation occurred during pre-estrus and standing estrus, indicating that these phases are particularly important for sensor-based estrus monitoring. By representing estrus as a sequence of biologically defined phases rather than as a simple estrus versus non-estrus condition, the present study provided phase-specific information on changes in movement intensity, rotational activity, posture allocation, and resting behavior before, during, and after standing estrus.

Among the evaluated activity features, VeDBA and Gyro_mag showed more biologically informative responses than SVM_acc for characterizing estrus-related movement changes under the conditions of this study. VeDBA showed the greatest response during standing estrus, whereas Gyro_mag increased during both pre-estrus and standing estrus, indicating that dynamic acceleration and rotational movement captured different aspects of estrus expression. This pattern is consistent with previous studies reporting increased activity, restlessness, mounting-related behavior, and social interaction during estrus [1,2,5,11]. The phase-related patterns of VeDBA and Gyro_mag reflected their association with distinct movement components: VeDBA represents dynamic body acceleration after reducing the influence of static acceleration, whereas Gyro_mag captures rotational head and neck movements associated with restlessness, locomotion, and mounting-related interactions [15,25,26]. In contrast, SVM_acc includes both static and dynamic acceleration components and can be affected by posture, gravity-related acceleration, and neck position in a neck-mounted sensor configuration [27,28]. These findings indicate that dynamic acceleration and rotational movement features provide more specific information on estrus-associated movement changes than simple acceleration magnitude alone [29,30].

Baseline-adjusted activity features, including baseline-standardized and baseline-difference variables, improved the interpretation of estrus-related changes by expressing activity deviations relative to each cow’s normal activity level. This approach is consistent with previous automated estrus-monitoring studies, in which estrus detection was commonly based on identifying deviations from an individual cow’s usual activity pattern rather than relying only on fixed population-level thresholds [1,7,8]. This individual-reference approach is biologically relevant because activity expression during estrus varies among cows and is influenced by cow-level and cycle-level factors, including baseline activity, parity, milk production, health status, social behavior, and estrus intensity [2,5,20,31]. In agreement with this concept, both baseline-standardized and baseline-difference VeDBA and Gyro_mag showed clearer phase-related patterns than the corresponding SVM_acc variables. Baseline-standardized features described deviations from the normal phase in standard deviation units, whereas baseline-difference features preserved the original measurement scale and described the absolute change from baseline. These findings indicate that cow-specific baseline adjustment helped distinguish estrus-associated changes from normal inter-cow variability and supports within-cow baseline adjustment as a biologically meaningful preprocessing step for future estrus-monitoring models [32,33].

The exploratory Combined Activity Index provided preliminary evidence for the value of integrating dynamic body acceleration and rotational movement into a single composite indicator. This index was elevated during pre-estrus and standing estrus, consistent with the phase-related increases observed for VeDBA and Gyro_mag. This pattern indicates that combining these features captures a broader representation of estrus-associated behavioral arousal than either movement dimension alone, because estrus expression involves coordinated changes in locomotor activity, restlessness, mounting-related behavior, and social interaction [2,5,34]. Previous estrus-monitoring studies have shown that combining multiple sensor-derived variables, such as acceleration with location or localization data, can improve estrus-detection performance compared with single-sensor approaches [35,36]. However, previous sensor-based estrus studies have not specifically examined a weighted composite activity index integrating VeDBA and Gyro_mag for differentiating biologically defined phases of natural estrus. Related cattle behavior studies support the complementary value of accelerometer- and gyroscope-derived information for behavior classification and reproductive monitoring [14,15]. Therefore, the Combined Activity Index should be interpreted as an exploratory but biologically grounded indicator, and further validation in independent dairy cow populations and management systems is required before practical estrus-monitoring application.

Posture-related behaviors complemented IMU-derived activity features by showing how natural estrus altered posture allocation, locomotion, and resting structure. The decrease in lying proportion during standing estrus is consistent with previous studies reporting reduced resting behavior and increased behavioral restlessness during estrus [2,5,11]. In parallel, the increased standing proportion during pre-estrus and standing estrus supports previous observations that cows spend more time in an upright and behaviorally active state during estrus expression, particularly because standing to be mounted is recognized as the primary behavioral sign of estrus [21,37]. Walking proportion also increased most clearly during standing estrus, agreeing with previous reports that estrus is associated with increased locomotor activity and movement-based activity alerts [1,2,11]. Unlike general activity-monitoring studies, the present study evaluated walking as a posture-related behavioral measure across biologically defined estrus phases, allowing more direct interpretation of the locomotor component of standing estrus. Therefore, walking serves as a complementary behavioral indicator to IMU-derived activity features, while its interpretation should consider housing space, social context, and individual cow variability [20,31].

The lying bout findings further clarified the structure of resting behavior during natural estrus. Previous studies have reported that cows reduce resting behavior and increase activity, restlessness, and locomotion around [2,5,18]. Consistent with these reports, lying proportion decreased during standing estrus in the present study. However, lying bout rate did not differ significantly across phases, whereas mean lying bout duration was shorter during standing estrus. This pattern indicates that the reduction in lying behavior was related mainly to shorter and less sustained resting periods rather than to more frequent lying episodes [38,39]. Therefore, mean lying bout duration complements lying proportion by describing the temporal organization of resting behavior and providing a more detailed interpretation of resting disruption during standing estrus [40,41].

Previous studies have used different reference standards or criteria to define estrus, resulting in varying levels of biological and temporal resolution. Studies using activity sensors, visual estrus observation, or activity-monitor alerts have shown that estrus is associated with increased activity, reduced inactivity or rumination, mounting-related behavior, and a biological relationship with ovulation timing [5,11,19]. However, most sensor-based studies have focused on broad estrus versus non-estrus comparisons, general activity increases around estrus, or ovulation-related outcomes rather than the temporal progression of natural estrus across distinct behavioral phases [2,7]. The present study addressed this limitation by evaluating IMU-derived activity features, posture-related behaviors, and lying bout characteristics across biologically defined phases of natural estrus. Natural estrus was divided into six ovulation-anchored phases including normal, pre-estrus, standing estrus, late estrus, peri-ovulation, and early post-ovulation phases, using video-derived behavioral signs and ultrasound-confirmed ovulation. This phase-based framework enabled behavioral changes to be interpreted according to their temporal relationship to standing estrus and ovulation rather than as generalized increases in activity alone.

This phase-based approach showed that behavioral differentiation was most evident during standing estrus, when standing-to-be-mounted behavior was accompanied by increased dynamic movement, rotational activity, baseline-adjusted activity deviations, increased standing and walking proportions, and disruption of lying continuity. Walking proportion and the Combined Activity Index were significantly elevated during standing estrus compared with most non-standing phases, indicating that locomotor expression and integrated dynamic–rotational movement contributed to standing-estrus differentiation. These findings are consistent with previous studies identifying standing-to-be-mounted behavior as the primary behavioral sign of estrus and reporting increased mounting-related activity, locomotion, and behavioral restlessness during standing estrus [2,5,24]. Because standing estrus precedes ovulation, accurate identification of this phase is biologically and practically relevant for insemination-timing decisions [17,19]. Previous studies have reported that ovulation occurs approximately 24–33 h after estrus onset and 15–22 h after estrus end, although these intervals vary according to cow-level factors, estrus definition, observation frequency, and ovulation-monitoring method [42,43]. In the present study, the 6-h peri-ovulation window was estimated to begin approximately 29.18 h after the first observed standing-to-be-mounted event and 17.77 h after the last observed standing-to-be-mounted event, with the midpoint occurring approximately 32.14 h and 20.73 h after these respective reference points. These estimates are broadly consistent with previous reports of ovulation timing relative to estrus onset and estrus end [44,45]. Therefore, phase-specific differentiation of activity and posture-related features, particularly walking proportion and the Combined Activity Index during standing estrus, provides a biological basis for future models designed to estimate the ovulation window and support insemination-timing decisions.

Several strengths and limitations should be considered when interpreting these findings. A major strength was the use of naturally occurring estrus under routine farm conditions, allowing behavioral expression to be evaluated without hormonal synchronization. In addition, the combination of continuous video observation and ultrasound-confirmed ovulation strengthened the biological reference standard for phase classification [5,17].

However, the study was based on a relatively small dataset of five cows and eleven natural estrus cycles from a single loose-barn facility. In addition, because some cows contributed multiple estrus cycles, the repeated structure of cycles within cows was not fully modeled. Therefore, the findings should be interpreted as observational and exploratory, and their generalizability to other herds, breeds, housing systems, management conditions, and sensor placements requires further validation. Future studies with larger datasets should apply mixed-effects models to account for repeated cycles within cows and to distinguish phase-related behavioral changes from cow-level and cycle-level variation.

5. Conclusions

This study demonstrated that IMU-derived activity features and posture-related behaviors varied across biologically defined phases of natural estrus in dairy cows, with the clearest differentiation occurring during pre-estrus and standing estrus. VeDBA and Gyro_mag were more informative than SVM_acc for characterizing estrus-related movement changes, while baseline-adjusted variables improved interpretation by expressing deviations relative to each cow’s normal activity level. Posture-related variables complemented activity features, as standing estrus was characterized by reduced lying proportion, increased standing and walking proportions, and shorter mean lying bout duration. The exploratory Combined Activity Index integrating VeDBA and Gyro_mag showed potential as a biologically grounded indicator of estrus-related activity, although further validation is required. Overall, the ovulation-anchored six-phase framework characterized natural estrus beyond a simple estrus versus non-estrus comparison and identified candidate variables for future estrus-monitoring models aimed at standing-estrus differentiation, ovulation-window estimation, and insemination-timing support.