See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/250719002

Extracting poultry behaviour from time-series

weigh scale records
ARTICLE in COMPUTERS AND ELECTRONICS IN AGRICULTURE JUNE 2008
Impact Factor: 1.49 DOI: 10.1016/j.compag.2007.08.015

CITATIONS

2 AUTHORS, INCLUDING:
Hongwei Xin
Iowa State University
288 PUBLICATIONS 1,690 CITATIONS
SEE PROFILE

Available from: Richard Stephen Gates

Retrieved on: 27 August 2015

ARTICLE IN PRESS

COMPAG 2049 17

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/compag

Richard S. Gates a, , Hongwei Xin b

Biosystems and Agricultural Engineering Department, 128 C.E. Barnhart Building, University of Kentucky, Lexington, KY 40546, USA
Agricultural and Biosystems Engineering Department, 3204 NSRIC, Iowa State University, Ames, IA 50011-3310, USA

a r t i c l e

i n f o

a b s t r a c t

RO

OF

Extracting poultry behaviour from time-series

weigh scale records

Keywords:

Algorithms for determining individual bird feeding statistics and stereotyped pecking

Poultry

behaviour from time-series recordings of feed weight were developed and compared to video

10

Behaviour

observations. Data taken from two separate experiments involving broiler and laying hen

11

Animal well-being

chickens were used to evaluate the algorithms. The effects of algorithm tuning parameters

12

Welfare

including thresholds for changes in weight and sequential number of stabilized readings,

13

Feeding

arithmetic moving average for meal tare values, and the sampling frequency of feed weight

DP

recordings were evaluated. Results suggest that a minimum sampling frequency of 0.51 Hz
is recommended for discerning behavioural changes that include timing of feeding events

TE

and their duration. However, lower sampling frequencies are acceptable for determining

1.

EC

hourly (or greater) feed consumption.

Introduction

The question of how management or environmental stimuli

may inuence poultry behaviour and/or well-being is of considerable importance for fundamental studies of behavioural
response to stimuli, and as a means of assessing appropriate
management and environmental designs for commercial production. What responses should be measured and whether
bird response is correlated to well-being are active areas of
investigation. If multiple choices or stimuli are available, it is
a considerable challenge to assign behavioural outcomes to
these treatments. Discrimination between competing choices
or stimuli requires careful experimental design to assess birds
choice selection. Wathes et al. (2001) list a set of criteria postulated by Abeyesinghe (2000), which can be used to normalize
assessments of animal response. These response assessment
criteria include a need for sensitivity to all stimuli, responsive
over different time periods and levels of stimulus, and suitable
repeatability for scientic assessment.

14
15

RR

16
17
18
19
20

CO

21
22
23
24
25
26

UN

27
28
29
30

Q1
1
2

2007 Elsevier B.V. All rights reserved.

One means of assessing bird response to stimuli involves

careful analysis of characteristics of individuals or groups
over time. Monitoring individual behaviour during research
trials is typically performed with some type of video imaging system. For poultry, behavioural activities are categorized
into events such as eating, drinking, preening, resting,
and stereotyped activities directed at different targets. This
assessment methodology is time-consuming, hence costly,
tedious and prone to errors, even with modern commercially available research systems that compile the statistics
semi-autonomously. There is an increasing need for means
to further automate collection of event-based behavioural
responses (Gates et al., 1995; Gates and Xin, 2001; Persyn et
al., 2003, 2004; Xin et al., 1993).
With behavioural monitoring, it is not feasible to determine variation in feed and water use amongst individual birds
within a treatment, nor do they dynamically monitor feed
and water intake for the same bird as environment is modied. Recent measurements with the Individual Bird Unit (IBU)

Corresponding author. Tel.: +1 859 257 3000; fax: +1 859 257 5671.
E-mail address: [email protected] (R.S. Gates).
0168-1699/$ see front matter 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.compag.2007.08.015
Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48

ARTICLE IN PRESS

COMPAG 2049 17

54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81

Materials and methods

82

2.1.

Equipment

83

2.1.1.

IBU system

87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103

CO

86

The Individual Bird Unit (IBU) system consisted of 24 feeding/drinking stations divided into four groups of six stations.
Each group was located in one of two environmentally controlled chambers. Each feeding station (Fig. 1) consisted of a
precision electronic weighing scale (model CT1200, Ohaus Corporation, Florham Park, NJ) with a 1210 g capacity and a 0.1 g
resolution and a plastic feeder measuring 13 cm (L) 13 cm
(W) 15 cm (H) (5 in. (L) 5 in. (W) 6 in. (L)). The plastic feeder
had a u-shaped access side opening and its bottom was fastened to the electronic scale with Velcrotm strips. Each scale
had an RS232 serial interface connected to a custom-built
microcontroller with RS232 and RS485 communication ports,
digital input/output and analogue/digital converter (KG Systems Inc., East Hanover, NJ). The 24 microcontrollers were
networked to a master unit via the RS485 ports; the master
microcontroller assigned polling commands, collected information from each unit, and forwarded the data to a PC via
RS232. The weigh scales were located on a wooden stand in
front of the individual birdcages. The cages measured 25 cm
(W) 46 cm (D) 46 cm (H) (10 in. (W) 18 in. (D) 18 in. (H)).

UN

85

RR

2.

84

OF

53

RO

52

DP

51

system (Puma et al., 2001; Cook and Xin, 2004) indicate that
individual birds adjust their eating and drinking behaviour
quite differently for the same thermal treatment, and that
this effect is masked when comparing group means. Collection of data for variability between individuals, if practical,
may provide an efcient basis for assessing bird response
using population percentages, minimization (or elimination)
of extreme responses, or genetic improvement by individual
et
selection of previously unavailable selection criteria (Na as
al., 2000).
One set of behavioural assessment criteria is feeding activity. Measures include number of meals, meal size, meal
duration, ingestion rate, meal intervals, and proportion of
time spent eating. In addition, birds spend varying amounts
of time pecking without eating, dened as stereotyped pecking behaviour. Such information may be useful to understand
how to better design housing systems to satisfy birds inherent needs for food, and to study the space requirements
and the impact of competition in commercial settings (Cook
and Xin, 2004). Behaviour of individual birds at the feeder, if
quantied, could form a comparative basis for assessing alternative management and housing strategies (Persyn et al., 2003,
2004).
The objective of this research was to devise, test and validate two algorithms to determine individual bird activities
including time at station, activity at station, meal size and
duration, for use with time-series recordings of feed levels
from the existing Individual Bird Use system. In addition, one
algorithm was tested in its ability to extract time allotment
activities as compared to video recordings. In the following
sections, we describe the two algorithms, discuss the effect
of key parameters to tune the algorithms, and as appropriate
compare them to baseline data.

TE

50

EC

49

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

Fig. 1 The IBU feed scale system installed in one chamber

with six cages (top). View of a single cage, with feed
weighing system.

Complete details of the IBU system can be found in Puma et

al. (2001).
Readings of the feeder weighing scale were scanned on
periodic command and captured with a Visual Basic macro
executing in MS Excel. Maximum sampling frequency for
the IBU system is one sample each 4 s (i.e. Ts = 4 s) with
all 24 units operational. For purposes of the work reported
here, sampling times Ts of 4 and 30 s were selectively
used.
Behavioural data of two groups of four birds were acquired
with a video recording system that consisted of two CCD
cameras (Panasonic, AG-6730), a time-lapsed VCR (Panasonic,
PV-V4520) and a TV monitor. The cameras were mounted so
that full images of four neighbouring hens could be recorded.
The recordings were compiled on an hourly basis and used
for comparison purposes. The number, duration and type of

Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119

ARTICLE IN PRESS

COMPAG 2049 17

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

2.1.2.

122
123

High-frequency sampling system

142

2.2.

132
133
134
135
136
137
138
139
140

143
144

Experimental birds

Both laying hens and broilers were utilized in these studies.

The data from the IBU system was taken from a study on

TE

131

EC

130

RR

129

CO

128

UN

127

DP

141

Four measurement stations were used in a separate study to

obtain high-frequency sampling data. Each station had one
electronic balance (2000 0.1 g) (model HF-2000, AND Inc.,
Tokyo, Japan). The balance provided a 01 VDC analogue output for the weight range. A rectangular aluminium feeder
(20 cm (W) 10 cm (D) 5 cm (D)) was attached to the platform of the balance using Velcro. The analogue output signal
of each balance was connected to a differential analogue
input channel of an electronic data logger (model CR23X,
Campbell Scientic Inc., Logan, UT, USA). The CR23X contained a 4 MB extended storage memory, and it measured and
stored the output signals at Ts = 0.1 s intervals (10 Hz sampling frequency). At this sampling rate, about 30 MB of raw
weight and time data were collected per day for the four
measurement stations. These data were downloaded hourly
to a PC.

126

effects of drinking water temperature during diurnal, warmto-hot environments (Xin et al., 2002). The hens were W-36
layers, approximately 32 weeks old at the start of the test.
Lights were turned on at 5:00 and off at 21:00 each day. Feed
was replenished daily between 8:10 and 8:50. A single day
of data from the 3rd week of a 4-week heat stress event
(two hens), and 2 days from the 2nd week of a 2-week thermoneutral recovery period (four hens) were available. These
hen day combinations were selected for analysis because
they had few missing values and simultaneous video recordings of these birds.
The high-frequency data were taken from a broiler feeding
study (Japanese, Chunky breed) to assess feeding behaviour
and consumption of a specialized sesame diet. At 4 weeks
of age, birds with similar body mass (BM) were sub-grouped
for feeding behaviour measurement. Starting with the heavier BM, two birds of similar BM from each group at a time
were brought from the rearing house to the measurement laboratory. At the measurement lab, the birds were individually
housed in cages (24 cm (W) 43 cm (D) 40 cm (H)) at constant ambient temperature of 24 C and relative humidity of
46%. After a 2-day acclimation, feeding behaviour was monitored for the next 45 h, and those data associated with the
nal 24 h (6 am6 am) were taken to be representative of the
normal behaviour of the birds and thus used in the analysis.
Full details are available from Xin (2001, unpublished research
report). For purposes of this study, sample sequences from a
single day of four birds were used.

OF

125

121

RO

124

events (time at feeder, time at drinker, remainder of nonresting time) were tabulated from visual analysis of the
time-lapsed recordings. Four hens during 1 day of heat stress
and two hens during 1 day of the recovery period were utilized
for this activity.

120

Fig. 2 Flow chart of main steps of algorithm AL2.

Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171

ARTICLE IN PRESS

COMPAG 2049 17

176
177
178
179
180
181
182
183
184
185
186
187
188
189
190

191
192
193
194
195

196
197

comparison of Wk to a threshold weight to assess whether

a candidate feeding event occurred;
determination of event start times and duration;
determination of whether each event represents feeding or
non-feeding activity.
For AL1 (Ts = 0.1 s), key features and discriminant steps
include:

208

For AL2 key discriminant steps include:

202
203
204
205
206

209
210
211
212
213
214
215
216

217
218
219
220
221
222
223

RR

201

AL1 was considered optimally tuned for discerning feeding

activity statistics from the high-frequency data, and suitable
for use in assessing dietary and environment effects on individual birds. However, a reduction in sampling frequency was
of interest to reduce storage requirements and processing
times. Thus, 110 min of representative data sampled at 10 Hz
(Fig. 3) were decimated (Ts = 1, 2, 5, 10 and 30 s) and analyzed
with different algorithm parameters to determine if similar
conclusions could be drawn. Specically, the number of samples in the ARMA, R, was adjusted with sampling interval Ts
(Ts = 0.1, R = 100 and 200; Ts = 1, R = 10, 20, and 30; Ts = 2, R = 5,
10, and 15; Ts = 5, R = 2, 4, and 6; Ts = 10 and 30, R = 2, 3, 4, and
5). Additionally, the effect of the meal weight threshold (WT)
used to determine start of a feeding event was evaluated by

compare the sequential differences Wk = Wk Wk1 to a

threshold weight to assess whether a candidate event has
occurred;
determine candidate event duration and end time from a
threshold based on Ts ;
assign candidate event activity as either feeding or nonfeeding using an ARMA of time-series weights for beforeand after-meal tare.

CO

200

Fig. 3 Representative high-frequency time-series weigh

scale recording (Ts = 0.1 s) used for algorithm tuning. Circles
represent candidate event starts from AL2.

UN

199

EC

207

a stabilized baseline feed weight from an arithmetic moving average (ARMA) of R consecutive readings is used to
determine whether the next Wk is the start of a candidate
feeding event;
use of a forward-based R-pt ARMA to determine meal
event cessation;
feeding event assessment using a 0.2 g threshold between
start and end weights;
automated handling of tare, when feed was added to the
system.

198

OF

175

Two algorithms (AL1 and AL2) were developed to utilize timeseries recordings of feeder weights as the basis for assessing
individual bird meal activity. Both algorithms were designed
to post-process large volumes of feeder weight recordings. AL1
was designed to handle high-frequency (10 Hz) time-series
recordings, whereas AL2 was designed for lower frequency
(1/301/4 Hz; or sample times Ts of 30 or 4 s). The frequency
criteria were dictated by the instrumentation systems used to
acquire the data, and offered a unique opportunity to assess
how well each algorithm performed, and the importance of
sampling frequency on determining behavioural attributes. A
representative ow chart of the main decision steps of AL2 is
presented in Fig. 2.
Each algorithm processes a series of feeder weight readings
taken at discrete times tk , denoted by W(tk ) = Wk , where index
k = 1,2,. . .n denotes sequential recordings taken at a sample
rate of Ts sec. For both algorithms, the following key elements
were used:

RO

174

Algorithm development

DP

173

2.3.

TE

172

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

Representative weigh-scale readings were used to assess algorithm performance. One sample (Fig. 3) was obtained from the
high-frequency data set (Ts = 0.1 s, 110 min total), and the other
from a full day of recordings (Fig. 4, Ts = 4 s) in which the hens
activities were recorded with time-lapse video. Each algorithm
was studied to assess robustness to tuning parameters, and
sampling frequency, as described below.

Fig. 4 Representative time-series weigh scale recording

(Ts = 4 s) used for algorithm tuning, representing a full
(05:0020:00 inclusive). Circles represent candidate event
starts from AL2.

Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

224
225
226
227
228
229
230
231
232
233
234
235
236
237

ARTICLE IN PRESS

COMPAG 2049 17

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264

265

266
267

3.

Results and discussion

3.1.

Algorithm AL1

OF

245

RO

244

Ts , number of points in the ARMA, R, and two different weight

thresholds. Complete results are available in Table 1 of Gates
and Xin (2001) and not repeated here. The actual change in
feed mass during this 110 min period was 5.3 g, and the tuned
algorithm from the full data set (Ts = 0.1 s, R = 100) yielded the
following baseline information: four meals totalling 5.2 g, with
mean meal statistics of meal size (MS) = 1.3 g, meal duration
(MD) = 89 s, meal interval (MI) = 1088 s, and meal ingestion rate
(IR) = 0.9 g min1 .
The effect of doubling R was to increase mean MD by 6 s,
and did not affect other statistics. For Ts = 1 s and R = 20, the
baseline case was closely mimicked, whereas for R = 10 (i.e. a
10 s ARMA) IR was over-predicted. For Ts = 2 s, R = 15 produced
results similar to the baseline (MD was 8 s greater) and both
R = 5 and R = 10 matched actual feed used. At Ts = 5 s, a 6-pt
ARMA produced results most similar to the baseline except
that it missed one meal.
The effect of using a smaller meal weight threshold WT
(0.2 g vs. 0.5 g) was as follows. For shorter Ts , the shorter ARMA
gave similar MS but one extra meal and thus shorter MD and
MI, and greater IR. This trend of counting additional meals was
also noted at greater Ts , and combinations in which Ts R = 30
gave results most similar to baseline whereas Ts R = 10 overpredicted total feed used. Results for (Ts = 10, R = 4) also closely
matched baseline results. Surprisingly, Ts = 30, R = 2 predicted
total feed, MS and MD quite well.
The opportunity for 1020-fold reduction in sampling frequency when using AL1 thus appears reasonable. Attempting
to improve sensitivity by decreasing the threshold to 0.2 g creates erroneous counts of additional meals noted and changes
meal statistics. This latter point underlines the importance
of devising a set of denitions for what constitutes a meal.
Once dened, AL1 with reduced sampling interval time-series
data shows great promise in automating feeding activity
analysis.

DP

243

TE

242

The algorithm was tested on the representative feed sequence

in Fig. 3, and evaluated for different values of sampling interval

EC

241

Table 1 Hourly summaries of bird behaviour from video recordings (hen 4, thermoneutral conditions)

RR

240

performing the analysis at 0.5 g (original tuned value) and at

0.2 g. The effects of tuning these parameter combinations was
assessed from computed meal size (MS (g)), meal duration (MD
(s)), time between meals (MI: meal interval (s)) and ingestion
rate (IR (g min1 )). Performance of AL2 when analyzing the
lower frequency IBU data was also evaluated.
Parameters for AL2 that were adjusted included weight
threshold (WT (g)), end of event threshold (EET, no. samples),
and the number of points in the ARMA (ARMAnpoints). A comparative assessment of AL2 versus AL1 was made by analyzing
the same high-frequency (and decimated subset) data (Fig. 3),
and lower sampling frequency data obtained from the IBU
(Fig. 4). Observations of bird behaviour taken from time-lapse
video of the IBU data were used to direct the tuning of AL2
parameters.
AL2 was developed with a different set of criteria than AL1,
namely to identify both feeding and non-feeding activities
from lower sampling frequency data. Thus statistics on candidate events, i.e. activity at a feeder that may or may not
be feeding, were gathered. To discriminate feeding events, an
event key code (0 or 1) was assigned for each event by comparing feed disappearance for each event to the meal weight
threshold. Total feed consumed was obtained from a dot product between the event key code array and individual feed
changes for each candidate event. The sum of entries in the
event key code yields the number of meals, and other statistics
such of mean MS, MD and IR were computed.

Time (hh:mm)

5:00
6:00
7:00
8:00
9:00
10:00
11:00
12:00
5:0012:00 Total (%)
14:00
15:00
16:00
17:00
18:00
19:00
20:00

Number of feeding
events

Number of drinking
events

5
3
7
2
7
4
5
8

0
4
4
2
4
3
2
2

41

21

7
1
8
6
5
10
5

5
0
6
6
3
8
4

42

32

CO

239

UN

238

14:0020:00 Totals (%)

Total time (s) at

Feeder
160
350
550
498
950
528
572
507
4115 (14.3)
762
57
598
652
721
1425
620
4835 (19.2)

Waterer
0
221
191
174
248
208
114
153
1309 (4.5)
198
0
186
278
164
446
129
1401 (5.6)

Other
3441
3036
2857
2927
2403
2863
2911
2936
23374 (81.2)
2645
3543
2821
2672
2715
1573
2850
18819 (74.7)

Feed consumed
(g)
3
5.8
6.5
7.4
9.6
3.5
6
3.5
37.9
7.9
0.8
7
6.9
9.8
16.5
7.9
56.8

Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301

ARTICLE IN PRESS

COMPAG 2049 17

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

Table 2 Summary of results using AL2 on the data in Table 1, with parameters: WT = 0.5 g, EET = 12 and ARMAnpoints = 3
Total feeding time

5:0012:00
(s)
%
Error (%)

4436
17.6

3724
14.8
9.5

186

14:0020:00
(s)
%
Error (%)

5512
21.9

4760
18.9
1.6

176

Totals
(s)
%
Error (%)

9948
19.7

8484
16.8
5.2

362

3.2.

Mean meal duration

Algorithm AL2

3.3.

311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332

335
336
337
338
339
340
341

TE

310

EC

309

RR

308

CO

307

UN

306

39.5

621

53.5

1388

93.0
1.8

with Ts = 4 s. A plot of the time-series of weights is given

in Fig. 4.
For the rst time period, AL1 with R = 5 and a weight threshold of 0.2 or 0.5 g, detected respective values of total feed
consumed of 41.8 and 41.7 g, and 3180 and 2924 s spent eating (12.6% and 11.6%, respectively vs. 14.3% observed). During
the second time period, total feeding time was 4636 s and
4240 s (18.4 and 16.8% of total time) for the two values of WT,
respectively, compared to 18.9% observed; and predicted total
feed consumed was 55.8 g, compared with 56.8 observed (2%
under-estimate).
Table 2 presents summary results from applying AL2 to
the sample days data. Selected parameters for AL2 were
WT = 0.5 g, EET = 12 and ARMAnpoints = 3. Total period feed
consumption was 93 g compared with 94.7 g obtained by subtraction from the stored records. Predicted time at feeder
(Total Event Time) was 17.6% versus 14.3% noted from video
recordings, and 21.9% versus 19.2%, for the two time periods,
respectively. MD and MI averaged 362 and 1388 s for the day,
with 20 and 27 meals noted for period 1 and period 2, respectively. For the entire period, 85% of time at feeder involved
eating, and the remainder was classied as stereotyped pecking. This statistic could not be checked against the video
recordings, but demonstrates the ability of AL2 to discern not
only meal size and duration, but also non-meal activity at the
feeder.

DP

334

305

767

5.8

333

304

Total feed consumed (g)

4.2

The algorithm was tested on the representative feeding

sequence in Fig. 3, and evaluated for different values of sampling interval Ts , meal weight threshold WT, end of event
threshold EET, and number of points in the ARMA, ARMAnpoints. Complete results are available in Table 1 of Gates and
Xin (2001) and not repeated here.
For the full data set (Ts = 0.1 s), AL2 mimicked AL1 predictions closely (with parameters: WT = 0.5 g; EET = 100, 200
or 300; and ARMAnpoints = 120). Both algorithms give similar meal statistics, however, AL2 estimated meal interval to be
400 s, whereas AL1 estimated it to be 1088 s (and it cannot be
stated which is more accurate).
A 10-fold increase in sampling interval (Ts = 1 s), and
reduced weight threshold (WT = 0.2 g) resulted in overestimated number of meals and associated feeding statistics.
However, increasing WT to 0.5 g provided good estimates of
meal events, and with EET = 8 and 12, the results from AL1
were bracketed. Further increasing EET to 30 appeared to provide the best match with AL1. The number of non-meal events
predicted were 31, 26 or 20 for these values of EET, respectively. For sampling interval, Ts = 2 s, total feed consumed was
underestimated (EET = 8, 12, 30), but total feeding time was
reasonably matched. For Ts = 5 s, a combination of WT = 0.5
and EET = 4 was best, although total feed consumed and feeding time were slightly over-estimated. At Ts = 10 s, consumed
feed was slightly under-predicted and total feeding time was
over-predicted, with EET = 4 or 8 giving best results. At Ts = 30 s,
predictions were not very consistent. From these results, it
appears that using a 30 s sampling interval with AL2 is not
advisable, but Ts = 1, 2, 5 or 10 s were practical with appropriate
tuning of WT and EET.

303

Mean meal interval

OF

302

Total event time

RO

Time period

Comparison to video data

Sample summary behavioural data taken from time-lapsed

video recordings are presented in Table 1. The data include
number of feeding and drinking events, and time at feeder,
waterer, and other. The data are summarized by hour,
and by an 8 h morning and 7 h evening period (5:0012:00
and 14:0020:00 inclusive, respectively) covering the daylight hours. Weigh scale data were simultaneously recorded

3.4.

Other effects

A low frequency band pass lter (pass band 1e2 to 5e0 Hz)
was applied to the high-frequency data, and the resultant ltered data were subjected to analysis with both algorithms.
AL1 provided nearly identical results with the ltered data at
Ts = 0.1 s, if WT was held at 0.5 g; however, six meals and a
total of 5.8 g feed consumption were predicted if meal weight
threshold WT was reduced to 0.2 g. AL2 over predicted number
of meals and total feed consumed.

4.

342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367

368

369
370
371
372
373
374
375
376

Summary and conclusions

Applications of results from individual bird activity at the

feeding station include development of a frequency of occur-

Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

377
378

ARTICLE IN PRESS

COMPAG 2049 17

c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e x x x ( 2 0 0 7 ) xxxxxx

381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400

402
403
404

Both algorithms, AL1 and AL2, could be tuned to provide

similar predictions, thus a wide range in data sampling frequency could be analyzed.
AL1 was developed for 10 Hz time-series recordings
(Ts = 0.1 s); however, it was found to robustly determine meal
size, duration and interval information up to Ts = 130 s.
AL2 was developed for lower sampling frequency data, yet
reasonably matched AL1 results for data taken at Ts = 0.1
and 1 s.
Both algorithms were capable of predicting time at feeder
with good agreement with observed video recordings. They
provide the additional benet of discrimination between
eating at the feeder, versus stereotyped pecking.

405
406

TE

407
408
409
410

413
414
415
416
417

Q2

418

Uncited reference
Persyn et al. (2002).

Acknowledgements

Funding for the research was partially provided by Ajinomoto

Heartland Inc., and a grant from USDA NRI Animal Health and

421
422
423
424
425

426

Abeyesinghe, S.M., 2000. Aversion of the domestic fowl to

concurrent stressors: methodology. PhD Thesis. University of
Bristol, UK.
Cook, R.N., Xin, H., 2004. Effects of cage stocking density on
feeding behaviours of group-housed laying hens. In: ASAE
International Meeting, Ottawa, ON, Canada (ASAE Paper no.
044004).
Gates, R.S., Xin, H., 2001. Comparative analysis of measurement
techniques of feeding and drinking behaviour of individual
poultry subjected to warm environmental condition. In: ASAE
International Meeting, Sacramento, CA, USA (ASAE Paper no.
014033).
Gates, R.S., Turner, L.W., Chi, H., Usry, J., 1995. Automated
weighing of group-housed growingnishing swine. Trans.
ASAE 38 (5), 14791486.
I.A., Pereira, D.F., Curto, F.P.F., Behrens, F.H., Carvalho, J.C.C.,
Na as,
Amendola, M., Mantovani, E.C., 2000. Proceedings of The XIV
Memorial CIGR World Congress. Determining the Broiler
Female Breeder Behaviour Using Telemetry, vol. 1, CIGR,
Tsukuba, Japan, pp. 10141018.
Persyn, K.E., Xin, H., Gates, R.S., 2002. Effects of beak trimming on
poultry ingestion behaviour. In: ASAE International Meeting,
Chicago, IL, USA (ASAE Paper no. 024070).
Persyn, K.E., Xin, H., Ikeguchi, A., Gates, R.S., 2003. Feeding
behaviours and pecking force of chicks with or without beak
trimming. In: ASAE International Meeting, Las Vegas, NV, USA
(ASAE Paper no. 034005).
Persyn, K.E., Xin, H., Nettleton, D., Ikeguchi, A., Gates, R.S., 2004.
Feeding behaviour of laying hens with or without beak
trimming. Trans. ASAE 47 (2), 591596.
Puma, M.C., Xin, H., Gates, R.S., Burnham, D.J., 2001. An
instrumentation system for measuring feeding and drinking
behaviour of individual poultry. Appl. Eng. Agric. 17 (3),
365374.
Wathes, C.M., Abeyesinghe, S.M., Frost, A.R., 2001. Environmental
design and management for livestock in the 21st century:
resolving conicts by integrated solutions. In: Stowell, R.R.,
Bucklin, R., Bottcher, R.W. (Eds.), Proceedings of Livestock
Environment VI. 2123 May. ASAE Publications, St. Joseph, MI,
pp. 514.
Xin, H., Berry, I.L., Barton, T.L., Tabler, G.T., 1993. Feeding and
drinking patterns of broilers subjected to different feeding
and lighting programs. Appl. Poultry Res. 2 (4), 365372.
Xin, H., Gates, R.S., Puma, M.C., Ahn, D.U., 2002. Drinking water
temperature effects on laying hens subjected to warm cyclic
environments. Poultry Sci. 81, 608617.

UN

419

CO

412

RR

EC

411

420

references

DP

401

Well-Being Program, and is acknowledged with gratitude. This

paper no. 01-05-111 of the Kentucky Agricultural Experiment
Station, reports results of an investigation by the Kentucky and
Iowa Agricultural Experiment Stations, and is published with
the approval of the Directors.

OF

380

RO

rence estimate, or histogram, for statistics including number

of meals per unit time, MS, MD, MI and IR. These statistics,
derived from autonomous real-time or stored and postprocessed data, can provide an objective basis for evaluating
bird response to environmental or management stimuli. The
algorithms presented in this work have been successfully
employed in a series of subsequent research trials (Persyn et
al., 2003, 2004; Cook and Xin, 2004).
With signal processing such as outlined in this paper, meal
size, frequency of occurrence, and a measure of bird stereotyped pecking at the feed station can be obtained. While the
proposed technique cannot be made to distinguish between
resting, preening or other activities away from the feeder,
it can be used to assess impact of environmental stressors
on feeding behaviour, and to determine how dietary manipulation may be utilized to counteract deleterious effects of
adverse environmental conditions.
Additional signal processing is recommended for improvement of predictions. For example, a properly congured
low-pass lter applied to the data may enhance meal size
determinations without degrading meal duration and interval
information. Use of a self-tuning black box linear model, calibrated and validated against each birds behaviour, could be
used for real-time assessment of dynamic response to external stimuli.
In conclusion:

379

Please cite this article in press as: Gates, R.S., Xin, H., Extracting poultry behaviour from time-series weigh scale records, Comput. Electron.
COMPAG 2049 17
Agric. (2007), doi:10.1016/j.compag.2007.08.015

427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472