Crochard et al. Supplementay materials: Buzzy bees: Improving the monitoring of pollinator activity in sunflower fields with continuous acoustic recording and deep learning
Abstract
Supplementary materials : Buzzy bees: Improving the monitoring of pollinator activity in sunflower fields with continuous acoustic recording and deep learning - Supplementary informations - 3 data sets: One is the database used to train the pollinator buzz recognition classifier One used to study the relationship between the number of honeybees counted per transect and the number of flying insects detected on acoustic recordings on the same day One used to study the relationship between the average number of buzzes per day detected during flowering season and the average number of honeybees per observation session - The best pollinator buzzes recognition classifier obtained after all the steps described in "Buzzy bees, quantifying bee activity within sunflower fields using acoustic monitoring and deep Learning" - 1 R script - 1 README
Full text
Supplementary information for: “Buzzy bees: Improving the monitoring of 1 pollinator activity in sunflower fields with continuous acoustic recording and 2 deep learning” 3 4 Ludovic Crocharda,*, Léa Maritona, Colin Fontainea, Mathilde Baudeb,c, Maxime 5 Raguéd,e, Didier Basf, Sabrina Gabad,e, Vincent Bretagnolleg,e, Romain Julliarda, Yves 6 Basa,h 7 a Centre d'Écologie et des Sciences de la Conservation (CESCO), Muséum national 8 d'Histoire naturelle, Centre National de la Recherche Scientifique, Sorbonne 9 Université, CP 135, 57 rue Cuvier 75005 Paris, France 10 b Université d’Orléans, F-45100 Orléans, France 11 c Sorbonne Université, Université Paris Cité, Université Paris Est Créteil, CNRS, 12 IRD, INRAE, 13 Institut d'écologie et des sciences de l'environnement de Paris, IEES Paris, F-75005 14 Paris, France 15 d INRAE, CNRS, Université de La Rochelle, USC 1339, Centre d’Etudes Biologiques 16 de Chizé, F-79360 Villiers-en-Bois, France 17 e LTSER “Zone Atelier Plaine & Val de Sèvre”, CNRS, Beauvoir-sur-Niort, France 18 f none (volunteer), France 19 g Centre d’Etudes Biologiques de Chizé, UMR 7372, CNRS & Université de La 20 Rochelle, F-79360 Beauvoir-sur-Niort, France 21 h CEFE, Université de Montpellier, CNRS, EPHE, IRD, Montpellier, France 22 * Corresponding author 23 24 Ludovic Crochard 25 43 rue Buffon – Campus Buffon 26 3 allée des Crapauds – Bât 135 27 CESCO – UMR 7204 28 75005 Paris 29 [email protected] 30 31
Supplementary information A - Setting up the audiomoths 32 Firmware version: 1.4.2 33 Sample rate: 16 kHz 34 Gain: Low-Medium 35 Sleep duration: 5 s 36 Recording duration: 900 s 37 Start recording: 6 am 38 End recording: 10pm 39 40
Table S1 - Overview of records. The percentage of loss is due to a hard drive failure 41 and sound recorders that fell to the ground 42 Field Theoretical number of records of 10s Number of 10s records we have Number of hours recorded Percentage of loss Fields included in the comparison between acoustic monitoring and honeybeecounts in field transects 12 152 460 147 220 409 3.5 X 116 151 740 125 194 348 17.5 X 376 132 840 131 966 367 0.7 X 1823 151 110 110 758 308 26.7 2187 131 310 130 279 362 0.8 X 2230 143 190 141 997 394 0.8 X 2939 180 450 179 194 498 0.7 X 3107 207 180 157 846 438 23.8
3253 148 860 147 657 410 0.8 X 3271 159 840 118 384 329 25.9 X 4566 141 840 140 346 390 1.1 X 5572 171 630 135 085 375 21.3 X 5595 148 680 147 448 410 0.9 X 5867 192 690 191 389 532 0.7 X 6223 169 470 145 255 403 14.3 7029 139 860 105 708 293 24.4 7413 156 870 127 159 353 18.9 7507 162 450 144 321 401 11.2 X 8690 133 380 90 503 251 32.1
9523 99 450 69 789 194 29.8 X 10106 183 510 92 117 256 49.8 10486 175 950 105 273 292 40.2 10626 95 310 86 702 263 9 X 12446 145 620 143 887 400 1.2 X 12473 121 770 69 586 193 42.9 13138 138 240 114 444 318 17.2 X 13253 159 480 158 204 439 0.8 X 14499 108 360 82 404 229 24 X 15339 106 650 71 567 199 32.9 17019 126 180 93 872 261 25.6 X 43
44
Table S2 - Number of sound events labeled per class of the training dataset A 45 Sound event class Number of events labeled Apis mellifera 198 Bombus sp. 20 Unidentified pollinator 7 Bird 9 Human voice 201 Car 216 Noise (i.e. mechanical noise or experimenter noise) 1077 Wind 124 46 47
Table S3 - Number of recordings sampled in E) Classifier testing 48 For score classes 5 and 6, it was impossible to obtain the theoretical number of 49 records in certain fields due to an insufficient number of detections, which explains 50 why the number of sampled records does not correspond to the theoretical number of 51 records for these classes. 52 53 Confidence score group Confidence score range Theoretical number of records per field Total theoretical number of records per confidence score group Number of recordings 1 0-0.5 7 210 210 2 >0.5-0.6 9 270 270 3 >0.6-0.7 11 330 330 4 >0.7-0.8 13 390 390 5 >0.8-0.9 17 510 501 6 >0.9 21 630 399 54 55
Table S4 - Number of sound events labeled per class for the training datasets B, C 56 and D 57 Sound event class Number of events labeled Example of sound event class sonograms (duration of each sonogram is indicated under them) Apis mellifera 198 5.3s Bombus sp. 20 8s Unidentified pollinator 12651 6.3s Bird 734 6.2s Turtledove and pigeon 625 5s