BATScan: A radar classification tool reveals large‐scale bat migration patterns

We constructed the classifier using data from 10 radar deployments, covering a wide range of habitats on central bird migration flyway over 7-year period, comprising ~18 million observations. analysed animal above Hula Valley, home to 30 species bats spanning 5–150 g in size and exhibiting variety ecological characteristics. distinguished bat-labelled echoes for training according phenology, morphology movement ecology bats, birds insects. Several non-bat datasets were joined train classifiers under increasing levels difficulty. Class imbalance resulting was handled generative adversarial network up-sampling much smaller bat dataset. The classification tool reached high level accuracy precision, further scrutinized with an extensive set validations. בשנים האחרונות השימוש במכ”מ לחקר בעלי חיים אוויריים איפשר לענות על שאלות אקולוגיות בסיסיות בזמן ובמרחב בפירוט חסר תקדים. יחד עם זאת, עד כה טרם ניתן היה להפריד בין החזרי מכ”ם מציפורים ומעטלפים, ובכך הוגבלה מאוד האפשרות לבחינת השפע, התפוצה והתנועה של עטלפים והשלכותיהם היישומיות הנוגעות להאבקה, הפצת זרעים ושמירת טבע. כתוצאה מכך, הידע שלנו קבוצה חשובה זו יונקים מוגבל מאוד. במסגרת פרויקט מחקר זה, פיתחנו את ה-BATScan שהינו המסווג הראשון שפותח אי פעם לזיהוי במידע מכ”מי. מסווג זה מבוסס בינה מלאכותית ונבנה תוך שימוש במאגר מידע ייחודי שנאסף ע”י אנכי שהוצב במוקד מחקרי החולה, בסמוך לאגמון החולה. תהליך איסוף המידע ובניית נעשו בסיס בתקופה בה הפעילות האווירית ציפורים בלילה היא מינימלית, בשילוב סינונים מבוססי צורה ותנועה. תצפיות בודדו זמני פעילות, ביומכאניקה תנועת כנפיים וגודל. BATScan מגיע לרמות דיוק גבוהות (מעל 90%) ועבר בהצלחה סדרת ולידציות חיצוניות. שימש ליצירה מסד נתונים בעולם כ-60,000 אשר נאסף באמצעות במכ”מים האנכיים שהוצבו בעשרה מקומות שונים בארץ לאורך שבע השנים האחרונות. מציג לראשונה דפוס תנועה מלא כל חודשי השנה. מצאנו כי היקף נדידת העטלפים מהווה כ- 10-25% מהיקף ציפורי השיר בישראל וכן שלנדידת יש פנולוגיה שונה מנדידת הציפורים. מתחילה מאוחר יותר ביחס לנדידת הציפורים בשתי עונות הנדידה, שעונות הנדידה קצרות לציפורים. בנוסף נצפה טווח הגבהים בו נודדים (200-600 מטר מעל פני הקרקע) צר מטווח שיר (100-1,000 הקרקע). יישום נתוני מאתרים נוספים צפוי לייצר עולמי לחולל מהפכה בהבנתנו האקולוגיה זו. aerial habitat holds both opportunities challenges science ecology. On one hand, airspace is practically limitless free obstacles, perfect arena in-depth, large-scale analysis (Kunz et al., 2008). It serves as countless organisms, which take part diverse processes (Diehl 2017). Many vertebrates invertebrates migrate, forage, compete, commute, mate even sleep while airborne (Brunton, 2018; Frick 2017; Rattenborg, other this out our reach, we must rely technological means thoroughly study it (Robinson 2010). To grasp largely unexplored vast airspace, depend machinery software detect, visualize, document analyse organisms inhabiting habitat. Aero-ecologists products these instruments, can either be obtained by tracking single individuals large areas long periods animal-borne devices, or constantly monitoring fixed volume air, thus innumerable short predetermined frame reference (Phillips 2019). Using radars detection started about 80 years ago (Lack & Varley, 1945; Plank, 1956; Tolbert 1958) widely implemented ‘Eulerian sampling’ aeroecological research (Schmaljohann, 2020). Vertical-looking (hereafter VLR) provide fixed, point view enabling long-term studies (Jeffries 2013). emits electromagnetic pulses records returning then produce multiple parameters regarding passing animal's shape, (Figure 1). detailed documentation all animals that pass radar's describes distribution flying time altitude. main limitation wildlife difficulty accurately classifying detected objects (Nohara 2007; Rosa 2016). Given amount modern VLR insights they are intended provide, manual individual generally impractical. In recent years, developments computational hardware have facilitated automatic machine learning algorithms (Rosa 2016; Yeşil 2019; Zaugg Separating biological targets has been impossible so far due their morphological behavioural similarities (Bruderer Popa-Lisseanu, 2005; Kunz 2016), often lumped classifiers, including those (e.g. This because supervised require large, independently verified algorithms, learn classify unfamiliar data. Training compiled accurate methods like human visual identification themselves (Niemi Tanttu, 2016) expert based prior knowledge experience (Zaugg Visual almost high-flying nocturnal assessment, may include behaviour, phenology morphology. A will enable Bats highly mobile play role pollinators, seed dispersers, insect pest regulators nutrient cyclers geographical temporal scales, making them ‘mobile links’, connecting distant components biosphere (Castillo-Figueroa, 2020; Gnaspini Trajano, 2000; Lacher Lundberg Moberg, 2003; Muscarella Fleming, Tremlett Voigt 2015). Despite importance, very little known different properties long-range movement. perform seasonal migrations thousands kilometres which, far, only studied Lagrangian sampling (single tracked periods, Bach 2022; Caprio Cryan 2014; Hutterer Lagerveld O'mara 2014, Petit Mayer, Russell Sullivan 2012; Villa Cockrum, 1962; Wilkinson 1996). These sparse, sporadic observations usually small number individuals, limited migration, preventing comprehensive, large-scales analyses (Fleming, Implementing approaches expected substantially advance understanding movement, behaviour Here use BATScan, algorithm developed, separate birds, inanimate reflectors first Eulerian (short-term area) description migration. specifically describe altitudinal migratory movements, deduced directionality. Our work focuses major avian where movements previously documented (Levin 2013), throughout full annual cycle. represents systematic comprehensive abundance anywhere world. present deployments BirdScan MR1 radars. processing scheme included novel deductive implementation AI tools. overcame challenge producing clean ‘Bat’ labelled dataset decades ornithological valley, patterns fauna region. isolated period 2 weeks when no takes place, hence activity negligible. list Israel at time, assessed contamination risk posed each Finally, applied rigorous filtering process ensure minimal tested several types 90% maintaining cross-classification contamination. Standard diagnostic evaluations complemented series validations realistic assumptions well-established classifier. used (1) validate bat–insect separation (2) assess false-positive rate diurnal detections (which misidentified place (3) additionally whether movement-related class correspond actual properties. result robust learning-based VLR. demonstrated application utility developed Israeli dataset, describing directionality (>60,000 observations) locations representing spectrum habitats. wider revolutionize aeroecology global scale. All processing, testing done R version 4.2.1 (R core team, 2022). methodology given Supporting Information. Data collected VLR, 25 kW, X-band, 9.4 GHz vertical-looking pulse system, detect bat-sized up ~800 m ground. Detected automatically characterized terms flight track (heading, altitude, speed), wing-flapping characteristics size. Each target also classified few primary classes, insects non-biological 2008; Information entire based, did not involve any contact handling permit ethical approval. Classifying requires hypothesis rationale general inability actually gather measurements positively identified flight. Wingbeat kinematics found differ (Riskin Tian 2006). Furthermore, wings mainly comprised living tissue water content, reflective substance targets, affect wing-related echo parameters. accordingly hypothesized should flyers wing reflection kinematic properties, proceeded base approach wing-flapping-related would effected such differences. Wings discernible effects reflections (Addison 2022), noticeable signature WFF, average width, length together PP), pause 1) characteristics, speed cross section (RCS, related size, Chilson Mirkovic interact differently (Ellington, 1991; Grodzinski 2009; Norberg Rayner, 1987; Pennycuick, produced four datasets, corresponding (1. Full, 2. No speed, 3. 4. PP) combinations (Supporting 3.1). ability calculate depends strength duration. Some calculated some targets; hence, parameter maximize usage. Machine techniques externally validated characterize classes multidimensional space. heuristics ‘Non-bat’-labelled 18 varying duration 7 Israel, diversity extreme desert agricultural wetlands Mediterranean ecosystems (Table S1; Figure S1). Most site located Research Centre, (35°43‣ E, 33°03‣ N), stopover migrants Eurasian-African (Collins-Kreiner 2013) regional biodiversity hotspot. visually acoustically identify aloft couple identifications signals. Here, filtered specific conditions ranges best understanding, contain exclusively. robustness against 260 during summer information ringing centre (https://www.birds.org.il/he/species-checklist). assigned its WFF published literature 2010), Rayner (1995) local body mass species. potential pattern (during non-migratory season), occurrence Valley region, well might pose above-mentioned ground-dwelling unlikely relevant altitudes (two owls, Athene noctua, Otus scops, hunt near ground, Eurasian stone-curlew, forages closely guards territories ground), swift (Apus affinis) monitored group project, suggesting night rare does vicinity radar. conclude certainty low. Bat-labelled 2a; 4.1) taken 2018 2021 consisted only, season directional night's vertbrate rayleigh.test function package ‘circular’, Agostinelli Lund, 2022) inactive night. removed signals ‘insect’, ‘aeroplane’ ‘flock’ classifier, generated kept 6–18 Hz represent inverse relation largest (Rousettus aegyptiacus) smallest (Pipistrellus pipistrellus) area, respectively, formula (2012) Rousettus aegyptiacus similar dimensions P. pipistrellus (Bullen McKenzie, 2002; Carpenter, 1986). increased upper limit ~25% accommodate foraging (Aldridge, 1987). Notably, still outside flapping frequency area estimations taxa, region (Schaefer, 1976). restricted RCS corresponds sizes 0.25 m2, aegyptiacus, area. measure (amount reflected radiation) object ~0 depending various technical factors, 3.1) but never exceeds threshold values. altitude basis 50 avoiding ground clutter, below 800 m, insectivorous (Shi 2021). ‘non-bat’ 2b; Table extracted daytime locations. proportion day easily distinguishable (small insects, raptors, etc.). aimed create accentuates cases distributions challenging instances Non-bat (in chosen parameters, passerine-sized possibly Lepidopterans) most important bats. However, naturally containing morphologies, etc. Accordingly, sub-datasets 4.2). sub-set restrictions, layered more samples represented disproportionality higher frequency, five times incidence 2). enabled separating similar, non-bat, targets. ‘bat’ form overall data, divided availability fit selected 2c). Classification sensitive relative introduced stage. ‘Non-bat’ severely imbalanced, outnumbering observations, heavily biasing (Tables 3; (GAN) balanced equal 4.3). GAN generates artificial original confronting two neural networks iterative gradually increases sample (Zhai generation sets ‘ganGenerativeData’ 10,000 iterations discrimination probabilities 0.95 (Goodfellow 2014). adequacy necessity implementing GAN, complex computer intensive, parallel up-sampled, up-sampled SMOTE (synthetic minority oversampling technique), considered standard imbalanced problems (Fernández 2018). bagged trees (BT) combination 5; 5). trained keeping 10% testing, ‘train’ ‘caret’ (Kuhn, 2008) ‘treebag’ method, ‘Accuracy’ metric cross-validation. GAN-generated test before algorithm's handle bias caused natural addition diagnostics, performed validation classification, evaluation class) indeed correctly (see Section 3.2). Bat allowed us examine presence proxy through compare analogous passerine (based classifier). nightly mean azimuth error taxonomic every between August November ‘circular’ (version 0.4-93; 6). spaces 3), initial exploratory analysis, successful bat/non-bat separation, space performing (balanced accuracy, arithmetic specificity accounts data-94.5%) without pulse-pause being least accuracy—85.8%; 4). Classifiers consistently superior procedure including: nature likely classification. excluding last hour sunset (Calculated ‘sunset’ bioRad package, Dokter 2019) locations: Sde-Boker, 262 km south Negev Highlands, within Israel. Notably previous stages construction, classifiers. Except Classifier 4 (No than 6% less half lack PP probably prevented passerines proportions decided forego results. supports proper birds. purpose step classifier's class. distinguishes reliable composition BirdScan's results revealed notable absence bat-classified except 3 WFF). lacks clearly susceptible points importance case consequently WFF) compared airspeed, deliberately chose preparation look spontaneous performance created separated process. radar-generated (average length, length/pause ratio airspeed) calculation (number pauses/duration signal). displays existence distinct Pulses pauses seem longer shorter (these do necessarily class, nonetheless comparable), pause, slower rate. suggests narrower airspeed makes sense light non-bats. faster line active propulsion medium. phenological 2007) explored time-series yearly Activity plots depict features pronounced taxa 4a). First, winter strong reduction activity, emigration hibernation, many reduced appearance evident taxa. Second, sharp increase mid-June arrival species, foremost Greater mouse-tailed (Rhinopoma microphyllum), emergence young resident born spring begin independent flights year conforms clear northwards months March June southward September November. appears shifted passerines, starting terminating later displaying delayed start autumn season. comparison (also 5), foraging) night, absent migrating passerines. mostly elevations 200 600 observed 5, bottom). reveals markedly diminishes elevation months, depicted complete edges bins. Based tests analyses, formulated protocol (no PP, respectively), night-time appropriate 4) observation missing applying consists ~430,000 (about 2% Israel), roughly 15% (63,630) objective aeroecology, what unique, continuously elusive inhabitants fine detail. Without it, merely scratch bottom habitat, billions supplier crucial ecosystem functions services humanity (Bauer An differentiate greatly promote practice conservation biology, management aspects critically assembly, implementation, diagnostics evaluations, adapting stage problem reflectivity, wingbeat para