03750naa a2200661 a 450000100080000000500110000800800410001902200140006002400540007410000200012824501350014826000090028352018820029265000230217465000140219765000170221165000230222865000170225165000190226865000180228765000290230565000110233465000220234565000170236765000200238465000220240465000260242665000250245265300090247765300190248665300090250565300240251465300240253865300220256265300150258465300100259965300170260965300230262665300150264965300200266465300150268465300150269965300130271465300200272765300190274765300370276665300190280370000250282270000190284770000240286670000200289070000200291070000210293070000230295170000220297470000230299677300690301921241292021-06-28       2020    bl uuuu     u00u1 u    #d  a0378-11277 ahttps://doi.org/10.1016/j.foreco.2020.1183972DOI1 aFERREIRA, M. P.  aIndividual tree detection and species classification of Amazonian palms using UAV images and deep learning.h[electronic resource]  c2020  aInformation regarding the spatial distribution of palm trees in tropical forests is crucial for commercial exploitation and management. However, spatially continuous knowledge of palms occurrence is scarce and difficult to obtain with conventional approaches such as field inventories. Here, we developed a new method to map Amazonian palm species at the individual tree crown (ITC) level using RGB images acquired by a low-cost unmanned aerial vehicle (UAV). Our approach is based on morphological operations performed in the score maps of palm species derived from a fully convolutional neural network model. We first constructed a labeled dataset by dividing the study area (135 ha within an old-growth Amazon forest) into 28 plots of 250 m×150 m. Then, we manually outlined all palm trees seen in RGB images with 4 cm pixels. We identified three palm species: Attalea butyracea, Euterpe precatoria and Iriartea deltoidea. We randomly selected 22 plots (80%) for training and six plots (20%) for testing. We changed the plots for training and testing to evaluate the variabilityn, in the classification accuracy and assess model generalization. Our method outperformed the average producer?s accuracy of conventional patch-wise semantic segmentation (CSS) in 4.7%. Moreover, our method correctly identified, on average, 34.7 percentage points more ITCs than CSS, which tended to merge trees that are close to each other. The producer's accuracy of A. butyracea, E. precatoria and I. deltoidea was 78.6 ± 5.5%, 8.6 ± 1.4% and 96.6 ± 3.4%, respectively. Fortunately, one of the most exploited and commercialized palm species in the Amazon (E. precatoria, a.k.a, Açaí) was mapped with the highest classification accuracy. Maps of E. precatoria derived from low-cost UAV systems can support management projects and community-based forest monitoring programs in the Amazon.  aAerial photography  aArecaceae  aBiogeography  aEuterpe precatoria  aRain forests  aRemote sensing  aTropical wood  aUnmanned aerial vehicles  aAçaí  aAerofotogrametria  aBiogeografia  aEspécie Nativa  aFloresta Tropical  aPopulação de Planta  aSensoriamento Remoto  aAcre  aAerial surveys  aAmaz  aAmazonia Occidental  aAmazônia Ocidental  aBosques lluviosos  aDeepLabv3+  aDrone  aEmbrapa Acre  aFotografía aérea  aImagem RGB  aMadera tropical  aMapeamento  aPalm trees  aPalmeira  aRio Branco (AC)  aTeledetección  aVehículos aéreos no tripulados  aWestern Amazon1 aALMEIDA, D. R. A. de1 aPAPA, D. de A.1 aMINERVINO, J. B. S.1 aVERAS, H. F. P.1 aFORMIGHIERI, A.1 aSANTOS, C. A. N.1 aFERREIRA, M. A. D.1 aFIGUEIREDO, E. O.1 aFERREIRA, E. J. L.  tForest Ecology and Managementgv. 475, n. 118397, p. 1-11, 2020.