Deep Learning Architecture for Forest Detection in Satellite Data

GABRIEL CAFFARATTI; MARTÍN MARCHETTA; RAYMUNDO FORRADELLAS; LEONARDO D. EUILLADES; PABLO A. EUILLADES

Río Cuarto

Congreso; XXV Congreso Argentino de Ciencias de la Computación; 2019

Red de Universidades Nacionales con carreras en Informática (RedUNCI) Dpto. de Computación ? FCEFQyN ? Universidad Nacional de Río Cuarto

Deep Learning algorithms have achieved great progress indifferent applications due to their training capabilities, parameter reduction and increased accuracy. Image processing is a particular area thathas received recent attention promoted by the growing processing powerand data availability. Remote sensing devices provide image-like datathat can be used to characterize Earth?s natural or artificial phenomena. Particularly, forest detection is important in many applications likeflooding simulations, analysis of forest health or detection of area desertification. The existing techniques for forest detection based on satellitedata lack accuracy or still require human expert intervention to correctrecognition errors or parameter setup. In this work a Deep Learningarchitecture for forest detection is presented, that aims at increasing accuracy and reducing expert dependency. A data preprocessing procedure,analysis and dataset composition for robust automatic forest detectionis described. The proposed approach was validated with real SRTM andLandsat-8 satellite data.