Extraction of Phenological Features using Deep Convolutional Neural Networks and Random Forest models using MODIS EVI Time Series

FRANCISCO GRINGS; ESTEBAN ROITBERG; MERCEDES SALVIA; VERONICA BARRAZA

curitiba

Congreso; IUFRO 2019; 2019

IUFRO

Understanding the spatiotemporal dynamics of phenology is critical for forest monitoring and management. In South America, much of the effort in these studies with remote sensing data has been concentrated on tropical rainforests (i.e. Amazonas), while subtropical formations (xerophytic Chaco forests) have been given less attention, although they account for a large fraction of South American forests. In this study, we exploit the 18-year long MODIS EVI data to study time series similarity measures, and to use them to develop a methodology to classify areas with similar phenology over the Chaco Forest in Argentina. The idea to develop a phenology-based classification is not new; several models have been developed to estimate phenology metrics from NDVI time series, from simple linear smoothing methods to more complicated spectral methods (i.e. TIMESAT). Nowadays, machine learning algorithms provide a powerful tool to develop classifiers, even on data sets with poor signal-to-noise ratios or characterized by a high dimensionality. Since we opted for a supervised training approach, we defined six pilot areas inside every major ecosystem, homogeneous in terms of phenology (as seen by MODIS EVI). From each area we extracted EVI time series and used them to train classifiers based on Deep Convolutional Neural Networks and Random Forest. Then, we used these models to classify the whole study area. In this way, we obtained several annual maps of phenologically-similar areas, in which is it possible to study its spatiotemporal dynamics and it is also possible to recognize the anthropogenic impact, mainly related to deforestation.