The Dry Chaco Forest Near Real-Time Deforestation Detection System

ESTEBAN ROITBERG; MERCEDES SALVIA; FRANCISCO GRINGS; PABLO PERNA; VERONICA BARRAZA

Maryland

Simposio; ForestSAT; 2018

University of Maryland

The Dry Chaco region (DCF) has the highest absolute deforestation rates of all Argentinian forests. The most recent report indicates a current deforestation rate of 200,000 Ha year−1 (0.85 % year−1). In order to better monitor this process, DCF was chosen to implement an early warning program for illegal deforestation. To fulfil the requirements, we chose to develop several near real time change detection models to identify abrupt changes in vegetation dynamics associated with deforestation events. The Near Real-Time Deforestation Detection System takes as input MODIS 16-day enhanced vegetation index (EVI) and/or 8-day land surface temperature (LST), with 250 m and 1 km of spatial resolution, respectively, enabling an early warning system to support surveillance and control of deforestation. Three models were developed: i) a temporal pattern classification model based on convolutional neural networks, ii) a bayesian change point detection model and , iii) a statistical model based on EVI and LST historical values for each pixel. The model (i) is based on the supervised classification of ?time series segments?, in order to estimate change point detection date using both previous and current values of EVI (2 month window period). Model (ii) is based on the identification of abrupt changes in the generative parameters of sequential data. Based on Bayes´ theorem, we can compute the posterior distribution to make online predictions robust to underlying changes in the time series. The time series was modeled as a set of 23 uniformly distributed and correlated random variables corresponding to each of the 16-day period of the year for EVI. The model is trained using historical data, in which data that are accepted are included in order to further refine the model. The last model (iii) relies on a semiempirical approach based on the fact that a deforestation event should produce a lower value than  the historical value of EVI and a larger value than the historical LST. Incoming values of EVI and LST are compared with their historical values and marked using a decision rule in order to be flagged as potential deforestation. Finally, the three models are combined using a simple voting approach. Results shows that deforestation was detected with a F-score of 0.78, and with a mean time lag not higher than 30 days combining all models.