EVI Time-Series Breakpoint Detection Using Convolutional Networks for Online Deforestation Monitoring in Chaco Forest

Abstract

The Dry Chaco Forest has the highest absolute deforestation rates of all Argentinian forests (current deforestation rate of 150 000 ha yr-1, 0.85% yr-1). The deforestation process is seen as a breakpoint in the enhanced vegetation index (EVI) time series, associated with the change from a typical forest phenology pattern to something else (e.g., bare soil, pasture, and cropland). Therefore, to monitor this process, a near real-time time-series breakpoint-detection model is needed. In this article, we exploited the 18-year-long MODIS EVI time-series data to train a temporal pattern classification model based on convolutional neural networks. Model architecture parameters (optimizer, number of hidden layers, number of neurons, and so on) were selected using an optimization procedure. The trained model then tries to estimate the probability that a given 'time-series segment' corresponds to a deforestation event. The model was validated using in situ data derived from high-resolution images. Results are promising, since the model presents good performance for the validation data set [F1-score = 0.85, {fpr} = 0.0012 (of the order of the true deforestation rate), {tpr} = 0.8 , for a sample size = 50 × 10{3} ] and average performance in a yearly analysis (F1-score = 0.6, sample size = 1120 × 10{3} ). Model performance was studied using two diagnostic tools: activation maps and model ensemble error estimations. Results show that proposed model presents good extrapolation capabilities, but its maximum F1-score is bounded by error in the available data set (in particular, mislabeled deforestation events).