Predictive Analyses of Deforestation from Foret-Des-Pins in Haiti via Parametric and Neural Network Techniques

Data extrapolation via one-dimensional (1-D) parametric methods and a recurrent neural network (RNN) are investigated to study the evolution of the deforestation of Foret-des-Pins in Haiti. Satellite images collected by the National Aeronautics and Space Agency (NASA) from 1984 to 2022 are first used to extract entropy and correlation indexes and form time series from which losses in green space areas can be monitored. The Burg algorithm and state space are selected as parametric approaches, and the long-short term memory (LSTM) is chosen as the RNN to tackle the objective. Two extrapolation methods are considered: a traditional one, where the forecast is conducted via a one-step predicting scheme, and a second one with looping where each extrapolated sample is fed back to the input data, as time evolves, to compute the next one. Two experimental scenarios are conducted to judge the practicality of the extrapolation schemes, test the performance metrics of the three techniques, propose which one is best suited in each scenario and make a projection about the future of the forest if any mitigation measures are not adopted.