FOLU-Net: A novel framework using long short-term memory networks to predict future forestry and other land use

The objective of this study is to predict future tropical forest cover presence and types using multitemporal imaging spectroscopy data. Accurately predicting land cover changes with image time series is vital for assessing the effects of climate change and land management on forest resources. Data were obtained from the DLR Earth Sensing Imaging Spectrometer (DESIS) covering a region in the West Godavari district of Andhra Pradesh, India. DESIS, mounted on the International Space Station, records Earth observation data in 235 channels over a 400–1000 nm spectral range. Five overlapping cloud-free images were selected, capturing the seasonal variability among land covers to form the multitemporal image stack. 1070 randomly generated training points spanning the five dates were visually classified into four land cover classes: forest plantation, palm plantation, natural forest, and non-forest. Future land cover was predicted using the following steps: (1) A recurrent neural network with long short-term memory (LSTM) was used to predict future reflectance values of the 235 bands for each point. This model had an R 2 coefficient of 83.0 %. (2) A multi-layer neural network using Keras was trained on the classified points from each image with 5-fold cross-validation, achieving an accuracy of 73.0 %. (3) The classification model from step 2 was then applied to the reflectance data generated from the LSTM (step 1) to predict the future land type at each point. The combined land cover prediction framework, titled Forestry and Other Land Use Neural Network (FOLU-Net), enables predictions of land use change without the need for potentially error-prone land use classifications at each prior time step necessitated by approaches such as Markov chain analysis. Our findings demonstrate a robust framework for characterizing the evolution of land cover using multitemporal imaging spectroscopy. • Multitemporal imaging spectroscopy data from DESIS used to predict land cover change. • Long short-term memory networks successfully predict future hyperspectral data. • Deep learning provides robust framework for characterizing future land use. • FOLU-Net eliminates error-prone classifications per timestep unlike Markov analysis.