Improving global land cover fraction change mapping using temporal deep learning

Monitoring land cover changes becomes increasingly vital in light of climate change and urban and industrial expansion. Land cover, however, is often expressed in discrete categories, which can oversimplify transitional and heterogeneous areas. Instead, by mapping land cover types as fractions, one can make a more accurate representation of not only the spatial distribution but also the temporal variability, which is crucial for pronounced change detection. However, current methods produce inconsistent time series or oversimplify change dynamics. At the same time, the promising potential of temporal deep learning and dense time series modelling has not been previously explored. Bridging this gap, this study compares novel Long Short-Term Memory (LSTM) and post-processing LSTM (PostLSTM) models with baseline Random Forest (RF) and Markov Chain models for multi-class land cover fraction regression and post-processing, respectively. Landsat-8 data were used to produce fraction time series at a 100-meter spatial resolution in both annual and, in a novel approach harnessing intra-annual characteristics, dense (16-day) time steps. The models were trained and validated using yearly land cover fraction records collected from globally distributed sample sites. The LSTM models performed better than the RF models when focusing on land cover classes predicted or validated to have a fraction of at least 1%. However, they did not outperform because of poorer handling of the remaining 0% fractions, which are naturally inflated in number since the majority of classes are commonly absent from a given area. Since the RF models produced fluctuating time series, the PostLSTM models achieved the highest yearly Overall Accuracy (OA) ( > 67 % ). Compared to the Markov Chain models constrained by their Markov property, the PostLSTM models stabilised time series while preserving actual change. All dense models achieved similar yearly OA to their annual counterparts, but can offer end-users valuable information on intra-annual land cover fraction change. This research showcases the potential of temporal deep learning and provides important new insights into increasing accuracy in annual and dense land cover fraction mapping. • LSTM and post-processing LSTM models were tested for modelling fraction time series. • RF and Markov chain models were tested for comparison. • Time series were modelled in yearly and, novel, 16-day dense time steps. • PostLSTM models stabilised predictions, preserved change and had the highest accuracy. • Dense models had similar accuracy to annual models but model more detailed dynamics.