Assessing The Temporal Dynamics Of Drivers Of Deforestation With Deep Learning In Southern Nations, Nationalities, and Peoples' Region, Ethiopia
By Hugo Poupard
The assessment of the temporal dynamics of land-use following deforestation, or follow-up land-use (FLU), can help countries to comply with REDD+ directives by monitoring the second most emitting source of greenhouse gases i.e., tropical deforestation, and its related land-use change. The traditional approaches to assess land-use dynamics show many limitations such as the inability to differentiate land-use from land-cover or to account for spatial heterogeneity, as well as performing poorly on large-scale analysis. To alleviate those issues, recent studies have aimed to develop deep-learning methods for thematic segmentation to classify FLU. However, no studies have been produced to assess the FLU temporal dynamics, which is key to identify more detailed and accurate patterns of deforestation. Therefore, in this work, we evaluate the potential usage of deep learning and high-resolution imagery to assess the yearly dynamics of FLU in the Southern Nations, Nationalities, and Peoples' Region (SNNPR), Ethiopia by using an Attention U-Net model developed by (Masolele et al., In review) and Planet & NICFI bi-annual images for tropical deforestation monitoring. We used the year 2014 from the Hansen Forest Cover Loss dataset as a baseline area for the analysis of FLU. From Planet data, we created yearly median-composite from 2016 to 2021. We histogram matched the yearly composites from 2017 to 2021 based on the 2016 yearly composite since it was noticed that Planet satellite sensors changed over the study period, as well as the data availability frequency. As a reference dataset, we sampled 65 locations and visually interpreted the dominant FLU of the region over the study period, resulting in a dataset of 390 samples. For validation, we use three types of F1-scores, namely macro-average F1-score, micro-average F1-score, and weighted-average F1-score. The validation phase allowed us to validate the use of yearly composite images on the deep learning model, to emphasize the gain in accuracy due to the histogram matching process, and to exclude yearly composite i.e., 2020 and 2021, on which the model was performing poorly. The results showed that the source data and the compositing process, as well as diverse physical factors such as climate variation over time, presence of clouds, and vegetation regeneration were influencing the performance of the deep learning model notably over time. The analysis of FLU temporal dynamics showed that the patterns observed in the produced graphs were odd. We demonstrated by gathering information about the results, such as the confusion matrices, the transition diagram, and ground truth examples that there was very-low, to no dynamics after the land-use settled in one location during the study period. This study gave promising results on the use of a deep learning model for analyzing the FLU temporal dynamics, while giving a comprehensive insight about the limitations of Planet & NICFI data in the context of land-use temporal assessment. It also highlighted the points to consider in developing a deep learning model, especially regarding training data.
Keywords: Tropical deforestation; Deep Learning; Follow-up land-use; PlanetScope; NICFI; Temporal analysis; Attention U-Net; Histogram-matching; Ethiopia