Patterns of disturbance dynamics within the Cerrado-Amazon Transition using time series data and Residual Neural Networks

The Cerrado-Amazon Transition (CAT) is the world’s largest tropical ecotone and separates the Cerrado Savannah from the Amazon Rainforest. Deforestation and degradation of large swathes of the dense Amazon rainforest and Brazilian Savanna is leading to irreversible transformation and a critical loss of biodiversity. An increase in wildfire and agriculture-led deforestation makes the CAT a dynamic ecological border within the internationally known ‘Arc of Deforestation’. Yet, our understanding of the impacts of deforestation and degradation in the CAT is hampered by a lack of knowledge as to where and when these disturbances occur. Here we combine time-series segmentation and deep learning algorithms to identify and quantify disturbances in the CAT over a 35 - year period. Using a combination of the Landtrendr algorithm, Landsat time series data and a Residual Neural Network (ResNet), we identified four different forest disturbance types (forest clearance, savannah clearance, forest wildfire, savannah wildfire) occurring within the CAT, based on their temporal spectral trajectories. Using our approach, we identified  more than 384,000 km² of disturbance between 1985 and 2020, with forest clearance accounting for the most significant proportion (35%) of the identified change. The accuracy of disturbance detection ranged from 88to 93%, while the accuracy of disturbance type classification reached ~ 95%, although disturbance events occurring within savannas are more difficult to identify, often due to lower initial vegetation cover. The greatest period of disturbance occurred between 1995-1998, due to increased agricultural activity.