Machine and deep learning as tools for monitoring forest recovery from disturbance in the western United States

Changing disturbance regimes in the western United States have raised concerns that forests may not recover following disturbance. We used 20 years of land cover change data and 23 years of summer maximum NDVI to assess forest recovery. The land cover change data were used to identify where forest was lost early on in the 20-year period and the NDVI data were used to monitor recovery at disturbed locations. Two different target classifications were developed to categorize post-disturbance NDVI trends and convolutional neural networks were used to detect outliers – locations for which the NDVI response did not fit as cleanly into its assigned target class. Convolutional neural networks were trained on a 1% sample of the data and then used to score (classify) on the entire dataset. Agreement between the target (nominal) class and the modelled class was > 97% for both classifications. Nominal-modelled agreement and modelled class membership probability were used to identify outliers. We supported our use of convolutional neural networks by (1) evaluating performance of well-known temporal unsupervised clustering algorithms to identify native patterns in the temporal NDVI data and (2) through the development and demonstration of a GIS dataset of ecologically relevant indicators and map accuracy indicators to further inform temporal NDVI patterns. Temporal unsupervised clustering results (machine learning) did not match the nominal classification. The GIS dataset included patch size (a map accuracy proxy), fire severity, vapour pressure deficit (VPD), and a proxy for distance to seed source. Use of the GIS data showed that about 28% of the ~445,000 ha of disturbed forest experienced moderate or severe fire followed by dry post-fire conditions, and were greater than 90 m from a seed source.