Remote sensing of tropical forest degradation from selective logging

Remote sensing is the most accurate and cost effective way to monitor forests at large spatial scales. The preceding decade has seen incredible progress in accurate forest monitoring from space, with operationalized deforestation and fire alerts available in near-real-time globally. In contrast, methods for detecting and mapping forest degradation from selective logging have lagged behind; despite recognition that selective logging is a key driver of both deforestation and forest degradation. In this these I develop novel methods that utilize detailed spatial and temporal logging records to train machine learning algorithms to detect and map tropical selective logging. First, I utilized optical satellite data from the Landsat program and show that imagery acquired before the cessation of logging activities (i.e. the final cloud-free image of the dry season during logging) was best for detection, displaying a 90% detection rate (with roughly 20% commission and 8% omission error rates). Next, I tried extending this methodology to the detection of logging with synthetic aperture radar (SAR) data, but poor performance made logging predictions too uncertain. I go on to show that SAR data from Sentinel-1 display a distinct breakpoint in the time series of pixels logged under higher intensities (> 20 m3 ha-1) and could be used to detect more intensive selective logging within the Amazon. I then assess if combining optical and SAR data improve the detection of logging over the use of either on their own. I show that a combined model performs worse than optical data alone and including SAR data adds uncertainty that lowers model performance. Finally, I refine the optical approach developed in the beginning, generalizing the methodology to facilitate a large spatial and temporal scale assessment of selective logging. We create annual estimates of selective logging between 2000 and 2019 over the Brazilian state of Rondônia. I estimate that 41.0% of the State of Rondônia remained undisturbed forest through 2019, with 3.4% having undergone selective logging and 25.7% being deforested (with 13% Commission Error and 45% Omission Error over the twenty year period). In general, rates of selective logging were twice as high in the first decade relative to the last decade of the period. My results show improved access to data and technologies will enable advances in space-based forest monitoring and reiterate the value of free and open data access policies. Our approach is step in the direction of an operationalized selective logging monitoring system capable of detecting subtle forest disturbances over large spatial scales.