Michael Marshall; Thomas Groen
The Role of Forests in Climate Change Mitigation and the Use of Multi-sensor Remote Sensing to assess Carbon / Advanced Image Analysis
Statistics, forestry and programming background/experience highly recommended
We need to know where and by how much tree cover has changed through time in order to quantify the impact of deforestation on the Earth System. Predictive maps of such information suffer from some or all of the following problems: (i) they do not do not go far enough back in time; (ii) they are categorical (not quantitative) in nature; (iii) they are based on two snapshots in time; and (iv) they have a very coarse spatial resolution. Since the turn of the century, studies increasingly incorporate large satellite image archives consisting of wall-to-wall coverage of surface conditions through time, high-performance cloud computing and data-driven methods to address these knowledge gaps. Several hurdles must be overcome before historical reconstructions such as these are fully realized, including how to: geo-reference and screen historical imagery for clouds, harmonize data from different sources through time, and test/train models.
The purpose of the study is to develop and test an automated technique to map the fraction of forest cover at 30-60m resolution from the late-1960s to present. Images will be acquired over a predetermined study area from the Landsat and Corona Spy Satellite image archives. The Landsat archive consists of Landsat 1-3 Multispectral Scanner (MSS), Landsat 4-5 Thematic Mapper (TM) and Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images in the visible to near infrared at 30-60m resolution from 1972-present. The recently declassified Corona images provide panchromatic coverage at high (<5m) spatial resolution from 1967-1972. The Landsat and Corona images can be acquired and processed on cloud-computing platforms, such as Google Earth Engine. A few options are worth exploring to geo-reference, cloud-screen and harmonize the imagery. Any number of machine-learning techniques, such as support vector machine, random forest and boosted regression can be adopted to reconstruct a tree cover time series from the harmonized dataset. Finally, existing tree cover benchmarks, such as the University of Maryland’s continuous fields of tree cover or ESA TREES products can be used to train and test the harmonized dataset and selected machine learning technique.
Dan-Xia S, Huang C, Sexton JO, Channan S, Feng M, Townshend, JR 2015 Use of Landsat and Corona data for mapping forest cover change from the mid-1960s to 2000s: Case studies from the Eastern United States and Central Brazil ISPRS. J. Photogramm. 103 81-92