Forest cover is an indicator of species habitat and biodiversity that can be monitored effectively using satellite images. The benefits of using satellite images for large scale forest monitoring are that they are freely available globally and frequently updated, which reduces the need for extensive field data collection. Field data collection to monitor forest change can be prohibitively costly in many places around the world. A challenge of working with these images is missing data due to clouds, particularly in tropical regions where forest monitoring is essential. Existing methods for interpolating missing data based on only past observations, such as compositing, are effective for stable land cover but inaccurate for dynamic and substantially changing landscapes. In this talk I present joint work with Dr Kate Helmstedt and Distinguished Professor Kerrie Mengersen: our new machine learning method Spatial Stochastic Random Forest (SS-RF). Our method accurately interpolates missing forest and land cover under simulated forest clearing scenarios by taking spatial relationships in the landscape and past and current data into account to produce probabilities of land cover classifications. This is necessary because monitoring changing landscapes and modelling missing data are highly uncertain problems.
We found our SS-RF method detected different land clearing scenarios accurately, and importantly offers more accurate and robust estimates with associated uncertainty measurements not possible with traditional compositing approaches. This method has promise for use for other remotely sensed environmental monitoring cases, and has been presented to a UN expert group as a potential method to inform adaptive ground truth sampling plans.
University of Adelaide
Friday 14 October 2022, 4pm
Zoom link below (Passcode: 017349)
Statistics Across Campuses