Continuous land cover change detection in a critically endangered shrubland ecosystem using neural networks

Abstract

Existing efforts to rapidly detect land cover change in satellite image time-series have mostly focused on forested ecosystems in the tropics and northern hemisphere. The notable difference in reflectance that occurs following deforestation allow for unsupervised methods, often with manually determined thresholds, to detect land cover change with relative accuracy. Less progress has been made in detecting change in low productivity, disturbance-prone vegetation such as grasslands and shrublands, where natural dynamics can be difficult to distinguish from habitat loss. Renosterveld is a hyperdiverse, critically endangered shrubland ecosystem in South Africa with less than 5-10% of its original extent remaining in small, highly fragmented patches. I demonstrate that supervised classification of satellite image time series using neural networks can accurately detect the transformation of Renosterveld within a few days of its occurrence, and that trained models are suitable for operational continuous monitoring. A training dataset of precisely dated vegetation change events between 2016 and 2020 was obtained from daily, high resolution Planet labs satellite data. This dataset was then used to train 1D convolutional neural networks and Transformers to continuously classify land cover change events in multivariate time-series of vegetation activity from Sentinel 2 satellites as new data becomes available. These models reached a f-score of 0.93, a 61% improvement over the f-score of 0.57 achieved using an unsupervised method designed for forested ecosystems. Models have been deployed to operational use and are producing updated detections of habitat loss every 10 days. There is great potential for supervised approaches to continuous monitoring of habitat loss in ecosystems with complex natural dynamics. A key limiting step is the development of accurately dated labelled datasets of land cover change events with which to train machine learning classifiers.

Regional ecological forecasting across scales: A manifesto for a biodiversity hotspot

“…In South Africa, legislation has required the use of ecosystem threat status as a tool to guide conservation and land use decision‐making since 2004 (Botts et al, 2020). Remote sensing has revolutionised our ability to map and track land cover change with relatively standardised classification algorithms (Wulder et al, 2018), and has been used to develop at least one near‐real‐time change detection system for some ecosystems in the CFR (Moncrieff, 2022). However, existing remote sensing analyses in the CFR typically only provide information on complete habitat loss, and rarely tell us anything about the relative condition of the habitat that remains, hindering accurate ecosystem threat assessments (Skowno et al, 2021).…”