Posted August 5, 2025 · Submitted by Amara Okeke, University of Cape Town
Accurate monitoring of coastal wetland extent is hindered by gradual, sub-pixel land-cover transitions that conventional classification methods fail to capture. We train a convolutional segmentation model to map wetland extent across four decades of medium-resolution satellite imagery. The approach resolves incremental conversion adjacent to expanding aquaculture and reveals accelerating wetland loss. We release the annotated training dataset to support reproducible, large-scale monitoring of vulnerable coastal ecosystems.
Trouble viewing? Open the PDF in a new tab.
Amara Okeke, Priya Nair, Liam O'Connor (2025). Mapping Coastal Wetland Loss with Deep Learning on Decadal Satellite Imagery. TerraNova preprint.
Amara Okeke, Priya Nair
Excessive groundwater extraction drives land subsidence that compounds relative sea-level rise across low-lying, rapidly urbanising deltas. Combining satellite geodesy with piezometric records, we map subsidence rates and attribute them to…
Read preprint →Amara Okeke, Sofia Reyes
Whether grazing management can durably enhance soil carbon storage remains contested. We report a three-year replicated field experiment contrasting continuous and rotational grazing on semi-arid grassland. Rotational grazing significantly…
Read preprint →