High-resolution wall-to-wall land-cover mapping and land change assessment for Australia from 1985 to 2015

Computational and data handling limitations have constrained time-series analyses of land-cover change at high spatial resolution over large (e.g., continental) extents. However, a new set of cloud-computing services offer an opportunity for improving knowledge of land change at finer grain. We constructed a historical set of seven high resolution wall-to-wall land-cover maps at continental scale for Australia and analyzed temporal and spatial changes of land-cover from 1985 to 2015 at 5-year time-steps using Google Earth Engine (GEE). We used 281,962 Landsat scenes for producing median cloud-free composites at each time-step. We established a pseudo ground truth dataset and used a PCA-based outlier detection method to reduce its uncertainty. A random forest model was trained at each time-step for classifying raw data into six land-cover classes: Cropland, Forest, Grassland, Built-up, Water, and Other areas, using 49 predictor datasets and nearly 20,000 training points. We further constructed uncertainty maps at each time-step as a proxy of per-pixel confidence. The average overall accuracy of the seven 30 m-resolution land-cover maps was-93%. Built-up and Water areas displayed the highest user and producer accuracies (>93%), with Grasslands and Other areas slightly lower (similar to 82-88%). Classification uncertainty was lower in more homogeneous landscapes (i.e., large expanses of a single land-cover class). Around 510,975 km(2) (+/- 69,877 km(2)) of land changed over the 30 years at an average of similar to 17,033 km(2) yr(-1) (+/- 2329 km(2) yr(-1)). Cropland and Forests declined by similar to 64,836 km(2) (+/- 16,437 km(2)) and similar to 152,492 km(2) (+/- 24,749 km(2)) over 30 years, mainly converting to Grassland. Built-up areas experienced the highest relative increases, increasing from 12,320 km(2) in 1985 to 15,013 km(2) in 2015 (similar to 19.2%, +/- 3.1%). The sensitivity, i.e., proportion of pixels correctly classified as having changed, was over 96%, whereas the specificity, i.e., the proportion of pixels correctly classified as no-change, was over 68%. Numerous potential applications of these first-of-their-kind, detailed spatiotemporal maps of land use and land-change assessment exist spanning many areas of environmental impact assessment, policy, and management. Similarly, this methodological framework can provide a useful template for assessing continental-scale, high-resolution land dynamics more broadly.

Automated summary

The study utilized Google Earth Engine to construct historical land-cover maps for Australia and classified land-cover using a random forest model, finding significant decreases in cropland and forests over 30 years, while built-up areas showed a noticeable increase. Additionally, classification uncertainty was lower in more homogeneous landscapes.