Continental-scale predicted ecosystem condition for Australia using deep learning

These spatial datasets represent a continental-scale approach to ecosystem condition monitoring using Earth observations and deep learning. This was undertaken for the years 2010, 2015, 2020, 2021 and 2022 for Australia at 100 m resolution. For each predicted ecosystem condition dataset (e.g. predicted_ecosystem_condition_2022.tif), a condition score of 1 indicates at or near reference condition, where a condition score of 0 indicates a fully degraded condition. Accompanying the predicted ecosystem condition datasets are QGIS colour schema (.qml). We also include datasets on spatial and temporal model sensitivity, including mean absoluate error (MAE) of predicted ecosystem condition by bioregions, as well as standard deviation of each pixels predicted ecosystem condition value over time, for all years mapped.\n\nThese datasets are associated with a manuscript currently undergoing peer-review.\nLineage: The methods used to generate these datasets are described in full in Owers et al. (in review). In this study we use an innovative deep learning architecture to pair time series satellite imagery to predict ecosystem condition across Australia for several years (2010, 2015, 2020, 2021, 2022) at 100 m resolution. Each dataset represents a continuum of ecosystem condition from close to reference condition (1) to fully degraded (0), where a pixel’s value corresponds to the prediction of ecosystem condition.\n\nThese datasets were generated using EO data and deep learning techniques. Our model was developed using 209,041 on-ground records of ecosystem condition, coupled with Landsat time series data and topographic and climatological datasets. Input data was collated from several sources including the harmonised Australia vegetated plot (HAVPlot) dataset for at or near reference condition (Mokany et al., 2022). Further input data considered as fully degraded condition were obtained by using features such as roads, parking lots and buildings from Open Street Map (OpenStreetMap contributors, 2017), plantation and vineyard locations (AARSC, 2021; ABARES, 2021), cultivated areas (Owers et al., 2022).\n\nEarth observation (EO) data were accessed using the CSIRO Earth Analytics Science and Innovation (EASI) platform. Spectral reflectance values of 6 reflectance bands (Blue, Green, Red, NIR, SWIR1, SWIR2) were extracted for each observation in specified calendar years and 17 common spectral indices were generated relevant for landscape and vegetation characteristics relating to ecosystem condition (NDVI, EVI, LAI, SAVI, MSAVI, TCG, NDWI, MNDWI, TCW, BUI, NBI, NDBI, TCB, BAI, NBR, BSI, NDMI). Additional nationally available spatial data layers were also used to provide topographical and climatological context. These included a digital elevation model (Farr et al., 2007) and topographic wetness index (Gallant and Austin, 2015), latitude and longitude, as well as climatology datasets including mean annual potential evaporation, mean annual precipitation, mean maximum temperature of the warmest month, and mean minimum temperature of the coldest month (Harwood et al., 2018).\n\nOur deep learning model architecture was inspired the temporal convolutional neural network proposed by Pelletier et al. (2019). The selected final model was applied to make predictions across the Australian continent for the calendar years 2010, 2015, 2020, 2021, 2022. EO time series data were generated using the same approach as input data for model training, including normalisation of data variables and interpolation of values where required in time series observations. The model was applied to Landsat data resampled to 100 m resolution.\n\nSpatial and temporal sensitivity of the model was evaluated across the Australian continent for the five calendar year maps. Spatial variation was mapped, stratified on bioregions, where mean absolute error (MAE) of predicted and actual ecosystem condition values were summarised. Temporal variation was mapped to evaluate predicted ecosystem condition values over time. This was achieved by mapping the standard deviation of each pixel for all years mapped. These sensitivity datasets are also included in this collection.\n\nThe datasets including in this collection are geotiffs of predicted ecosystem condition predicted_ecosystem_condition_.tif with an accompanying QGIS colour schema. Spatial and temporal model sensitivity datasets include a vector file for bioreogions (predicted_ecosystem_condition_MAE_bioregions.gpkg) and geotiff for standard deviation (predicted_ecosystem_condition_stddev.tif) with accompanying QGIS colour schema.\n\nAll geotiffs as continental mosaics at 100 m resolution projected in the Australia Albers coordinate system for Australia (EPSG:3577).\n\nReferences\n-- see attached document --