Deep learning convolutional neural networks for Landsat-derived land cover mapping

Deep learning is an area of machine learning research that can be applied to image analysis and feature extraction. It uses neural networks (NNs) with many layers (hence deep learning) that can extract feature representations from data and, for example, undertake image segmentation into categorical classes as required to produce land cover maps from satellite imagery. A major advance in the accuracy of convolutional neural networks (CNNs) for computer vision image classification occurred in 2012, supported by advances in processing such as GPUs. This was subsequently adopted by the remote sensing community from 2014 for a variety of tasks including: image classification, segmentation, object detection, image fusion, change detection, building or road extraction from high resolution imagery and pan-sharpening. To date, almost 90% of studies have applied CNNs to higher resolution imagery (less than 10m resolution) with a majority of studies using imagery finer than 2m. Landsat imagery which is freely available globally enables long-term studies of changes in land cover. Such studies have been undertaken for over 40 years at continental and, more recently, global scales using a variety of machine learning methods. This project chose to use Landsat data for a number of reasons: its long time series enables estimation of land cover change over time; application of CNNs for land cover mapping using medium resolution Landsat data in Australia was limited; and benchmark land cover datasets built using performant pixel-based methods such as Random Forest (RF) decision trees were available for comparison. Label/ training data for model development supporting supervised machine learning were derived using current best-available national-scale government mapping including: Catchment scale Land Use of Australia (CLUM), Forests of Australia (FoA) from ABARES; and National Vegetation Information System (NVIS) from DCCEEW. CNN models were built using input annual Landsat 8 geomedian data for 2018 from Geoscience Australia to produce land cover maps for a study area in SE Australia, and compared to leading alternative machine learning techniques such as pixel-based RFs. The most accurate CNN model was applied for 34 years (1987-2020) using input annual Landsat 5, 7 and 8 geomedian imagery and tested for its spatial and temporal stability. A third experiment compared the capacity of CNNs and RFs to map natural vegetation and major forest types in the study area. Key findings include the ability of CNNs to detect land cover in the absence of accurate label data using proxy data (e.g., land use translated into land cover), superior spatial and temporal stability of CNN models compared to RFs, and the ability of CNNs to partially compensate for label data errors. Of broad-scale land cover and forest types mapped using CNNs, Eucalyptus/ Forest, Plantation and Grassland showed high accuracy (>80%), Built-up, Crop, Horticulture and Water as well as Callitris, Mangrove and Rainforest moderate accuracy (20-80%) and Bare, Acacia, Casuarina and Melaleuca exhibited low accuracy (<20%). CNNs are capable of broad-scale land cover mapping and change detection using Landsat resolution input data. CNNs produce coherent maps operating at a coarser spatial resolution than RFs, whereas RFs are better for fine-scale variations, for example, vegetation in urban areas. CNNs have potential for validation of existing national and global land use and land cover datasets. CNNs rather than pixel-based mapping approaches have sufficient temporal stability and spatial consistency for use in natural resource management at regional, continental and global scales and applications such as environmental accounting. This project has demonstrated that deep learning CNNs have the potential to exploit the long time series of Landsat data globally to produce robust broad-scale land cover maps of change over time.

Author dc:creator

• Boston, Tony

Language dc:language.iso

en_AU

Handle dc:identifier.uri

https://hdl.handle.net/1885/733716324

OAI identifier oai:identifier

oai:openresearch-repository.anu.edu.au:1885/733716324