Richard Thackway

Satellite imagery is the only feasible approach to annual monitoring and reporting on land cover change. Unfortunately, conventional pixel-based classification methods based on spectral response only (e.g., using random forests algorithms) have shown a lack of spatial and temporal stability due, for instance, to variability between individual pixels and changes in vegetation condition, respectively. Machine learning methods that consider spatial patterns in addition to reflectance can address some of these issues. In this study, a convolutional neural network (CNN) model, U-Net, was trained for a 500 km × 500 km region in southeast Australia using annual Landsat geomedian data for the relatively dry and wet years of 2018 and 2020, respectively. The label data for model training was an eight-class classification inferred from a static land-use map, enhanced using forest-extent mapping. Here, we wished to analyse the benefits of CNN-based land cover mapping and reporting over 34 years (1987–2020). We used the trained model to generate annual land cover maps for a 100 km × 100 km tile near the Australian Capital Territory. We developed innovative diagnostic methods to assess spatial and temporal stability, analysed how the CNN method differs from pixel-based mapping and compared it with two reference land cover products available for some years. Our U-Net CNN results showed better spatial and temporal stability with, respectively, overall accuracy of 89% verses 82% for reference pixel-based mapping, and 76% of pixels unchanged over 33 years. This gave a clearer insight into where and when land cover change occurred compared to reference mapping, where only 30% of pixels were conserved. Remaining issues include edge effects associated with the CNN method and a limited ability to distinguish some land cover types (e.g., broadacre crops vs. pasture). We conclude that the CNN model was better for understanding broad-scale land cover change, use in environmental accounting and natural resource management, whereas pixel-based approaches sometimes more accurately represented small-scale changes in land cover.