Near Real Time Monitoring of Land Cover Using Satellite Imagery

Abstract

Accurate identification of historic and current land cover changes is vital for monitoring and managing the environment. Earth Observation satellite archives provide cost-effective and consistently measured time series of data at a scale
appropriate for land managers. Current time series change detection approaches require a data density not present in historic archives and can be computational complex, limiting application to large-scale areas. In light of these challenges, the ‘map-to-image’ approach was developed, requiring only a base map and a satellite image to identity change features (i.e. map-to-image change detection). The aim of the study was to assess and develop the ‘map-to-image’ change
detection technique as a component of a multi-class land cover monitoring system for Wales. The approach assumes that within a base map class area the pixel values (e.g. reflectance or backscatter) of the satellite image are homogeneous, therefore any land cover change features will exhibit different responses. Three methods to categorise these changes were developed using; histogram distributions, probabilities from repeated random forest classifiers, and outlier analysis. The approaches were applied to a time series of Landsat (5, 7, and 8) and Sentinel 2 data covering Wales, UK acquired between 1990 and 2017. They were initially tested and compared on coniferous forest change, performing well when identifying clear-felling and producing accuracies of 89.4-94.1%. When applied to a full land cover map change detection accuracies of 69.2 %-89.1% were produced. Outlier and histogram distribution-based approaches outperformed repeated random forest classifier probabilities in most cases. The method performed well when detecting large-scale changes for spectrally homogeneous classes. However, in cases where change vectors were spectrally similar limitations in accuracy were identified. The change map use case, and likely land cover change drivers and types, must be considered when selecting a change detection approach. The ‘map-to-image’ approach should be used when a computationally efficient, robust algorithm is needed to analyse large-scale high magnitude changes in a time series of data where the time period between capture dates varies greatly or data volume is low.

Date of Award	2021
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Andy Hardy (Supervisor), Pete Bunting (Supervisor) & Richard Jensen (Supervisor)