This paper describes a statistical modelling scheme for detecting vegetation cover change events using Landsat time series data supplied by Digital Earth Australia (DEA). Vegetation cover change activities, such as urban development and mine building, often have great environmental impacts. Identifying timing and locations of such events is essential for environmental monitoring and land account audits.
Digital Earth Australia (DEA) is a series of data structures and tools which organise and enable the analysis of large Earth observation satellite data. DEA uses spatial and temporal metadata to link calibrated and standardised geoscience data sets from various sources to provide governments, individuals, and businesses access to over 30 years of Earth observation satellite imagery and related datasets over the Australian region.
This method of change detection in vegetation cover relies on two steps: noise removal and statistical time series modelling. Most remote sensing data is affected by noise, such as abnormal sensor reading, clouds, cloud shadows and intermittent surface water cover. To obtain accurate change detection results, such noise in the time series must be detected and excluded from subsequent data modelling processes. A noise detection algorithm is developed to classify each pixel in a time series into clear, cloud and cloud shadow categories. The algorithm identifies noise by comparing reflectance data to statistics of local time series neighbours, thus does not rely on predefined thresholds. The advantage of using such unsupervised approach is that the implicit thresholds are self-adaptive to various underlying landcover types. The result of this step is per-pixel per-observation noise detection, and observations that are flagged as cloud or cloud shadow are excluded from time series analysis in the second step.
The second step applies a statistical time series model to detect long term ground cover change. For our purpose, two time series of Landsat surface reflectance indices are created, of albedo (mean of 6 spectral bands) and the Enhanced Vegetation Index (EVI). A moving average time window is applied to bothtime series , with the window size depending on observation density, with the lengths of the windows ranging from 1 to 2 years. This creates a pair of adjacent moving windows that scans through each time series, and the differences of the average and the standard deviation of the pair of the moving windows are recorded. These statistics are then fed into a statistical change detection model, which outputs the locations and the timing of vegetation cover change events.
The proposed method was tested on several areas where known vegetation clearing events have occurred. Experimental results show that the method successfully identifies locations and timings of these events. Figures below show experimental results in 2 sites: mine development at Maules Creek, NSW and urban development at Gungahlin, ACT. The cool colours represent events occurring at an earlier date while the warm colours represent events occurring more recently.
Presented at the 22nd International Congress on Modelling and Simulation (MODSIM2017)
