Integrating Geographic Object Based Image Analysis with the Open Data Cube for multi-resolution, multi-sensor hyper-temporal image analysis using CCDC for deforestation monitoring in Canada
by Wolfgang Lück (a), Masroor Hussain (a), Andrew Dyk (b), Stephanie Ortlepp (b), Christiane Schmullius (c), Sally Tinis (b)
(a) PCI Geomatics, 490 St Joseph Blvd. Suite 400, Gatineau, Quebec, Canada, J8Y 3Y7
(b) Pacific Forestry Centre, Canadian Forest Service, Victoria, British Columbia, Canada
(c) Friedrich Schiller University, Jena, Thuringia, Germany
Deforestation is a permanent land use change, caused by the direct human-induced conversion of forest land to other land uses. Land cover changes such as forest clearings due to harvesting as part of sustainable forest management, or due to natural disturbances, are not deforestation, as they are usually temporary and followed by forest re-growth. Current pixel based change detection techniques are sensitive to tree cover change and specific sensor characteristics and are, therefore, ill-suited for automated land use change classification. With a limited number of temporal observations, Geographic Object Based Image Analysis (GEOBIA) has been widely used for deforestation mapping in projects such as REDD+, but has not yet been implemented within a data cube architecture, such as the Open Data Cube (ODC). The combined use of hyper-temporal, multi-sensor, multi-resolution optical and SAR data within a data cube has also not been widely explored for deforestation monitoring.
The purpose of this research project is to use analysis-ready data (ARD) acquired by the Landsat Program and Sentinels 1 and 2 over Canada, to detect deforestation using the ODC with a GEOBIA data abstraction layer. Using both SAR and optical observations, the Continuous Change Detection and Classification (CCDC) algorithm is then applied on a per segment basis to characterise historic land use versus current land use, and to identify the resulting change date.
To generate a quality layer for the ODC, a pixel based spectral pre-classification delineating spectral classes such as cloud, haze, shadow, water, snow, dark bare soil, bright bare soil, herbaceous vegetation and woody vegetation, is applied to every image. A reference object layer is generated from segmented Sentinel 2 10m resolution imagery acquired in summer of 2017 by incorporating multi-date acquisitions until all segments have been formed from clean data. For the analysis, only segments are used which do not contain any cloud, shadow, haze, or snow pixels. For the time series analysis, features consisting of mean spectral values of pure pixels falling entirely within a segment are calculated for each available image date. In the time series for each segment, when a discontinuity in pixel based class mixtures is identified, the segment is split along the discontinuity boundary. In the ODC, once segments have been established across the full time series, and spectral/textural features have been calculated for each segment, CCDC is used to characterise the phenology of each segment. Abrupt changes in phenological behavior can be identified with given break points. When the land cover classes on each side of the temporal divide are identified, they can be used to determine if the non-forest land cover reverts back to forest or not. This will distinguish between land cover change and land use change. Only a change in land use is flagged as true change, and this can then be used to assist in deforestation monitoring.
The study area is situated in northeastern Alberta, Canada, and is historically subjected to frequent deforestation activities, mainly attributed to oil sands mining and fracking. Boreal forests comprise the natural vegetation of the area with dominant species being black spruce, jack pine, with some birch and popular. An active timber industry as well as natural factors such as fire introduce fluctuations in tree cover over the forest land use classes. Oil sand mining contributes to major deforestation events, whereas fracking leads to small scale deforestation. Using manual image interpretation techniques, the Canadian Forest Service has recorded deforestation in a systematic sampling design at frequent intervals for the small events, and by a complete coverage for large events. This highly accurate reference data is used to validate the performance of automatic deforestation mapping techniques developed in this study.