Fig 1.
Study area located in the UNESCO Kafa Biosphere Reserve in the Southern Nations, Nationalities and Peoples Republic (SNNPR) state of southwestern Ethiopia.
Biosphere Reserve zones and location of local expert disturbance reports (deforestation and degradation) and additional reference data (no-change) are shown. The locations of map tiles from Fig 9 are shown as boxes labeled A to D.
Fig 2.
Flowchart of methods used in this study.
Processes are shown as rectangles and data and results are shown as parallelograms.
Fig 3.
Photo evidence from local disturbance reports documenting deforestation (A) and degradation (B).
The location shown in panel A corresponds to the time series shown in Fig 4, and the location shown in panel B corresponds to the time series shown in Fig 5.
Table 1.
Spectral bands on the Landsat TM, ETM+ and OLI sensors.
Table 2.
Spectral indices used in this study.
Fig 4.
Time series over a deforested site for four spectral bands: SWIR2, NDVI, NBR and TCW.
The RLM-fitted season-trend model for each segment is shown as a dotted line. Local disturbance photo evidence for this site is shown in Fig 3A.
Fig 5.
Time series over a degraded forest site for four spectral bands: SWIR2, NDVI, NBR and TCW.
The RLM-fitted season-trend model for each segment is shown as a dotted line. Local disturbance photo evidence for this site is shown in Fig 3B.
Fig 6.
Flowchart demonstrating the iterative updating of random forest models.
Time of acquisition of local expert data (parallelograms) are shown in the box on the right hand side. In each phase, a subset of the local expert data were used for model calibration (grey), and another subset was used for model validation (white).
Fig 7.
Iterative calibration and validation of change classes.
Boxplots of random forest class probabilities for the deforestation (P(DEF)), degradation (P(DEG)) or no-change (P(NOCH)) computed for in situ data having DEF or DEG labels are shown for the training phase (top panel) and operational phase (bottom panel) of the monitoring activities. The model updating approach is shown in Fig 6.
Fig 8.
Importance scores (S) for each band based on overall accuracies and class accuracies for deforestation (DEF), degradation (DEG) and no-change (NOCH).
Fig 9.
Maps of deforestation and degradation at four sites.
The probability of deforestation and degradation are shown as red and blue colour maps, respectively. Local expert reports of deforestation (X) or degradation (+) collected between 2012 and 2015 are overlaid on the maps. The base images are SPOT5 images (band 2; 2.5m spatial resolution) acquired between 2009 and 2011. Dark shaded areas represent forest in the SPOT5 image, and light areas are non-forest land cover types (e.g. cropland or wetland). The locations of each tile (A to D) are shown in Fig 1.
Fig 10.
Random forest class probability histograms for deforestation (red) and degradation (blue) at each of the four sites shown in Fig 9.