Deforestation & Degradation Mapping
Tool 3: Utilising dense time-series of Sentinel-2 for continuous change monitoring and proxies of forest change
The reliable measurement of deforestation in tropical forests is an ongoing challenge, and remote sensing of forest degradation is still in a research phase. These challenges are particularly great in tropical dry forests, which have strong seasonal cycles, are highly heterogeneous and show large inter-annual variations. New data from the Sentinel-1 and Sentinel-2 platforms offer unprecedented resolution and frequency for freely available data, and may be capable of reducing uncertainty in rates of deforestation and degradation in tropical dry forests.
The SMFM DEnse FORESt Time series tool (deforest) will aim to use data from Sentinel-1 and Sentinel-2 to overcome some of the monitoring issues of tropical dry forests through analysis of dense time series of data. The tool will aim to be both useful for accounting of historical forest changes along with near real-time monitoring of forest change for management purposes.
Approaches for extracting information on forest change from dense time series are still experimental, with a range of methods that have been proposed. Three major challenges exist in classification of dense time-series of imagery from dry forests and woodlands:
Cloud cover is a problem in optical data, especially during the wet season;Cloud
Inter and intra annual variability in rainfall, tree-leaf display and grass growth leads to significant variations in land surface phenology, as does the occurrence of fire. Such changes can often create false positives in change detection algorithms
Dry forests are characterised by landscape heterogeneity, which means that the signal associated with deforestation and degradation is likely to differ across a landscape.
A recent paper by Reiche et al. (2018) offers some potential solutions to these challenges through multi-sensor integration and image normalisation. Reiche et al. (2018) used a multi-sensor dense time series of data (Landsat, Sentinel-1 and ALOS PALSAR) to generate near real-time estimates of deforestation in Bolivia. This method uses a Bayesian framework to flag forest changes to generate early warnings of forest change, and confirms them over time as more data becomes available. This method offers an elegant means of combining long term monitoring with shorter term alerts that can be actioned by forest managers.
For the purposes of SMFM project, this method has a number of advantages. Foremost is that there is no requirement for long time series of data to calibrate the model, a desirable property where using data from recently launched satellite sensors. Limitations of cloud cover are addressed through integration of multiple sensors, and the method includes an approach for removing the effects of seasonality from imagery (‘de-seasonalisation’) which is not reliant on prior knowledge of the seasonal cycle. Finally, the data processing and memory requirements are greatly reduced relative to time-series approaches such as BFAST, requiring each image to be loaded into memory only once, rather than building a profile of each pixel through time.
Firstly, the development of the tool will focus on the reliable detection of deforestation events, which, being of greater magnitude is a comparatively straightforward problem to solve. Secondly, following the collection of field data identifying locations of degradation (fieldwork planned for May-July 2018), the tool will be adapted to classify degradation events. Even if the higher risk goal of detecting forest degradation does not prove reliable, the tool will still provide novel utility that will be useful to the tropical dry forest monitoring community.
The prototype deforest tool has three steps:
Data download and pre-processing;
Bayesian detection of change events.
Data Download and Pre-Processing
The download and pre-processing steps for the deforest tool will use the SMFM sen2mosaic and sen1mosaic tools (section 2 and section 3, respectively). The cloud masking and artefact removal algorithms within these tools are particularly important as extreme pixel values may be interpreted as deforestation event. In terms of data processing and storage this is the most computationally intensive processing step, with the volumes of data required by the deforest tool necessarily large (approximately 1 terabytes for 100 x 100 km over 3 years).
The Bayesian change detection algorithm requires that all input images are converted into maps of the probability of forest (or degraded forest) being present. This transformation of images from multiple sensors into the same units allows them to be compared, and for progressive certainty for the detection of forest changes.
Classifiers of forest cover will vary widely, and in the same way that different countries will be unlikely to use the same classifier for land cover mapping, solutions will likely differ for forest probability estimation. Therefore, the scripts for calibration were separated from those for change detection, which will be usable independently of the classification method. Here is described the classification approach for converting input images into probabilistic maps of forest.
Feature Selection- Optical sensors have multiple inputs bands which contain information about land cover. For example, in optical data (e.g. Sentinel-2) the red and near-infrared spectral bands contain information on the green vegetation components. Spectral bands can be combined into spectral indices that highlight the property of interest (e.g. NDVI), in this case the presence or condition of vegetation cover. As well as highlighting vegetation, spectral indices have the property of mitigating against differences in scene illumination or topography.
Radar sensors do not have an equivalent of spectral bands, but many are equipped to transmit and measure radiation with different polarisation (orientation). For radar imagery (e.g. Sentinel-1) the tool uses input features of backscatter in each available polarisation and ratios between them.
Image normalisation- The SMFM deforest tool applies an image normalisation technique called ‘de-seasonalisation’ (Hamunyela et al. 2016) to input images, which aims to normalise for seasonal effects in imagery. In its original form this involves running a moving window over the image and taking the 95th percentile of an index (e.g. NDVI), and dividing by this value. The rationale is that this value should always represent a forest area (given a large enough window), and therefore is indicative of the phenological state of forests at that location. Reiche et al. (2018) used a simplified version of this method where, instead of a moving window, the 95th percentile of a baseline map of forest for each input image was taken, and subtracted this value from other pixels in the image.
De-seasonalisation is potentially powerful as it can be applied to images at any time of year without a long calibration period, it is robust to drought years, and with a moving window it should take into account landscape heterogeneity. Both the moving window approach of Hamunyela et al. (2016) and the forest baseline method of Reiche et al. (2018) will be implemented in the deforest tool, allowing their outputs to be tested and compared.
Logistic Regression- Logistic regression is used in the deforest tool to convert input features into a probability of forest cover. Logistic regression is an appropriate classifier as it returns well-calibrated probability estimates , and where input features are correlated the model can be made robust to overfitting through regularisation (which prioritises a ‘simple’ model to a more complex one). A separate model must be fit for each input sensor (Sentinel-1 and Sentinel-2), and different models for each acquisition mode (e.g. Sentinel-1 single polarisation, Sentinel-1 dual polarisation).
Each model will be calibrated with data from training regions representing forest and non-forest areas (or degraded and non-degraded) locations, with model predictions tested using a cross-validation process. The resultant model parameters are used to classify all input images, outputting an ordered series of GeoTiff files showing the probability of forest at each location in each image.
Bayesian Detection of Change Events
The Bayesian change detection approach of Reiche et al. (2018) is used by the deforest tool. The method takes a time-series of forest probabilities, combining new information on forest state with previous observations to generate updated probability estimates of forest being present in a pixel at each point in the time series (Pforest). Where an observation with < 50% probability of being a forest is encountered, a potential deforestation event is flagged. Pforestt is updated by future observations, leading to the eventual acceptance of the event where Pforest passes a user-specified threshold (depending on the degree of confidence required), or rejection of the change as a false positive where Pforest falls back below 50%. Observations are subject to a ‘block-weighting function’, which limits the probability of forest in an individual observation to the range 10% to 90%, preventing extreme observations from resulting in false detections of forest change.
The original module (available on github) is written in R, but by its design as an experimental system is computationally inefficient. The deforest tool improves the efficiency of the original script significantly by processing entire arrays simultaneously in place of the original pixel-by-pixel approach.
Outputs of the Bayesian change detection algorithm from the prototype scripts for monitoring and early-warning are shown in the Figures, below.
Outputs from the prototype deforest tool, showing ‘confirmed’ deforestation events in blue for the year 2017 in the vicinity of Moribane forest reserve in Central Mozambique, one of the locations selected to perform fieldwork.
Outputs from the prototype deforest tool, showing ‘early warning’ deforestation events in red for the year 2017. Many of these events will later be rejected as false positives, but this near real-time view is useful as a management tool.
User Interface and Usage
The deforest tool is provided as a Python module with a command-line interface. The command line tools can be operated in a similar fashion to those of the SMFM sen2mosaic and sen1mosaic tools.
Like image classification for land cover mapping, users will require a country-specific classification algorithm, which cannot be generalised. Therefore some knowledge and guidance is required to calibrate the tool for an individual location, whereas ongoing monitoring should be more straightforward once tools are calibrated.
Documentation and Distribution
The scripts for the deforest tool (under development) are available on bitbucket, including related documentation (under development).
Documentation and worked examples are available online: