Deforestation & Degradation Type Mapping
Tool 4: Developing a Method for Identifying Cause & Drivers of Forest Change using Earth Observation
In addition to identification of the locations of land cover change, managers of dry forests may also be interested in identification of the causes and drivers of land cover change. Attributes such as the size, shape and intensity of disturbance events may provide clues about the type of change, which may be used to classify change events by their cause. The aim of this tool is to identify the cause to individual forest cover change events detected by remote sensing analysis.
Classification of forest change type is experimental and has not yet been attempted in tropical dry forests. Here, a novel method to identify the cause of forest change events in remote sensing data products is outlined. It is noted that the novel and innovative nature of this research comes with some risk and uncertainty, with very little similar existing work available the SMFM team.
This tool will be developed after field data collection has been conducted.
This tool will have three components:
Fieldwork data collection;
Training a classifier and validation;
Prediction of change events.
Fieldwork Data Collection
This tool will require data from the field to calibrate and validate its predictions. It is common to use data from Google Earth to calibrate and validate land cover / land cover change maps (e.g. Collect Earth tool), but as it is not possible to determine the cause of forest changes or to reliably identify degradation even using high-resolution imagery, the tool will always require local field data to be calibrated.
Given the effort and cost associated with field data collection, the field data collected should be usable for multiple purposes. The objectives of field data collection (May-July 2018) will therefore be (in order of importance):
Objective 1: Obtain reference data that identifies locations of woody loss (deforestation and degradation), and their causes (e.g. logging, charcoal, agriculture etc.) to provide training data;
Objective 2: Collect validation data to test the Producer’s accuracy of the deforestation/degradation predictions of tool 2 and tool 3 (i.e. the probability of an event on the ground being detected by the remote sensing product);
Objective 3: Collect validation data to test the User’s accuracy of the deforestation/degradation predictions of tool 2 and tool 3 (i.e. the probability of an event identified by the remote sensing product being present on the ground).
The plan will involve stratifying the landscape into representative land cover types (e.g. tropical dry forest and woodlands), sampling to test and compare the tool’s performance in both land cover types. The main field activity will be searching for areas of recent (2017 or later) woody cover change. Upon identifying a change event, its extent will be recorded on a GPS unit, pictures will be taken, its cause will be identified and documented through consultation of local experts, and the time at which it occurred will be estimated. A number of change event measurements will be necessary for each type of change (n = 30), aiming to include as many different causes as possible. For some causes this will be relatively straightforward (e.g. smallholder agriculture), and others likely much harder (e.g. logging). This data will be used for objectives 1 and 2. Objective 3 is more problematic, as it should only be performed once the tools are finalised. Additionally it is much harder to travel to a specific event than to find one of many examples of forest change. For objective 3, methods will be trialled with the aim of informing future fieldwork efforts.
Training a Classifier and Validation
GPS tracks will be used to match change events in remote sensing datasets to their cause. The tool will extract a range of features from each change events (e.g. size, shape, intensity etc.), which will be used to train a classifier to identify the cause of change events in satellite data (Figure, below). A range of classification approaches will be tested, including machine learning techniques, with the eventual choice of classifier will depend on the field data and through quality assessment of the classifier. The algorithm will be validated as a cross-validation, where a proportion of the field data are held back for validation purposes.
Examples of individual change events identified using ALOS data in Mozambique (Ryan et al. 2014), which will be used to classify change types in remote sensing imagery.
Prediction of Change Events
The tool will use functions within Python to extract individual change events based on criteria of minimum extent and intensity. Functions will extract key features from each change (e.g. size, shape etc.), which will be input into the classification algorithm to identify its cause. The tool will output GeoTiff files for visualisation in addition to reports of the area attributable to each change. The classifier will be applied to the outputs of tool 2 and tool 3, but should be equally applicable to other commonly used remote sensing datasets (e.g. Hansen et al. 2013).
User Interface and Usage
The tool will be provided as a series of Python scripts operated from the Linux command line. For more advanced users there will be a Python interface, and detailed methodological documentation to assist in further development of the method in other programming languages (e.g. R).
Documentation and Distribution
Documentation will follow finalisation of the tool.