Q&A

Choices and Principles

Q: Why does Forest Foresight work on a 0.004°x0.004° (about 400x400m) resolution?

A: This was chosen together with our executive partners in Indonesia and Gabon as a workable scale. This means that it is not too large so that it cannot be easily covered in the field and not be too small so that there would be way too many predictions.

Q: Why was six months in the future chosen as the timeline?

A: This was chosen together with our executive partners in Indonesia and Gabon as a workable timeframe. Not too far in the future to be too vague to handle on, but not too short-term to not be able to make preparations for an intervention. We do produce preprocessed ground-truth data for one, three and twelve months as well for anyone who wants to build a model on those timeframes by using the open source path.

Q: How often are the predictions updated?

A: We make the predictions every month, in the first week of the month. So every month we make predictions for the next six months.

Q: Why is Forest Foresight only covering the region between 30° North and 30° South?

A: this has two reasons. The first is the coverage of the Early Warning System (the Integrated Alerts of Global Forest Watch) that we rely on to make our predictions. The second and more important reason is the huge biodiversity of the forests in this region and the urgency to preserve this area.

Open

Coverage of the Integrated Alerts of GFW

Q: In what regions is Forest Foresight doing its predictions on a high resolution?

A: We only make predictions in areas that have forest according to the GLAD classification of 2000 with all tree cover loss removed between 2000 and 2020. This is because GLAD-L (a Landsat-based Early Warning System) also predicts in newly grown forest, which can also be plantations. We from Forest Foresight focus on primary tropical forest in our deforestation predictions and thus do not take all Integrated Alerts into account by using this mask.

Q: Why are the Integrated Alerts used as an Early Warning System and not another system?

A: We deem Global Forest Watch to be an independent and reliable source. Also, we work together with the Wageningen University in the Netherlands to further improve our predictions and they are the developers of the RADD alert system which is integrated in the Integrated Alerts.

Q: Which alerts do you use from the integrated alerts based on the confidence level?

A: We use all alerts, regardless of confidence level . If we would not include the low confidence alerts in our predictions/labeled (or ground-truth) data then this would be a loss because most of these alerts turn out to be actual deforestation. We also do not want to exclude them from our training data because then the system would be trying to predict these low confidence alerts as well without having the full knowledge available from the Early Warning System. We do include the confidence level as a feature for training but deem this to be most fair.

Q: Why do you not only monitor illegal deforestation?

A: The definition of illegal is very hard to know without full local knowledge. Still then it can be vague and debatable, depending on the stakeholder. We find it important that whoever has the resources to prevent deforestation has all the information available and can then filter our predictions based on what they deem to find important.

Q: Why do you use the Integrated Alerts as ground-truth while they have errors themselves?

A: Since our predictions have to cover everything and over a large area it is better to have wall-to-wall covering of our ground-truth dataset and for a longer period of time than to have high-quality sparse data for a short timeframe and a small area. The underlying datasets of the Integrated Alerts have all shown to be accurate, especially with regards to precision/user’s accuracy, for over 90%.

Q: Why do you prioritize a high precision (certainty that a prediction actually takes place) over a high recall (predict all forest loss events)?

A: Given the work that goes into preparing a field intervention by our partners we have found that it is more important to be very sure about where deforestation will take place than to cover everything. Experience has taught us that most implementing partners do not have the resources to tackle the majority of the predicted forest loss

Q: Why do you say you prevent deforestation when it is actually forest degradation?

A: The two major reasons are that deforestation is a term well known to the general public. The second one is that it is really hard to prove with a Predictive Warning System based on an Early Warning System whether forest loss is permanent and whether the spatial scale means that it is complete or only part removal.

Q: Why do you call yourself a Predictive Warning System as opposed to an Early Warning System?

A: An Early Warning System is used as a general term in two instances: to alert before an event happens or to alert as soon as possible when an event has happened. Though the term Early Warning System is used way more often than our Predictive Warning System we have experienced that the term is mostly used to describe the latter (report as soon as possible after the fact). By changing the name of our System we want to differentiate ourselves from that.

Q: Can the predictions be used for monitoring purposes?

A: No, they are predictions and not events that have necessarily already happened. We suggest using an Early Warning System or yearly forest map classification for this purpose. We do develop a risk map along with our binary classification that can be used for risk analysis. However, we have no metrics on accuracy of this risk map, only on the binary classification that is done based on this risk map.

Q: Isn’t it so that with these predictions it could also help people who deforest to find the right areas of deforestation or to avoid patrols that would go to these areas?

A: This is a tough question to answer. We have written a document discussing this here and find that it is safe to say that there is a lot more positive expected impact than negative expected impact with having Forest Foresight Open Access and Open Source

Q: Can I use the tool for predictions of other phenomena than deforestation?

A: Yes! the software tool is data-agnostic, meaning that you can change the input features and groundtruth (on which it is trained) to something else and it will always try to predict the groundtruth with the input features. So you can make this anything you like as long as it follows the required data structure as explained here Preprocessing your own datasets .

Q: My question is not here! What to do?

A: Just get in touch! send an e-mail to ff@wwf.nl and we will respond as soon as possible

Technical Questions

Q: What hardware specifications are required to run the Forest Foresight package?

A: Since we use R we almost never have memory issues. Regardless we would advise to have at least 16GB of RAM to process an average-sized country (32-64GB for a country like Brazil). No GPU acceleration is used so not important. Any moderately sized CPU will do (i5-i7). Storage depends on the area size but for an average-sized country you need about 2GB of storage space. Note that this is only required for training, for predicting using an existing model you can use almost any computer.

Q: Why was R chosen as a scripting language for Forest Foresight?

A: In our experience it is very easy to package and install an R package. Other software packages, like python, tend to have more issues with package dependencies and non-flexible memory issues (as opposed to lazy loading in R). We wanted to prevent the need of containerization because this would also increase the required level of programming skills needed to run the package.

Q: Why does the Forest Foresight package not have a GUI?

A: We find it more important that the user has full control over the scripts to make it their own than to give the easiest experience possible. For non-technical people we provide the pre-made predictions monthly that can be downloaded with a GUI. Do note that it requires very little scripting skills to get started with building and running a model, as explained here.

Q: what are the predictive features of Forest Foresight?

A: We have quite a few! They can be found on this page.

Q: Do you need to have scripting skills to add new features?

A: No, our package was built data-agnostic. In our case this means that the process will fetch all data from the data input folder, which does not require changing any of the functions or code of Forest Foresight.

Q: What is the license of the Forest Foresight package?

A: At the moment it is distributed under GNU-GPLv3