sustainacraft Co., Ltd. Joint research report by Sustainacraft, National Institute for Environmental Studies, and Hitotsubashi University was selected as Best Paper at NeurIPS, a difficult international conference in the field of AI and machine lea

Sustainacraft Co., Ltd.
Joint research report by Sustainacraft, National Institute for Environmental Studies, and Hitotsubashi University was selected as Best Paper at NeurIPS, a difficult international conference in the field of AI and machine learning
Received Best Paper at NeurIPS 2022 Workshop “Tackling Climate Change with Machine Learning”

Sustainacraft Co., Ltd. (Headquarters: Chiyoda-ku, Tokyo;
Representative Director: Hiroshi Suetsugu; hereinafter
“Sustainacraft”); National Institute for Environmental Studies (Location: Tsukuba City, Ibaraki Prefecture; President: Masahide Kimoto; Environmental Research Institute”) and Hitotsubashi University (Location: Kunitachi City, Tokyo, President: Satoshi Nakano, hereinafter referred to as “Hitotsubashi University”) are promoting joint research to create high-quality forest-derived carbon credits. increase.
Hiroshi Suetsugu and Keisuke Takahata from Sustainacraft, Shoichi Fukaya from the National Institute for Environmental Studies, and Shinichiro Shirota from the Center for the Promotion of Social Data Science Education and Research, Hitotsubashi University, will participate in one of the world’s top international conferences in the field of AI and machine learning. One of them, Neural Information Processing Systems 2022 (hereinafter “NeurIPS 2022”), presented a research report in the workshop “Tackling Climate Change with Machine Learning”, which was selected as the best paper of the workshop. [Image 1

research summary
In this study, we attempt to apply time-series causal inference methods to the evaluation of carbon credits derived from forest conservation projects. We are proposing a statistical model that uniformly handles both post-evaluation of projects from the viewpoint of dealing with junk carbon credits, which has become a problem in recent years, and forecasting from the viewpoint of dealing with initial financing problems of projects. increase.
Details and report materials can be viewed from the following URL. Keisuke Takahata, Hiroshi Suetsugu, Keiichi Fukaya, and Shinichiro Shirota. Bayesian State-Space SCM for Deforestation Baseline Estimation for Forest Carbon Credit. NeurIPS 2022, Tackling Climate Change with Machine Learning workshop.
In addition, we have created a commentary article in Japanese as follows. Research background and purpose
Carbon credits are an economic incentive system to promote projects that have additional effects on climate change countermeasures. Recently, many companies and organizations have set a goal of net zero, but carbon credits play an important role in offsetting the gap from zero emissions that remains after each entity’s efforts to reduce carbon emissions. is expected to fulfill Among them, deforestation and forest degradation control projects are considered one of the most effective approaches, and REDD+ (Reducing Emissions from Deforestation and Forest Degradation) is one of the frameworks to promote such efforts through credits. .
However, there are some criticisms of REDD+ derived carbon credits. Credits issued for projects that do not actually have a positive effect on climate change countermeasures are called “junk carbon credits”, but in fact many REDD+ projects may generate junk carbon credits. suggested in previous studies.
One of the key issues regarding carbon credits is the validity of the baseline (the amount of deforestation that would have occurred if the project had not occurred). It is possible to increase the amount of credits issued by setting a higher baseline, or because there are changes in the external environment that were not initially
anticipated after the project started (for example, changes in policies at the national level). , we often see cases in which initial assumptions are considered to be overestimations. Against this background, a concept called dynamic baseline has recently been advocated. Since the baseline is updated every time forest cover changes are observed after the start of the project, it is expected that the impact of changes in the external environment can be considered and corrected.
On the other hand, the financing issues often pointed out in carbon projects still remain. In principle, credits are issued after a certain amount of project performance has been observed, but in this case, the project owner has to wait several years until the first credits are issued. So financing is a big challenge, especially in the early stages. Also, from the investor’s perspective, project risk needs to be quantified for investment decisions, and in that sense, it remains important to improve the validity of baseline forecasts before project commencement.
From the above, we can see that it is necessary to handle both pre-project baseline predictions before the start of the project and sequential updates of the post-project baseline when observation data are collected after the start of the project in a unified manner. In this study, we proposed Bayesian modeling of new time-series causal inference as one approach to this problem.
In this research, we refer to a method called SCM (Synthetic Control Method) [1], which is one of the causal inference estimation methods for time series data, and a kind of Bayesian structural time series model called CausalImpact [2]. I proposed a model that treats the two points in a unified manner. The former method is only intended for ex-post evaluation after the project starts, but it includes an important mechanism in causal effect estimation of covariate adjustment. The latter, on the other hand, is relatively easy to extend to prediction because it is built on state-space models, but unlike SCM, it cannot take into account covariate adjustments. In the proposed model, CausalImpact’s state-space-based causal effect estimation model is extended to include prediction, and the equivalent of covariate adjustment in SCM is reflected in the posterior distribution using the idea of ​​generalized Bayesian updating. . As a result, while adjusting covariates, it is possible to make baseline forecasts based on past trends before the start of the project, and to update them using actual values ​​after the start of the project. Figure 1 shows an example of applying the proposed method to the Valparaiso project [3], one of the REDD+ projects in Acre, Brazil. From left to right: (a) Prediction using information up to the start of the project, (b) Baseline update/prediction using information up to 4 years after the start of the project, (c) Prediction up to 8 years after the start of the project It is a baseline update/prediction based on information. The vertical axis is the annual deforestation rate. The following can be read from this figure.
The baseline prediction confidence interval (blue area) is large when there is little data before the start of the project, but the interval gradually narrows each time the baseline is updated with more data after the start of the project.
The prediction confidence interval at each time point includes the posterior mean value (solid blue line) for at least 3 years, so it can be seen that the prediction works to some extent.
Since the project actual value (black solid line) exceeds the baseline post-event average, no clear project effect can be seen for several years after the project commencement (2011-2015), but after that the curve reverses up and down. This indicates that there was a small positive effect. This suggests that the project may have had a partial deterrent effect on the upward trend in deforestation rates across Brazil that has been reported since 2012.
Although we omit the details, it is also important to note that the estimates by the proposed method are within a realistic range compared to the baseline estimates based on existing methodologies.
As described above, the strength of the proposed method is that the prior baseline prediction and post-event evaluation, which were the background of the research, can be handled in a unified manner within the framework of Bayesian modeling.
[Image 2

Figure 1: Baseline forecast values ​​and ex-post evaluation values ​​after the start of the Valparaiso project (x-axis: year, y-axis: annual deforestation rate, solid black line: deforestation rate) Observations, solid blue line/blue area: posterior mean/90% confidence interval of baseline estimate, dashed black line: posterior mean of baseline estimate without covariate adjustment, project started in 2011)
[1] Alberto Abadie, Alexis Diamond, and Jens Hainmueller. Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program. Journal of the American Statistical Association, 105(490):493-505, 2010.
[2] Kay H. Brodersen, Fabian Gallusser, Jim Koehler, Nicolas Remy, and Steven L. Scott. Inferring causal impact using Bayesian structural time-series models. The Annals of Applied Statistics, 9(1):247-274, 2015 .
[3] The Valparaiso Project.
Direction of future research
The proposed method assumes the existence of forest parcel polygons, and estimates the baseline in the project area by taking each polygon as a unit of the control sample and taking their weighted average. Therefore, estimation results may depend on the properties of polygons. In addition, the output of the proposed method does not give any indication of how deforestation spreads geographically. We believe that one means of solving these problems is spatio-temporal modeling, which treats forest cover data not in units of polygons but in units of pixels, and describes their spatial dependencies. are currently running in parallel.
[Image 3d97343-8-08e76749d70aa00986ca-1.png&s3=97343-8-d861a38135f48782ec0bba30f13f889a-485x455.png
Figure 2: Pixels and Polygons. A pixel refers to each pixel in the original forest cover data (green is forest, white is non-forest). A polygon is a polygonal area (red frame), and in this research, we used forest plot polygons for each ownership called CAR published by the Brazilian government to obtain statistics of forest cover in each area. is calculated and used as the unit of analysis.
Another topic is JNR (Jurisdictional and Nesting REDD+) support. JNR calculates a baseline for each administrative district (national, state, prefecture, etc.) and allocates it to each project according to deforestation risk, but the evaluation is performed at the aggregated level for each administrative district. As a result, the
overestimation of the baseline, which was an issue for individual projects, and the problem of the (mainly negative) ripple effect to the surroundings called leakage will be greatly resolved, so JNR will become mainstream in the long term. is expected. It is expected that it will take some time for the project in line with the JNR to be implemented, but we believe that how to align the proposed method with the JNR framework will be an important research theme in the future. This research is supported by NEDO (JPNP14012).
About SustainaCraft
We will contribute to nature conservation by providing two solutions: inexpensive and wide-ranging carbon stock monitoring of natural resources using satellite remote sensing technology, and evaluation of complex reference levels and leakage unique to carbon credits based on causal inference technology. We are working to create a sound flow of funds. Currently, we are promoting the social implementation of this carbon stock monitoring technology in collaboration with multiple NGOs and business companies that are engaged in environmental conservation, mainly in Latin America and Southeast Asia.
Sustainacraft company profile
Company name: sustainacraft Co., Ltd.
Date of establishment: October 1, 2021
Representative Director: Hiroshi Suetsugu
Location: 8th floor, US Building, 1-6-15 Hirakawacho, Chiyoda-ku, Tokyo Website:
Inquiries regarding this release
○ Sustainacraft Co., Ltd.
○National Institute for Environmental Studies
Planning Department Public Relations Office
Phone: 029-850-2308
e-mail : kouhou0 (Please add “” at the end)
○Hitotsubashi University
Public Relations Section, Public Relations and Social Collaboration Section, General Affairs Department
Phone: 042-580-8032
e-mail : pr1284 (Please add “” at the end)
Details about this release:


%d bloggers like this: