Thread by @PlayfulTechnol1

Thread: #livetweeting from #AIforSustainability

First talk: Melissa Chardet and Henk Jan Faber (PWC) - Environmental Footprint Insights

Providing employees with insights into work mobility CO2 emissions

PwC has a global commitment to net zero by 2030.

Mobility is the main cause of PwC's CO2 emissions (90%)

Colleagues perceive several barriers to lowering their footprint

Didn't know alternatives, thought working from home affected reputation, perceived productivity loss when travelling by train, mobility decisions made on project level by managers.

Forecasts CO2 emissions, travel costs, and travel times associated with teams emissions.

Provides personal and organisational insights.

Pulls data from existing systems to avoid burdening users.

Data sources: BCD (internal travel agency) -flight and rail use data.

Workday - HR data -travel to work distances

Alphabet - vehicle fuel.
IPower - travel to client sites

Collect data: Excel Sheets
1 line per trip

Cleanse data: extract relevant data in normalised form.

Calculate: KPIs include CO2 emissions, distance, number of trips. Calculated on individual and company wide level.

Visualise: dashboard. Breaks down CO2 emissions by travel type.

Question: Quick question. Do you handle PII using anonymization/pseudonymization methods, to ensure the privacy of the people involved?

Personal data filtered out early in the pipeline and not aggregated on company level.

Question: How this individualization of responsibility (by tracking the sustainable behaviour of employees) has impacted their satisfaction/happiness?

Answer: We don't want to put pressure on anyone or penalize them. Everyone has different demands on them.

Question: Do you plan to analyze if showing people this information has a causal effect on their CO2 footprint?

Answer: At the moment it's more insights and awareness. We will be able to analyse changes in behaviour over time, but haven't started yet.

Question: How do you cope with incompleteness of your data?

Answer: Started with transport data because most of it is easy to obtain.

Second talk: Otto Fabius - Impact of urban trees with mobile mapping and #AI

Sobolt - Innovation company

Impact of trees: CO2 absorption, reduce air pollution, water retention, shade and cooling.

Policy makers need hard numbers

Trees are assets
Inspections are already digital
Automate with AI?

Gather data => Recognize trees => automated analysis => digital delivery

Mobile mapping: 360 degree cameras, GPS, Lidar

Recognize trees:
What defines a tree?
Photos or point clouds?
Digital trees?

Rule based system would be very complex => deep learning approach.

Photo or point cloud
Photos have colour, point clouds are 3d. More occlusion in photos, location easier in point clouds, direct measurements from point clouds

Point cloud classification used to detect crowns and trunks. Created digital tree objects.

Locates trees with high accuracy.

Limited by cars : can only measure 90% of trees from road. Potential for other management systems.

Occlusion by parked cars can make it difficult to identify some trees.

From tree object, tree dimensions can be measured. For public safety, need to know whether branches are at safe height above paths.

Creates geodatabase, from which tree measurement can be accessed. Sustainability metrics can be measured.

Used published modes for environmental impact and safety.

@sobolt_ai

Question: Classification or Segmentation? If classification, how were the points assigned to a particular class?

Answer: Classification of points followed by rule based identification of objects.

Question: How does diversity of trees affect sustainablity metrics?

Answer: Tree species can mostly be identified and published models used?

Question: How does this compare to existing datasets, such as maps.amsterdam.nl/bomen/?
I assume this has been annotated manually, can you speculate on the biggest differences in annotations at the moment between your system and manually annotated trees?

Where possible to access trees, accuracy compares favourably to manual measurements. Pollarded trees usually misclassified.

Question: Trees do grow over time. So you need to acquire new data to reflect the current state of tree height on a regular basis. How have you designed your iterative process for data acquisition, deep learning , etc?

Answer: Legal requirement for urban trees to be monitored regularly in NL

Third talk: Aljen Viielink (Satelligence)
Monitoring deforestation with satellite data.

Machine Learning for classifying satellite data

Satellite data: more channels than typical image data, typically 8 or 32 bits. Often obscured by clouds.

Different bands at different resolutions.

Machine learning algorithms Random forests, XGBoost - tree based algorithms

Use ensemble methods to avoid overfitting of decision trees.

RF uses decision trees in parallel, XGBoost uses them in series (each tree fitted to residuals of last)

RF faster than XGBoost for this problem, similar accuracy (95%)

Once model is trained, can map deforestation over entire countries.

Other algorithms considered - Isolation forest, K-means clustering

Deep learning: needs much more training data, or more models.

But can be used to despeckle radar images.

Spatial images become much clearer.

QGIS : Threshold, frequency, colourmap

Understand data and choose best tool

@arjenvielink

Question: Hi Arjen, you mentioned the difficulty in using pre-trained models because those models require RGB input. Would it be possible to extend those pre-trained models with a small input module that is tasked with transforming satellite data into RGB values?

Answer: Usually reduce data to RGB, but RadiantEarth is producing satellite specific models

Question: What’s your hardware infrastructure to train your models in terms of GPU’s?
How long is your training cycle? (days, weeks, hours?)

Uses cloud services. Always start with a simple pipeline that you can debug quickly

What's the feature-set like? Are the features mostly continuous variables, nominal, binary, etc? Cheers

Intensity of different bands of the EM spectrum. All continuous data.

Question: Are you training the tree models on raw pixels? If so, then why not use deep-learning models?

Answer: Preprocessing and cleaning data is the most important step, before models are applied

Question: What is the impact you have made so far with your customers? How does your work influence supply chains? What are the plans for the future in that respect?

Answer: We have paying customers who are using the system to stop illegal deforestation in their supply chain.

Question for Otto @sobolt_ai: How long does it take using TreeTracker to measure all trees in The Netherlands? Could it be done by a single car or do you need multiple cars driving around all the time?

Answer: A few thousand driving days, but can only be used in summer

Question: what is the name of the classification algorithm you used?

Answer: KP-conf

Question: Did you try to use GAN to deal with partially obscured trees? Or Encoder-Decoder approach?
Like Pix2Pix or similar algorithm?

Answer: No, it would be interesting but probably not feasible.

@threadreaderapp #unroll

KPConv arxiv.org/abs/1904.08889
github.com/HuguesTHOMAS/K…

@arjenvrielink