Swiss Territorial Data Lab - geo Data Science - colloque FHNW
1. FHNW COLLOQUIUM
March 16, 2021
SWISS TERRITORIAL DATA LAB
APPLIED DATA SCIENCE
Raphael Rollier - Nils Hamel – Huriel Reichel
Adrian Meyer - Christian Dettwiler
2. Introduction about the Swiss Territorial Data Lab
4D Platform
Register of Buildings and Dwellings
Objects Detection
A Cantonal perspective
Q&A
Program
3. A co-creation
project
A way to share
and replicate
A space for
experimentation
Swiss Territorial Data Lab (STDL) is...
4. If you want to go fast, go alone. If you want
to go far, go together
STDL partners are…
6. What is the most effective way to find out the construction
date of buildings to complete the Building Register ?
How can I detect automatically changes in the field
in order to update the Land Register more rapidly ?
How can I improve the monitoring of solar energy
usage by detecting automatically panel installations ?
How can I monitor more effectively the
development of mining ?
The type of challenges we are exploring…
7. FHNW COLLOQUIUM
March 16, 2021
THE TIME DIMENSION
RBD COMPLETION RESEARCH PROJECT
Nils Hamel – Huriel Reichel
12. REGISTER OF BUILDINGS & DWELLINGS
Federal Statistical Office (OFS/BFS) & STDL
●
Federal Register
●
Missing buildings construction years
●
Automating construction years gathering
Two complementary research approaches
PROJECT
13. NATIONAL MAPS
Using swiss 1:25’000 national maps
●
Tracking the buildings on the maps
●
Detection of their appearence
●
Covering 2020 to 1950
APPROACH
21. STATISTICAL
Using statistical urban model
●
Workaround the lack of maps
●
Improve construction years approximation
●
Covering years before 1950
APPROACH
27. ●
National Maps Approach
84.7% within ±5.8 years
●
Statistical Approach
95 % within ±31 years
●
The importance of Time Dimension
Provides relevant information
CONCLUSION
29. STDL
OBJECT
DETECTION
FRAMEWORK
Generating a Model from
Cadastral Vectors and
Aerial Images
Predicting Objects in the
Same or a New Area of Interest
TRAINING
with Known Objects
INFERENCE
for Unknown Objects
BASIC IDEA
34. TRAINING
Ground Truth Generation:
Dataset Evaluation Split
80% Training
Used to Train Model Weights
10% Validation
Tuning Model Parameters
10% Test
Unbiased Assessment
Ground Truth Labels
44. Swimming Pools: Zoom Level Results
Zoom Level 15
≈ 480 cm/px GSD
16
≈ 240 cm/px GSD
17
≈ 120 cm/px GSD
18
≈ 60 cm/px GSD
19*
≈ 30 cm/px GSD
File System Load
1’949 Tiles
= 0.4 GB
4’216 Tiles
= 1.3 GB
12’817 Tiles
= 4.0 GB
42’990 Tiles
= 11 GB
154’861 Tiles
= 40 GB
Processing Duration
(Prep. / DL+Pred. / Postp.)
± 25 min ± 40 min ± 120 min = 2 h ± 240 min = 4 h ± 1’200 min = 20 h
Max. F1 Score on
TST Dataset
55.5 % 75.3 % 82.5 % 83.4 % 84.1 %
* Zoom Level 20 contains 583’014 Tiles = 179 GB and was expected to take ~120h, but aborted prematurely due to bandwidth memory errors
45. What’s Next for the
End Users?
Update Cadastre
Send Letters to Swimming Pool Owners
Think about Detectable Objects and
Contact Us!
46. Silage Bale Detector
Labeling Strategy
- Manually Digitizing 200 Stacks of
Silage Bales in QGIS
- Training a Preliminary Detector
- Use 300 Highest-Confidence
Predictions as New Labels
- Manual Correction and Filling In of
Complete Training Tiles
Result
700 Labels in 1.5 Days
Wikimedia Commons (2021)
47. New Elements: Silage Bales
Zoom Level 16
≈ 240 cm/px GSD
17
≈ 120 cm/px GSD
18
≈ 60 cm/px GSD
19
≈ 30 cm/px GSD
20
≈ 15 cm/px GSD
File System Load
8’000 Tiles
= 1.3 GB
25’000 Tiles
= 8.0 GB
84’000 Tiles
= 26 GB
310’000 Tiles
= 80 GB
1’310’000 Tiles
= 320 GB
Processing Duration
(Prep. / DL+Pred. / Postp.)
± 40 min ± 120 min = 2 h ± 240 min = 4 h ± 900 min = 15 h ± 6’000 min = 100 h
Max. F1 Score on
TST Dataset
52.5 % 74.7 % 87.2 % 92.3 % 90.9 %
50. SWISS TERRITORIAL DATA LAB
APPLIED DATA SCIENCE
Sicht eines Kantons: Thurgau
(Einbezogen in die Projekte 4D, Schwimmbäder, Ko-
Produktion Hoheitsgrenzen, Siloballen)
51. A Cantonal perspective
Potenziale aus Sicht eines Kantons: 4D-Plattform
Die amtliche Vermessung ist Referenzdatensatz!
Änderungen wirken sich auf viele «aufbauende
Geodaten» aus. => Information wertvoll
Datenmigration auf neue IT, neues Datenmodell:
Bleiben die Daten korrekt und vollständig?
Bauverwaltungen: Entsprechen die Änderungen
den neu bewilligten Bauten? Gibt es nicht bewilligte
Bauten?
SECTION
52. A Cantonal perspective
Potenziale aus Sicht eines Kantons: Objektdetektion
Schwimmbäder: Im TG eher sekundär, als Test der
KI-Tools wertvoll
Es gibt viele andere Potenziale: Bsp: Siloballen,
Echo des Landwirtschaftsamtes: Begeisterung
Die Objektdetektion sollte nicht auf das Thema
amtliche Vermessung eingeschränkt werden. Die
Kunden haben vielfältige Interessen, diese sollten
wir «abholen». Das stärkt unsere Position.
SECTION
53. A Cantonal perspective
Potenziale aus Sicht eines Kantons: STDL = Toolbox
Die STDL-Tools können mit etwas Phantasie und
vor allem mit Kenntnis der Kundenbedürfnisse breit
eingesetzt werden.
Leitsatz Amt für Geoinformation Thurgau:
Wir schaffen mit Geoinformation volkswirtschaft-
lichen Nutzen.
Die STDL-Toolbox hilft uns dabei.
SECTION
54. FHNW COLLOQUIUM
March 16, 2021
SWISS TERRITORIAL DATA LAB
APPLIED DATA SCIENCE
Nils Hamel – Huriel Reichel – Adrian Meyer
Raphael Rollier – Christian Dettwiler
info@stdl.ch
stdl.ch