forestry scientist with a wealth of expertise in old trees
In order to match the environmental benefits that an old tree with a crown circumference of 20 meters provides – such as air filtration
cooling and carbon storage – you need about 400 saplings
“The magnitude of this ratio also came as a surprise to me,” says Andreas Roloff
“But it emphasizes how much more we need to respect and care for the old trees in our environment and how the decision to cut them down
for instance to make way for construction projects
should not be made lightly.” Felled trees are required by law to be replaced by one to three saplings
this serves as a pretext only.
we can define eleven biological properties specific to trees which impact their life expectancy
The more of these a tree species can combine
the older it can theoretically become – if conditions
the tree is not cut down before its time,” explains Roloff.
Even the best of databases can’t help if trees are not properly cared for
forestry graduates and professionals in Dresden can also learn tree inspection and care from scratch in certification courses from the Deutsches Baum-Institut (German Tree Institute)
“We now have very exact provisions to ensure high quality
It is not uncommon for the seemingly cheapest methods
to turn out to be the most expensive in the long run due to the tree damage and follow-up costs they result in,” explains Andreas Roloff
With regard to the costs of planting new trees and the immense environmental benefits of old trees
it would be fair to ask what will end up costing you more paper
Andreas Roloff was able to witness an extremely unusual tree rescue thanks to professional care
is the oldest tree in Saxony and has been listed as a national heritage tree since October 2022
it was in danger of collapsing when its leaves started sprouting
eight arborists carefully pruned the ancient tree branch by branch in order to save it
“This is of course an exceptional situation and often unaffordable for regional municipalities,” said Andreas Roloff
it was an incredibly awe-inspiring and moving experience
the Collm Linden is now secured and equipped for the next decades of its life
Andreas Roloff (2023): Methusalembäume - Wie und warum können manche Baumarten 1000 Jahre alt werden
(Methuselah Trees - How and Why Can Some Tree Species Reach 1,000 Years of Age
Secondary Effects and Consequences.) Forstwiss
Andreas Roloff (2023): Inspiration Natur im Jahreslauf
Mentale Stärkung und Motivation durch bewusstes Erleben
Drawing Mental Strength and Motivation from Conscious Experience.) Verlag Quelle & Meyer
Andreas Roloff (2017): Der Charakter unserer Bäume
Ihre Eigenschaften und Besonderheiten Beschreibung
Their Properties and Distinctive Features.) Verlag Ulmer Eugen Verlag
are not responsible for the accuracy of news releases posted to EurekAlert
by contributing institutions or for the use of any information through the EurekAlert system
Copyright © 2025 by the American Association for the Advancement of Science (AAAS)
Dresden Magazin
Die Vielfalt von Dresden Elbland im offiziellen Stadtmagazin
Winter at Forstbotanischer Garten Tharandt is over
The park opened to the public again at the beginning of April
you can walk through forests as they used to be – and as they one day may become
You do not necessarily have to book a long-haul flight to visit the Rocky Mountains
they are just two stops away on the S-Bahn
warden of the Forstbotanischer Garten (Forest Botanical Garden) for the past 24 years
is waiting for us at the station in Tharandt
The 35-hectare site was planted in 1811 by Heinrich Cotta who ran it as a private forestry college
It has since been taken over by the Faculty of Environmental Sciences at TU Dresden
More than 3,000 different tree species grow here
only a few of which are actually native to Central Europe
We are taken by car to the other side of this small town
which was a fashionable spa resort in Cotta’s time (celebrity guests included the German literary giants Goethe and Kleist)
narrow bridge that divides the garden into two sections: the historic and the North American
the left turn will take us to Utah and Nevada
Pietzarka has travelled extensively in the USA
studying the forests there and collecting the seeds of local tree varieties
The subject of Pietzarka’s research is the forest of the future
That’s because forest terrains will have to adapt to climate change
Pietzarka and his colleagues at TU Dresden are therefore looking for alternatives to the trees that have traditionally dominated the German landscape
“We are studying the behaviour of domestic varieties
for example how they react to drought.” He currently has high hopes for some types of oak and ash which he believes could make the German forest future-proof
the group of trees in front of us looks sad and distressed
“Juniper which hasn’t survived the winter.” So not a candidate for the German forest of the future
A few hundred metres further on is a rocky crag
from the top of which the surrounding villages are visible
“Our Rocky Mountains,” Pietzarka explains with an ironic smile
but it offers a panorama of the different trees
The researchers know the origin of each individual one; after all
they have collected their seeds by hand in China
The locations of the mother trees are recorded in a database
“Each of these trees is a point on the world map,” says Pietzarka
A display board shows a picture of the Indian summer: the brightly coloured crowns of sugar maple
“That’s where we’re going,” says Pietzarka
they can spread very quickly; all it takes is for one beetle to land on the branch of a neighbouring tree
But if it has to fly a hundred yards to find the next tree of the variety
we pass a sign pointing to giant sequoia trees
The trees behind it are really pretty – but giants they aren’t
He collected the seeds in 1999 in the Sierra Nevada but didn’t plant them here until in 2003
It will be several decades before they live up to their name
Many of the results of the research being done here in the Forstbotanischer Garten will only become available to future generations
“A forest generation takes about 100 years,” says Ulrich Pietzarka
a cycle of ten or 20 years is nothing.” But if he didn’t have the patience
Forstbotanischer Garten Tharandt, Am Forstgarten 1, 01737 Tharandt
you both grew up in Florence and have been living in Dresden for many years now.…
OnlyOne: Molecular hocus-pocus at the bar First made famous by Spanish top chef Ferran Adrià at his restaurant
The event gets underway on Saturday 13th January with the individual sprints and continues on Sunday with the team sprints.…
Experience Dresden Elbland in 360-degree panoramas:More than 60 locations were recorded using drones and high tripods
Produced with funding from the Free State of Saxony within the framework of the Tourism Promotion Plan
Essenzielle Cookies ermöglichen grundlegende Funktionen und sind für die einwandfreie Funktion der Website erforderlich
Inhalte von Videoplattformen und Social-Media-Plattformen werden standardmäßig blockiert
Wenn Cookies von externen Medien akzeptiert werden
bedarf der Zugriff auf diese Inhalte keiner manuellen Einwilligung mehr
The world’s leading publication for data science
Sparse Convolution plays an essential role in LiDAR signal processing. This article describes how the sparse convolution works, which used a quite different concept and Gpu calculation schema compared with traditional convolution
In this article, the theory part is based on the paper "3D Semantic Segmentation with Submanifold Sparse Convolutional Networks" [1]. The implementation part, i.e. SpConv() is based on the paper "SECOND: Sparsely Embedded Convolutional Detection" [2]
the author made a more general discussion about sparse convolution
Convolutional Neural Network(CNN) has been proved very effective for 2D image signal processing
the extra dimension Z increases the calculation significantly
most of the voxels of the 3D point cloud are empty
which makes point cloud data in the 3D voxels often sparse signals
The question is whether we can only calculate the convolution with the sparse data efficiently instead of scanning all the image pixels or spatial voxels
regular image signals are stored as matrix or tensor
And the corresponding convolution was calculated as dense matrix multiplication
The sparse signals are normally represented as data lists and index lists
We could develop a special convolution schema that uses the advantage of sparse signal representation
the traditional convolution uses FFT or im2col [5] to build the computational pipeline
Sparse Convolution collects all atomic operations w.r.t convolution kernel elements and saves them in a Rulebook as instructions of computation
which explains how sparse convolution works
In order to explain the concept of sparse convolution step by step and let it easier to read
I use 2D sparse image processing as an example
Since the sparse signals are represented with data list and index list
2D and 3D sparse signals have no essential difference
such nonzero elements are also called active input sites
the input tensor has a shape [1x3x5x5] with NCHW order
The convolution kernel of sparse convolution is the same as traditional convolution
The dark and light colors stand for 2 filters
we use the following convolution parameters
The output of the sparse convolution is quite different from traditional convolution
The sparse convolution has 2 kinds of output definitions [1]
calculate the output sites as long as kernel covers an input site
The other one is called the submanifold output definition
the convolution output will be counted only when the kernel center covers an input site
Fig.4 illustrates the difference between these two kinds of outputs
A2 stands for the active output sites which are calculated from P2
A1A2 stands for the active output sites which are the sum of outputs from P1 and P2
Dark and light colors stand for different output channels
the output tensor has a shape [1x2x3x3] with NCHW order
which are similar to the input representation
Traditional convolution normally uses im2col [5] to rewrite convolution as a dense matrix multiplication problem
sparse convolution [1] uses a Rulebook to schedule all atomic operations instead of im2col
the input hash table stores all active input sites
Then we estimated all possible active output sites
considering either regular output definition or submanifold output definition
using the output hash table to record all involved active output sites
Thus which one should be the key depends on the program
This is the key part of sparse convolution
The purpose of Rulebook is similar to im2col [5]
which converts convolution from mathematic form to an efficient programmable form
Rulebook collects all involved atomic operation in convolution
then associate them to corresponding kernel elements
Fig.6 is an example of how to build Rulebook
Pₒᵤₜ has the corresponding output sites index
Then we collect the atomic operations from the convolution calculation process
consider the convolution process as many atomic operations w.r.t kernel elements
we record all atomic operations into Rulebook
In Fig.6 the right table is an example of the Rulebook
The first column is the kernel element index
The second column is a counter and index about how many atomic operations this kernel element involves
The third and fourth column is about the source and result of this atomic operation
The last step is the overview of the programming calculation pipeline
not like the sliding window approach but calculate all the atomic operations according to Rulebook
In Fig.7 red and blue arrows indicate two examples
In Fig.7 red and blue arrows indicate two calculation instances
The red arrow processes the first atomic operation for kernel element (-1
we know this atomic operation has input from P1 with position (2
This entire atomic operation can be constructed as Fig
the blue arrow indicates another atomic operation
The result from the red arrow instance and the blue arrow instance can be added together
the neural network calculates the sum of the linear transformation with activation function like f(∑ wᵢ xᵢ +b)
So we don’t really need the hash table for active output sites
just sum all of the atomic operation results together
The computation w.r.t the Rulebook can be parallel launched in GPU
So the sparse convolution is quite efficient because we do not need to scan all the pixels or spatial voxels
We only calculate convolution for the nonzero elements
We use Rulebook instead of im2col to rewrite the sparse convolution into a compact linear transformation problem
or we say a compact matrix multiplication problem
One additional computation cost is building the Rulebook
this building work can be also parallel processed with GPU
At last, I wrote a SpConv Lite library [4], which was derived from [2]. It was implemented by CUDA and packaged into a python library. I also made a real-time Lidar object detector [6] based on SpConv Lite
I wrote this article when my living city Dresden is under lockdown
Merry Christmas 2020 and Happy new year 2021
Step-by-step code guide to building a Convolutional Neural Network
A beginner’s guide to forecast reconciliation
Here’s how to use Autoencoders to detect signals with anomalies in a few lines of…
An illustrated guide on essential machine learning concepts
Derivation and practical examples of this powerful concept
Columns on TDS are carefully curated collections of posts on a particular idea or category…
The world’s leading publication for data science
and artificial intelligence professionals.