Scientific Data,
Journal Year:
2024,
Volume and Issue:
11(1)
Published: Feb. 10, 2024
Abstract
Accurate
building
extraction
is
crucial
for
urban
understanding,
but
it
often
requires
a
substantial
number
of
samples.
While
some
datasets
are
available
model
training,
there
remains
lack
high-quality
covering
and
rural
areas
in
China.
To
fill
this
gap,
study
creates
high-resolution
GaoFen-7
(GF-7)
Building
dataset
utilizing
the
Chinese
GF-7
imagery
from
six
cities.
The
comprises
5,175
pairs
512
×
image
tiles,
573.17
km
2
.
It
contains
170,015
buildings,
with
84.8%
buildings
15.2%
areas.
usability
has
been
proved
seven
convolutional
neural
networks,
all
achieving
an
overall
accuracy
(OA)
exceeding
93%.
Experiments
have
shown
that
can
be
used
scenarios.
proposed
boasts
high
quality
diversity.
supplements
existing
will
contribute
to
promoting
new
algorithms
extraction,
as
well
facilitating
intelligent
interpretation
Remote Sensing,
Journal Year:
2024,
Volume and Issue:
16(9), P. 1505 - 1505
Published: April 24, 2024
Crop
mapping
using
satellite
imagery
is
crucial
for
agriculture
applications.
However,
a
fundamental
challenge
that
hinders
crop
progress
the
scarcity
of
samples.
The
latest
foundation
model,
Segment
Anything
Model
(SAM),
provides
an
opportunity
to
address
this
issue,
yet
few
studies
have
been
conducted
in
area.
This
study
investigated
parcel
segmentation
performance
SAM
on
commonly
used
medium-resolution
(i.e.,
Sentinel-2
and
Landsat-8)
proposed
novel
automated
sample
generation
framework
based
SAM.
comprises
three
steps.
First,
image
optimization
automatically
selects
high-quality
images
as
inputs
Then,
potential
samples
are
generated
masks
produced
by
Finally,
subsequently
subjected
cleaning
procedure
acquire
most
reliable
Experiments
were
Henan
Province,
China,
southern
Ontario,
Canada,
six
proven
effective
classifiers.
effectiveness
our
method
demonstrated
through
combination
field-survey-collected
differently
proportioned
Our
results
indicated
directly
remains
challenging,
unless
parcels
large,
regular
shape,
distinct
color
differences
from
surroundings.
Additionally,
approach
significantly
improved
classifiers
alleviated
problem.
Compared
trained
only
samples,
resulted
average
improvement
16%
78.5%
respectively.
random
forest
achieved
relatively
good
performance,
with
weighted-average
F1
0.97
0.996
obtained
two
areas,
contributes
insights
into
solutions
highlights
promising
application
models
like
International Journal of Applied Earth Observation and Geoinformation,
Journal Year:
2024,
Volume and Issue:
130, P. 103929 - 103929
Published: May 25, 2024
In
the
quest
for
large-scale
photovoltaic
(PV)
panel
extraction,
substantial
data
volumes
are
essential,
given
demand
sub-meter
rooftop
PV
resolution.
This
requires
concept
of
Latent
Photovoltaic
Locations
(LPL)
to
reduce
scope
amount
subsequent
processing.
order
minimize
manual
annotation,
a
pioneering
weakly-supervised
framework
is
proposed,
which
capable
generating
pixel-level
annotations
segmentation
based
on
image-level
and
provides
two
datasets
required
classification-then-segmentation
strategy
without
more
annotations.
The
strong
noise-resistance
Segment
Anything
Model
(SAM)
discovered
in
extremely
difficult
rough
coarse
pseudo-label
refinement,
which,
after
integrating
probability
updating
mechanism,
achieves
seamless
transition
from
scene
classification
semantic
segmentation.
resulting
national
LPL
distribution
map,
rendered
at
2
m
resolution,
showcases
commendable
92
%
accuracy
F1-score
91
%,
advantages
terms
efficiency
have
been
verified
through
large
number
experiments.
process
explores
how
use
fundamental
models
accelerate
remote
sensing
information
extraction
process,
crucial
current
trajectory
deep
learning
sensing.
relevant
code
available
https://github.com/Github-YRQ/LPL.
Sensors,
Journal Year:
2024,
Volume and Issue:
24(9), P. 2889 - 2889
Published: April 30, 2024
In
the
realm
of
computer
vision,
integration
advanced
techniques
into
pre-processing
RGB-D
camera
inputs
poses
a
significant
challenge,
given
inherent
complexities
arising
from
diverse
environmental
conditions
and
varying
object
appearances.
Therefore,
this
paper
introduces
FusionVision,
an
exhaustive
pipeline
adapted
for
robust
3D
segmentation
objects
in
imagery.
Traditional
vision
systems
face
limitations
simultaneously
capturing
precise
boundaries
achieving
high-precision
detection
on
depth
maps,
as
they
are
mainly
proposed
RGB
cameras.
To
address
FusionVision
adopts
integrated
approach
by
merging
state-of-the-art
techniques,
with
instance
methods.
The
these
components
enables
holistic
(unified
analysis
information
obtained
both
color
D
channels)
interpretation
data,
facilitating
extraction
comprehensive
accurate
order
to
improve
post-processes
such
6D
pose
estimation,
Simultanious
Localization
Mapping
(SLAM)
operations,
dataset
extraction,
etc.
employs
YOLO
identifying
within
image
domain.
Subsequently,
FastSAM,
innovative
semantic
model,
is
applied
delineate
boundaries,
yielding
refined
masks.
synergy
between
their
scene
understanding
ensures
cohesive
fusion
segmentation,
enhancing
overall
precision
segmentation.
IEEE Transactions on Geoscience and Remote Sensing,
Journal Year:
2024,
Volume and Issue:
62, P. 1 - 15
Published: Jan. 1, 2024
Segment
anything
model
(SAM)
has
been
widely
applied
to
various
downstream
tasks
for
its
excellent
performance
and
generalization
capability.
However,
SAM
exhibits
three
limitations
related
remote
sensing
semantic
segmentation
task:
1)
the
image
encoders
excessively
lose
high-frequency
information,
such
as
object
boundaries
textures,
resulting
in
rough
masks;
2)
due
being
trained
on
natural
images,
faces
difficulty
accurately
recognizing
objects
with
large-scale
variations
uneven
distribution
images;
3)
output
tokens
used
mask
prediction
are
images
not
applicable
segmentation.
In
this
paper,
we
explore
an
efficient
paradigm
applying
of
images.
Furthermore,
propose
MeSAM,
a
new
fine-tuning
method
more
suitable
adapt
it
tasks.
Our
first
introduces
inception
mixer
into
encoder
effectively
preserve
features.
Secondly,
by
designing
decoder
remote-sensing
correction
incorporating
multiscale
connections,
make
up
difference
from
Experimental
results
demonstrated
that
our
significantly
improves
accuracy
outperforming
some
state-of-the-art
methods.
The
code
will
be
available
at
https://github.com/Magic-lem/MeSAM.
Crop-specific
land
cover
mapping
is
a
vital
application
in
agro-geoinformatics
with
the
proliferation
of
remote
sensing
data
and
machine
learning
techniques.
This
paper
presents
novel
approach
to
enhance
well-known
Cropland
Data
Layer
(CDL)
product
by
U.S.
Department
Agriculture
(USDA)
National
Agricultural
Statistics
Service
(NASS)
using
Meta's
Segment
Anything
Model
(SAM).
The
study
leverages
SAM's
zero-shot
generalization
capability
automatically
delineate
cropland
fields
from
Sentinel-2
images.
By
voting
for
major
crop
types
within
each
delineated
unit,
substantial
number
noisy
pixels
CDL
can
be
eliminated,
leading
notable
improvements
accuracy.
Preliminary
experimental
results
across
key
agricultural
regions
U.S.,
such
as
California's
Central
Valley
Corn
Belt,
suggest
that
SAM
significantly
quality
original
data.
ability
refine
crop-specific
data,
like
CDL,
demonstrates
practical
applicability
monitoring
systems.
Moreover,
result
showcases
promising
potential
integrating
into
existing
type
classification
workflows
create
high-quality
early-
in-season
maps
on
national
scale
minimal
effort.