WHICH FEATURES OF INVASIVE ALIEN PLANTS ARE CRUCIAL FOR THEIR MAPPING WITH AIRBORNE HYPERSPECTRAL AND ALS REMOTE SENSING DATA?

Full text

WHICH FEATURES OF INVASIVE ALIEN PLANTS ARE CRUCIAL
FOR THEIR MAPPING WITH AIRBORNE HYPERSPECTRAL
AND ALS REMOTE SENSING DATA?
Barbara Tokarska-Guzik, Dominik Kopeć,
Katarzyna Bzdęga, Anna Halladin-Dąbrowska,
Andrzej Pasierbiński, Beata Woziwoda,
Sylwia Szporak-Wasilewska, Jacek Jóźwiak,
Gabriela Kuc, Anna Jarocińska, Adriana
Marcinkowska-Ochtyra, Anita Sabat, Łukasz Sławik

2
Outline of presentation
•Introduction
•Invasive alien plant
identification and mapping
using remote sensing techniques
– role of plant traits
– role of spectral characteristics
– methods
– examples of preliminary results
– pros and cons
•HABITars project as an example
of testing new methodology
in habitats and species detection

3
Introduction
The fast and effective detection and mapping of
invasive alien plants at different spatial scales
is becoming increasingly important for their
management Dorigo et al. 2012, Bradley 2014
The application of hyperspectral and ALS remote
sensing data is a method complementary to
traditional field surveys, which additionally
allows coverage of large areas
Huang & Asner 2009, He et al. 2011
The use of invasive plant functional traits can
improve RS mapping, using ecological insights
on processes and functions associated with
invasion Niphadkar & Nagendra 2016, Müllerová et al. 2017
Remote sensing offers a means of fast and efficient
monitoring of invasive plants, but still optimal
methodologies remain to be defined
Müllerová et al. 2017

4
Invasive alien plant identification using remote sensing techniques
The study aims at answering the following questions:
which features are most useful
in the identification of individual species,
including at the early invasion stages?
which method of classification is most effective?
is it possible to select a set of features
and a methods which would be recommended
as optimal?

5
Invasive alien plant identification using remote sensing techniques
- role of plant traits
Plant
traits
morphological
/ structural / phenological
Life
history / growth form
annual
forb/grass, perennial forb/grass, shrub, tree, climber
Height
(cm)
low
/ medium / high
Foliage
structure
large
-small / distinct-indistinct / characteristics-not characteristics
Phenological
state / phase
flowering
/ fruiting / senescence period
Colour
changing
during growing season / unchanging
T
ype of spatial distribution
single
individuals / in patches / in dense stands
Cover
(%)
low
/ medium / high
S
eparateness / distinguishability
difference in the ratio to the background
i.e. surrounding vegetation, coexisting species
Photosynthetic
pigment contents
chlorophyll
/ carotenoids concentration
Condition
of plants
health
of plants / defoliations / leaf discolouration and damage

6
Hyperspectral data
MNF bands
Vegetation indices (e.g. mNDVI705, mCARI, NDII, CAI)
ALS data
BCAL products (BCAL Vegetation, BCAL Intensity)
SAGA GIS products
1MNF 6MNF 15MNF
Invasive alien plant identification using remote sensing techniques
- role of spectral characteristics
Spatial and spectral resolution of used sensors
HySpex data - 1 metre
ALS Intensity image with 0.5 m spatial resolution
RGB camera - 10 cm spatial resolution
Hyperspectral scanner
HySpex 400-2500 nm
VNIR – 0.5 M
SWIR – 1 m

7
•ALS derivatives:
•BCAL:
•Vegetation products (35):
•Minimum Height
•Mean Height
•Mean Absolute Deviation (AAD) from Mean
Height
•Height Variance
•Height Kurtosis
•Height Coefficient of Variation
•Number of LiDAR Returns
•…
•Intensity products (13):
•Minimum Intensity
•Maximum Vegetation Intensity
•Mean Vegetation Intensity
•St. Dev. Vegetation Intensity
•Minimum Bare-earth Intensity
•…
•SAGA GIS
•Topographic products (11):
•Topographic Position Index
•SAGA Wetness Index
•Direct Insolation
•Slope
•Aspect
•…
Invasive alien plant identification using remote sensing techniques
- role of spectral characteristics

8
Invasive alien plant identification using remote sensing techniques
- methods
In the present study we test:
selected features of a few species differing
in terms of life history and other features
oannuals: Echinocystis lobata and Erigeron annuus
operennial herbs: Heracleum sosnowskyi and Solidago spp.
owoody plants: Spiraea tomentosa and Padus serotina
against collected airborne and on-ground
botanical reference data
Classification methods
oMixture Tuned Matched Filtering (MTMF)
oSpectral Angle Mapper (SAM)
oRandom Forest (RF)
oSupport Vector Machines (SVM)
Echinocystis lobata Erigeron annuus
Heracleum spp.
Padus serotina Spirea tomentosa
Solidago spp.

9
Invasive alien plant identification using remote sensing techniques
- methods
Three times during the growing season were acquired :
airborne data
ohyperspectral images
oAirborne Laser Scanning (ALS)
ohigh-resolution aerial photographs (RGB)
on-ground botanical reference data
including species characteristics,
among others:
o percentage cover
o phenological stage
o height
o coexisting species
additionally species condition
(discolorations, damage)
and type of land use
•100 referential training plots
per species in one campaign
•research areas
of average size 5 km2

10
Invasive alien plant identification using remote sensing techniques
- methods
Significant criteria:
Timing of data acquisition
3 campaigns during growing season
including spectral properties,
acquired by ASD FieldSpec using 4 measurements
Synchronization of data acquisition
almost simultaneous
Scope of obtained data
large trial of reference points
for classification and the validation
of results
reference points (reference polygons)
- established in the field permanent plots

11
Taxon name Campaign 1 Campaign 2 Campaign 3 Total
Echinocystis lobata 109 110 96 315
Erigeron annuus 112 105 61 278
Heracleum spp. 104 104 108 316
Solidago spp. 103 107 107 317
Spiraea tomentosa - 106 106 212
Padus serotina 105 105 105 315
Assumption:
100 reference polygons for each plant species
Total number of reference polygons
for the analyzed plant species in each campaign
Botanical surveys
Invasive alien plant identification using remote sensing techniques
- methods
Reference polygons collected for surrounding
vegetation and other species
with similar morphology or frequently coexisting species

12
Examples of preliminary results
Echinocytis lobata
Data Processing Techniques
- Random Forest
- ratio of training polygons to validation - 50:50
- two class of polygons: Echinocystis lobata and background (surrounding vegetation and other species)
3 2 1
peak of flowering
2nd campaign
vegetative growth of juvenile individuals
1st campaign
yellowing of plant leaves
3rd campaign

13
Examples of preliminary results
Echinocytis lobata
3 2 1
MNF (21 bands manually choosen)
Overall Accuracy 99%
Kappa index 70.69
The most useful products of airborne data and Accuracy Assessment
Evaluation of the best results
results underestimated
– small coverage of E. lobata
– strong influence of co-occurring species
MNF (23 bands manually choosen)
and BCAL intensity (5 choosen)
overall accuracy 97%
Kappa index 83.05
peak of flowering
recommended time period
for detection
MNF (26 bands manually choosen)
and BCAL intensity (5 choosen)
overall accuracy 97%
Kappa index 85.55
results slightly worse than
in 2nd campaign
– strong influence of co-occurring
species and yellowing of leaves

14
Examples of preliminary results
Solidago spp.
Classification method: Random Forest
The most useful products of airborne data – manually chosen:
25 MNF bands + 4 BCAL intensity + CHM
3
2
1
1st campaign – vegetative growth
Results overestimated OA: 91% Kappa: 0.75
Accuracy Assessment
Evaluation of the best results
2nd campaign
Results overestimated OA: 92% Kappa: 0.78
3rd campaign – peak of flowering
The best results OA: 93% Kappa: 0.77
recommended time period
for detection

15
Examples of preliminary results
3
2
1
Heracleum sosnowskyi
Classification method: MTMF
The most useful products of airborne data – manually chosen:
25 MNF bands
1st campaign – vegetative growth
Results significantly overestimated
OA: 97% Kappa: 0.96
2nd campaign – flowering
Results slightly overestimated
OA: 98% Kappa: 0.98
3rd campaign – mowing
Results slightly overestimated
OA: 97% Kappa: 0.96
Evaluation of the best results
overestimated The best result
recommended time period
for detection
overestimated - mowing
1

16
Examples of preliminary results
Erigeron annuus
Classification method: Random Forest
The most useful products of airborne data MTMF – manually
chosen: 25 MNF bands + CMN + BCAL vegetation + BCAL
intensity + Vegetation Indices
2nd campaign
Class: Erigeron annuus
- overestimated
Class: Erigeron annuus < 70%
- overestimated
Class: Erigeron annuus >70%
Remarks
Conflict identified: other perennials
Classifies poorly at low coverage
Occurs in agricultural areas regularly mowed
Species classifies better in less fertile soils
Recommendations
Determining the minimum for percentage cover
Taking the land use into account
Increasing the number of training polygons for
non-target species

17
Examples of preliminary results
Spiraea tomentosa
Classification method: Random Forest
•six classes of polygons: Spiraea tomentosa,
background (other species), water, forest, roads, shadows
The most useful products of airborne data – manually chosen:
20 MNF bands + Vegetation indices (e.g. CAI, NDII, NDVI 705) +
CHM - Canopy Height Model + ALS products (e.g. SAGA, OPALS)
2nd campaign - September
Results slightly overestimated
OA: 97% Kappa: 0.81
3rd campaign - October
Results slightly overestimated
OA: 97% Kappa: 0.84
possible to detect the species in the non-blooming state

18
Examples of preliminary results
Padus serotina
Classification method:
RF (Random Forest) and SVM (Support Vector Machine)
The most useful products of airborne data – manually chosen:
25MNF, CHM, BCAL Intensity, BCAL Vegetation
Accuracy Assessment
Campaign
Method
O.A.(%)
Kappa F1(%)
Description
1 RF 98 98 82
overestimated
, strong influence of
fallows
2 SVM 98 98 79
overestimated
on fallows
3 SVM 99 99 92
strongly
overestimated, strong
influence of
fallows
Padus serotina

19
Examples of preliminary results
Padus serotina
Remarks
Conflict identified: other deciduous tree
species, difficult to detect in low vegetation
(perennials Tanacetum vulgare, Solidago spp.)
High diversity of tree flora in the research area
Classifies poorly on fallows
Difficult to detect the initial stage of invasion
Recommendations
It is necessary to account for other deciduous
tree species reference points
Include tree crown spread
Include tree crown overgrowth with other tree
species
It might be beneficial to acquire aerial survey
data in the Autumn (leaf discolourisation)
1st campaign
2nd campaign

20
Invasive alien plant identification using remote sensing techniques
- pros and cons
Several different datasets and methods suitable for detection and identification of each plant species
mentioned can already be recommended
Species with percentage cover lower than 30% are poorly detected.
Best results are usually obtained for those with over 60% coverage
(it is necessary to attempt to set a minimum threshold)
For species that do not form dense cover and coexist with others, it is possible to identify this mixed type of
vegetation
Detecting species in the initial stage of invasion is problematic
The timing of agricultural treatments are crucial in selecting the dates for field and flight campaigns
Random forest appeared the most effective classification method
Among the plant features that contributed to the best classification results were percentage cover, growth
stage and flowering stage
The optimal dataset includes combination of MNF, CHM, ALS Intensity products and vegetation
indices – depending on the plant species

21
Final remarks
Airborne
data
On-ground
data
Data analysis and classification
Methodology of species identification
Detection
of species
verification verification

The research has been carried out under
the Biostrateg Programme of
the Polish National Centre for
Research and Development (NCBiR),
project DZP/BIOSTRATEG-II/390/2015:
The innovative approach supporting
monitoring of non-forest Natura 2000
habitats, using remote sensing
methods – HABITars
2016-2018
HABITars project as an example of testing new methodology
in habitats and species detection

23
Team
Zastosowanie technik teledetekcyjnych w ocenie stanu środowiska…
Lider:
MGGP Aero Sp. z o.o. Łukasz Sławik – project coordinator
Jaromir Borzuchowski, Jan Niedzielko, Tomasz Kundzierwewicz, Adam Podsada,
Agnieszka Ptak, Agnieszka Mleczko, Anna Matusik, Konrad Duda, Adam Kania
Partners:
University of
Lodz
Dominik Kopeć – Lider UŁ
Dorota Michalska Hejduk, Anna Halladin-Dąbrowska,
Beata Woziwoda, Justyna Wylazłowska, Agnieszka Piernik
(UMK), Dariusz Kaminski (UMK), Agata Zakrzewska,
Karolina Kuświk
University of Technology and Life
Sciences
Hubert Piórkowski – Lider ITP.
Zuzanna Oświecimska-Piasko, Filip Jarzombkowski, Marek Rycharski, Paweł
Kalinowski, Beata Nasiłowska, Aleksandra Kazuń, Kamila Brzezińska,
Katarzyna Kotowska, Ewa Gutowska, Grzegorz Kaliszewski, Agnieszka
Gutkowska, Marta Wielgosz, Łukasz Krajewski, Katarzyna Topolska
University of
Warsaw
Anna Jarocińska – Lider UW
Bogdan Zagajewski, Adriana Marcinkowska-Ochtyra,
Adrian Ochtyra, Edwin Raczko, Marlena Kycko, Anita
Sabat
University of Silesia in Katowice
Barbara Tokarska-Guzik – Lider UŚ
Agnieszka Błońska, Beata Babczyńska-Sendek, Teresa Nowak, Katarzyna
Bzdęga, Andrzej Pasierbiński, Agnieszka Kompała-Bąba, Gabriela Woźniak,
Edyta Sierka, Barbara Fojcik, Alina Urbisz, Adrian Zarychta, Małgorzata Frelich
,
Katarzyna Kulik-Knapik, Ewelina Roszkowska, Małgorzata Szary, Łukasz Folcik
Warsaw
University of Life Sciences
Stefan Ignar – Lider SGGW
Sylwia Szporak-Wasilewska, Jarosław Chormański,
Małgorzata Kleniewska, Wojciech Ciężkowski, Piotr
Dąbrowski, Tomasz Gnatowski, Jan Szatyłowicz, Jacek
Jóźwiak, Maciej Góraj, Gabriela Kuc, Luca Demarchi
Warsaw
University of Technology
Katarzyna Osińska-Skotak – Lider PW
Bożena Michna, Joanna Pluto-Kossakowska, Aleksandra Radecka,
Przemysław Kupidura, Zdzisław Kurczyski, Krzysztof Bakuła, Łukasz
Jełowicki, Wojciech Ostrowski, Paulina Bartkowiak, Konrad Górski

24
Which features of invasive alien plants are crucial
for their mapping with airborne hyperspectral
and ALS remote sensing data?