ArticlePDF Available

OBJECT-BASED CHANGE DETECTION ON ACACIA XANTHOPHLOEA SPECIES DEGRADATION ALONG LAKE NAKURU RIPARIAN RESERVE

Article

OBJECT-BASED CHANGE DETECTION ON ACACIA XANTHOPHLOEA SPECIES DEGRADATION ALONG LAKE NAKURU RIPARIAN RESERVE

Abstract

Automated mapping of heterogeneous riparian landscape is of high interest to assess our planet. Still, it remains a challenging task due to the occurrence of flooded vegetation. While both optical and radar images can be exploited, the latter has the advantage of being independent acquisition conditions. However, and despite their popularity, the threshold-based approaches commonly used present some drawbacks such as not taking into account the spatial context and providing mixed pixels within class boundaries. In this study, we propose a novel methodology to avoid such issues by using an object-based image analysis approach on polarimetric radar data. We use our workflow to map the degrading Acacia x. species along lake Nakuru Riparian reserve, and obtain highly-accurate results.
OBJECT-BASED CHANGE DETECTION ON ACACIA XANTHOPHLOEA SPECIES
DEGRADATION ALONG LAKE NAKURU RIPARIAN RESERVE
A. Osioa,
, S. Lef`
evreb
aTechnical University of Kenya, Nairobi, Kenya
bUniv. Bretagne Sud, UMR 6074, IRISA, F-56000 Vannes, France
KEY WORDS: OBIA, SAR, Change Detection, Flood mapping
ABSTRACT:
Automated mapping of heterogeneous riparian landscape is of high interest to assess our planet. Still, it remains a challenging task due
to the occurrence of flooded vegetation. While both optical and radar images can be exploited, the latter has the advantage of being
independent acquisition conditions. However, and despite their popularity, the threshold-based approaches commonly used present
some drawbacks such as not taking into account the spatial context and providing mixed pixels within class boundaries. In this study,
we propose a novel methodology to avoid such issues by using an object-based image analysis approach on polarimetric radar data. We
use our workflow to map the degrading Acacia X. Species along lake Nakuru Riparian reserve, and obtain highly-accurate results.
1. INTRODUCTION
Wetlands are important in regulating both the aquatic and benthic
ecosystems. In recent time, Lake Nakuru situated along Kenya’s
great Rift valley has been facing challenges due to the rising water
levels (Onywere et al., 2013), hence causing destruction of the
National Park’s natural and cultural features. Amongst the natural
features that have been gradually affected by the floods, and that
need to be monitored, is the Acacia xanthophloea tree species
which is an important feed for the giraffes and vervet monkeys in
the park (Pellew, 1983). However, detecting and mapping flooded
areas remains particularly challenging when the target vegetation
or trees are immersed in water.
Previous studies have engaged optical and SAR imageries and
their derivatives to extract spatial information from riparian re-
serves. Wetland classification based on optical imagery remains
limited due to heavy cloud cover (Amani and Mobasheri, 2015).
Conversely, SAR sensors can provide observations irrespective of
time of the day or night, penetrating through cloud cover, hence
producing images of high quality and integrity (Anusha and Bharathi,
2020). However, their use in submerged vegetation classification
mostly relies on pixel-based thresholding(Schumann et al., 2010)
Such a strategy does not take care of the contextual information
of the pixels at the land cover class boundaries, thus leading to
low classification accuracies.
Thus, the goal of our study is to use time series of Sentinel-1 (VV-
VH) Synthetic Aperture Radar (SAR) with Single Look Complex
(SLC) acquisition modes from 2014 to 2020 within an Object-
Based Image Analysis (OBIA) framework (Franklin and Ahmed,
2017) in order to detect changes in Acacia Xanthophloea degra-
dation along Lake Nakuru Riparian Reserve. Although previous
studies have already used S1 Ground Range detected products
in flood mapping (Osio et al., 2020), few studies have involved
S1 SLC polarimetric SAR products together with an object-level
analysis for flood mapping.
Corresponding author.
2. MATERIAL & METHODS
2.1 Study Area
Lake Nakuru National park is located in Nakuru County, 170
kms from Nairobi. Its bottom initially before the recent flood-
ing was about 1756 metres above sea level while the surface of
the water at 1758.5m above sea level. The altitude ranges from
1760-2080m above sea level. Mean annual rainfall ranges be-
tween 876mm and 1050mm and has an inherent bimodal pattern.
The long rains start in March and end in June while the short
rains occur between October and December. Mean daily mini-
mum and maximum temperatures fluctuate between 8.2C and
25.6C. Lake Nakuru has no outlets and hence evaporation is the
only factor that accounts for water loss. Four seasonal rivers feed
the lake, ie Lamurdiak, Enjoro, Enderit and Enjoro. The soils are
volcanic and shallow in nature. Underneath the open grasslands
were soils and ashes which were well drained, friable to sandy
clay loams that supported the grasslands in the park. Siltation
from sand harvesting contributes to the fluctuation of the water
depth in the lake. However, recent flooding of the lake are hy-
pothesized to have been caused by additional factors i.e climate
change drivers and movement of tectonic plates.
2.2 Data
We consider Dual polarized Single Look Complex SAR, all from
ascending orbits with phase information between two cross po-
larized channels, i.e. VV and VH respectively. SAR data were
acquired at both wet and dry seasons, from Fall 2015 to Sum-
mer 2020 (12 dates). All the data were in the same datum and
projection i.e. WGS 84 / UTM Zone 36N.
2.3 Workflow
The workflow designed in this study aims to apply an OBIA
methodology for change detection on SAR data. It has the follow-
ing features: i) generating a baseline image using Landsat5TM
captured in 2010, ii) extracting features from SAR data, iii) con-
structing an OBIA ruleset for SAR data, iv) performing a multi-
temporal map analysis. We now explain the full process by pro-
viding details for each successive step.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2021
XXIV ISPRS Congress (2021 edition)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-347-2021 | © Author(s) 2021. CC BY 4.0 License.
347
1. Sentinel 1 SLC Preprocessing
Sentinel 1 Single Look Complex (SLC) of time series pre-
processing is conducted as follows: i) S-1 top split into
bursts; ii) orbit file alignment; iii) radiometric calibration;
iv) multi-date S-1 back-geocoding for image stacking; v) S-
1 top debursting; vi) AOI subsetting; vi) C2 covariance ma-
trix extraction; vii) multilooking: viii) refined Lee speckle
filtering; ix) range Doppler terrain correction; x) S-1 SLC
stack imagery generation.
2. ALOS PALSAR Preprocessing
We consider a single date of ALOS PALSAR of L-Band
and HH+HV channels with a 6.5m ground resolution cap-
tured in ascending mode from an incident angle of 38.7 de-
grees. The image was captured in September 2007 and pre-
processed using the following steps: i) radiometric calibra-
tion; ii) multi-looking using 1 Range look and 7 Azimuth
looks; iii) speckle filtering using refined Lee filter; iv) ALOS
deskewing for geometric corrections; v) terrain correction;
vi) application of land/sea mask for extraction of AOI.
3. Landsat 5TM Preprocessing
Landsat 5TM collected in 2010 was used as a baseline map
since it was still depicting the state of all riparian classes
along the Lake. The steps engaged here were: i) extrac-
tion of AOI of the study area; ii) atmospheric corrections
done by conversion of DN values to TOA reflectance; iii)
pansharpening to 15m spatial resolution using Ehlers fusion
technique; iv) production of a false colour image RGB us-
ing Band 4 (0.73-0.9um), Band 3 (0.63-0.69um) and Band
2 (0.52-0.60um). The latter combination make deciduous
trees such as acacia forest appearing bright red in colour,
urban area cyan while soil dark to light brown (see Fig. 5).
4. Sentinel 1 SLC Data Analysis
From the preprocessed S-1 SLC single date images, we gen-
erate bands C11, C22, C12REAL and C12IMAG using 22
(C2) matrix image in ESA SNAP toolbox (Lee and Pottier,
2017). The spectral bands derived from C2 matrix based
single date images consist in radar-based vegetation indices,
namely Radar Vegetation Index (RVI), Dual Pol Vegetation
Index (DpRVI), Polarimetric Radar Vegetation Index (PRVI)
and Degree of Polarization (DOP) for Dual Polarimetric SAR.
Time series temporal profiles were then generated from the
multi-date stack S1 SLC series built from datasets ranging
from 2014 to 2015 as shown in Fig. 1 and Fig. 2. Bands
with the highest backscatter response with other additional
variables were then used for further data analysis.
5. ALOS PALSAR Data analysis
The following bands were derived from ALOS PALSAR im-
age captured in 2007 before the floods with 6.5m ground
Resolution. Bands HH and HV were converted into σ(HH)
and σ(HV). These bands were transited to eCognition De-
veloper 10.1 where additional variables such as the PCA1,
PCA2 mean and standard deviation were added into the clas-
sification. Other variables include geometric features such
as Border Index and Ecliptic fit. Classifiers appended on the
classification include K-NN, Naive Bayes, Decision Tree
(DT) and Random Forest. Training and testing areas were
identified on this image and which were later used on multi-
date S-1 SLC image analysis.
6. Landsat 5TM Data analysis
Classification of the Landsat 5TM (with bands 4, 3, 2) data
was conducted with the aim of visualization as shown in
Fig. 5. The following classes were considered: Acacia For-
est, Bushed Themada Grassland, Chloris gayana Grassland,
Cynodon - Chloris - Themada Grassland, Cynodon niem-
fluensis Grassland, Cynodon niemfluensis - wooded Aca-
cia grasslands, Euphorbia Candelabra, Lake Nakuru, Olive
& Teclea Forest, Sand & Mudflats, Sedges & Marshes and
Tachonandus Bushland. These classes were then compared
to the classified image previously generated by (Ng’weno et
al., 2010) shown in Fig. 4.
Figure 1: Sentinel 1 SLC spectral response of RVI, DpRVI, PRVI
and DOP of the degraded Riparian Reserve 2014-2020 of Lake
Nakuru, Kenya.
Figure 2: Sentinel 1 SLC covariance matrix generated band av-
erages of C11, C12 and C22 response on the degraded Riparian
Reserve 2014-2020 of Lake Nakuru, Kenya.
3. RESULTS
The proposed methodology was designed to fill the following
knowledge gaps: i) to determine the spectral changes that have
taken place along the riparian reserve by using preprocessed and
co-registered multi-date stack time series images built from S-1
Single Look Complex SAR data; ii) to explore the capability of
the C2 covariance matrix derivatives (i.e. C11, C22 and RVI)
to characterize Acacia xanthophloea along Lake Nakuru riparian
reserve; iii) to explore the spatio-temporal changes that have oc-
curred on the Acacia strands from 2015 to 2020. We report here
the experiments conducted to answers these questions and pro-
vide some quantitative and qualitative results.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2021
XXIV ISPRS Congress (2021 edition)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-347-2021 | © Author(s) 2021. CC BY 4.0 License.
348
3.1 Determine polarimetric spectral response on time series
images detected along the riparian reserve via temporal
profiles
We generate 22 covariance matrix from SLC data to enhance
polarimetric characterization of the target. The covariance ma-
trix (C2) data (header and binary files) are used in the generation
of C11, C22, C12REAL and C12IMAG bands. We derive radar-
based vegetation indices, namely Radar Vegetation Indices (RVI)
(Mandal et al., 2020b), Dual Pol Vegetation Index (DpRVI), Po-
larimetric Radar Vegetation Index (PRVI) (Mandal et al., 2020a)
and Degree of Polarization (DOP) for Dual Polarimetric SAR.
The bands that had the highest backscatter response, i.e. C11,
C22 and RVI were then selected for multi-date image classifica-
tion across time series of Sentinel 1 Single Look Complex data
captured from the period 2014-2020. Fig.1 and Fig. 2 show the
multi-temporal average backscatter response of the C2 (22) co-
variance matrix generated bands.
3.2 The capability of C2 covariance matrix and spectral deriva-
tives in detecting the changes on the riparian vegetation
In this section the C11, C22 and RVI bands of each imagery cap-
tured after the floods between 2014 and 2020 are classified in
eCognition Developer 10.1 environment. For further class refine-
ment, the shapefiles consisting of the resulting classes, namely
Acacia Forest, Acacia undergrowth, Acacia crown gap and Other
were exported to ARCGIS 10.5. Later the tabular CSV files con-
sisting of multi-date Sentinel 1 SLC datasets were transited to
WEKA, an open source machine learning environment.
We report below the OBIA classification results conducted on
the Acacia Strands along the different parts of Lake Nakuru (i.e.
north-western and southern) at three different dates, namely De-
cember 2015, February 2017 and February 2019.
North-western, 2015. We consider both a multi-resolution seg-
mentation with parameters set to scale=5, shape=0.6 and com-
pactness=0.8 (Kavzoglu and Tonbul, 2017), and a spectral dif-
ference segmentation with scale=10 (Yang et al., 2017). Indeed,
using multiple segmentation algorithms has been proven to be
lead to better results (Tonbul and Kavzao˘
glu, 2019). Features
included in the classification were the mean and standard devia-
tion of C11, C22 and RVI. As far as geometric features are con-
cerned, we selected shape index, rectangular fit, roundness and
ecliptic fit (Kaplan and Avdan, 2019). The default K-NN clas-
sifier yielded an overall accuracy (OA) of 87% and a Cohen’s
Kappa score of 0.78. Using Random Forest with baggings and
100 iterations over 737 instances and 12 attributes with a 2/3-1/3
train-test split yielded higher values, with 96.1% of OA and 0.93
for Kappa. Cross-validation with a 10-fold approach improves
further the results, with an OA of 96.7% and a Kappa of 0.94.
Out of the 737 instances in the classification, 494 were detected
as belonging to the Acacia Forest class while 18 fall into the class
Other.
Southern, 2015. For this second experiment, we rely on multi-
resolution segmentation with scale=10, shape=0.6 and compact-
ness=0.8, and give more weight to bands C11 and RVI as shown
on Fig. 1 and Fig. 2. The mean and standard deviation of C11,
C22 and RVI were also included in the classification. Geometric
features such as shape index, ecliptic fit, roundness and rectan-
gular fit were completing the set of features. The default K-NN
algorithm in eCognition environment yielded an OA of 83% and
a kappa of 0.76. Comparatively, the Random Forest classifica-
tion with a 10-fold cross validation yielded an OA of 96.3% and
a Kappa of 0.94, this time lower than the 2/3-1/3 train-test split
yielding an OA of 96.8% and a Kappa of 0.95. Out of the 2491
instances used in the classification, 711 instances were classified
as Acacia Forest while 6 instances were confused as class Other.
True positive rate (TPR) and false positive rate (FPR) for Acacia
Forest were measured as 99.7% and 0.3% respectively.
North-western, 2017. Again, we used multi-resolution segmen-
tation with scale=10, shape=0.6 and compactness=0.8 (yielded
1014 objects), as well as spectral difference segmentation at scale
10, and consider the mean and standard deviation of C11, C22
and RVI, and integrate a set of geometric features (shape in-
dex, rectangular fit, roundness, border index and ecliptic fit). We
obtain an OA of 97% and a Kappa of 0.95 with the default K-
NN classifier, while Random Forest classification carried out in
WEKA environment using baggings with 100 iterations and 10-
fold cross-validation yielded an OA of 94.4% and a Kappa co-
efficient of 0.91. Out of the 1,140 instances used in this classi-
fication, 232 instances were classified as Acacia Forest while 22
instances confused as class Other. TPR for the Acacia Forest tar-
get class was 91% while FPR reaches 2.5%. Using the 2/3-1/3
train test split lowers the rsults, with an OA of 91.5% and Kappa
equals to 0.86. Out of the 388 test instances, 64 were classified
as Acacia Forest while 11 as class Other.
Southern, 2017. We continue to use C11, C22 and RVI bands,
but also apply a Principal Component Analysis (PCA) leading
to 3 principal components, i.e. PCA1, PCA2 and PCA3. We
then replace C22 by PCA1 and reorder the bands as C11, RVI,
and PCA1, before applying a multi-resolution segmentation with
scale=10, shape=0.6 and compactness=0.8, and a spectral dif-
ference segmentation at scale 10 to separate class Acacia For-
est from Sporobolus Spicatus and Chloris gayana grasslands. K-
NN classifier yielded an OA of 98.7% and a Kappa of 0.97. Let
us emphasize the positive role played by the PCA textural fea-
tures in the classification process, with an improvement of 15.4%
w.r.t. North-western, 2015 (where no such features were used).
Comparatively, Random Forest classification with 10-fold cross-
validation yielded an OA of 97.4% and a Kappa of 0.96. Out
of the 2668 instances used in this approach, 1250 were classified
as Acacia Forest and 3 belong to class Acacia undergrowth, 17 in
Other. TPR stood at 97.8% while FPR was 2.2%. Finally, the 2/3-
1/3 train-test split yielded an OA of 98.1% and a Kappa of 0.97.
1249 test instances were classified as Acacia Forest while 11 in-
stances were confused as either Acacia Undergrowth or Other.
TPR and FPR were measured to 98.8% and 1.2%, respectively.
North-western, 2019. Similarly to the previous case, we con-
sider C11, C22, RVI, and the Principal Components that can be
derived from them. Here, we give more weight to C11, RVI,
PCA1 and PCA2 according to their importance as determined by
their correlation coefficient (e.g. R= 0.91 between PCA2 and
RVI, R= 0.97 between C11 and PCA1, and R= 0.96 be-
tween C22 and PCA2). Some geometric features, namely eclip-
tic fit, shape index, rectangular fit, roundness, border index and
shape index were used. Best results were obtained with super-
pixel segmentation (SLICO) (Mi and Chen, 2020) and spectral
difference segmentation. The default K-NN classifier in eCogni-
tion Developer 10.5 yielded perfect results (OA of 100%, Kappa
of 1.0). Random Forest bagging with 100 iterations and 10-fold
cross-validation yielded an OA of 97.9% and a Kappa of 0.96.
However TPR of the target class (Acacia Forest) achieved stood
at 85.4% while FPR stood at 1.3%. Out of the 1634 instances, 70
were classified as Acacia Forest while 21 instances were misclas-
sified as either class Other or Acacia undergrowth. The 2/3-1/3
train-test split yielded an OA of 98.3% and a Kappa of 0.97, lead-
ing to TPR and FPR for the Acacia Forest class measured at 79%
and 0.5%, respectively. Out of the 624 test instances, 22 were
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2021
XXIV ISPRS Congress (2021 edition)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-347-2021 | © Author(s) 2021. CC BY 4.0 License.
349
classified as Acacia Forest while 3 instances were misclassified
as Acacia undergrowth.
Southern, 2019 The same approach and parameters used in
the previous setup were considered here: bands C11, C22, RVI,
PCA1, PCA2, PCA3; their mean and standard deviation; su-
perpixel (SLICO) and spectral difference segmentation. K-NN
achieves again perfect results. Random Forest with baggings, 100
iterations and 10-fold cross-validation realised an OA of 86.1%
and a Kappa of 0.78. Out of the 1014 instances, 141 were cor-
rectly classified as Acacia Forest, 21 misclassified as Acacia un-
dergrowth, 8 as Other.TPR for the target class (Acacia Forest)
was 90% and FPR was 4%. Random Forst with a 2/3-1/3 train-
test split resulted in OA of 86.7% and Kappa of 0.79. Out of the
345 instances involved in this classification, 54 instances were as-
signed the target class Acacia Forest while 2 were misclassified
as class Other and 8 as class Acacia undergrowth. TPR and FPR
for this target class were then 98% and 5%.
3.3 Spatio-Temporal Changes on Acacia Xanthophloea spp
Along Lake Nakuru Riparian Reserve
We have carried out statistical change detection on the different
training and test areas. This was made possible by calculating
the area of the Acacia X. Spp strands not affected by the floods
ranging from 2014 to 2020 as shown on the bar graphs in Figure
3. Site 1 corresponds to the Northern side of Lake Nakuru, di-
vided in a training site (north-western, in blue in the figure) and
a test site (north-eastern, in yellow in the figure). Similarly, site 2
corresponds to the Southern side of the lake (with the same color
code). Figure 3 reports the distribution of Acacia Forest strands
in the different training and test sites.
In site 1 (train), the spatial difference in Acacia X. Spp strands
between year 2014 and 2020 was accounting for an area of 6236
pixels, corresponding to an area initially covered by the Healthy
Acacia Forest that became degraded and submerged in the flooded
lake. For the test area, the difference in area covered by the
healthy Acacia X. Spp between year 2014 and 2020 was 22,552
pixels as of 26th July, 2020. Similarly, the Acacia forest tran-
sition for the training area of site 2 (southern part of the Lake
Nakuru) consists in a reduction of a total area initially covered
by Healthy Acacia X. Spp in 2014 of 43,100 pixels to an area
of 20,548 pixels by the year 2020, i.e. a loss in area of 22,522
pixels. As far as the test site is concerned, we observed a similar
loss, from 9,092 pixels in February, 2014 to 963 pixels by July,
2020. In other words, the area covered by degraded Acacia X.
Spp accounted for 8,129 pixels.
To determine if the Acacia Forest degradation was actually af-
fected by the floods, the area covered by the lake per each given
year (2014-2020) was calculated and correlated to the per pixel
area covered by the remnants of the Acacia along Lake Nakuru
Riparian Reserve. To do so, we used multi-resolution and spec-
tral difference segmentation approaches on a composite image
made of C11, C22, and C11/C22 bands. We obtained the follow-
ing measures: 52.2 km2in 2014, 54.11 km2in 2016, 57.24 km2
in 2018, and 59.8 km2in 2020. These figures were correlated to
the area covered by the Acacia Forest on train site 1 as follows:
16,528 pixels in 2014, 13,320 pixels in 2016, 10,650 pixels in
2018, and 10,302 pixels in 2020. We then use the correlation co-
efficient to measure strength of relationships between variables.
We observed a strong negative correlation of R=0.94 be-
tween the area of the Lake against the area covered by Healthy
Acacia Forest per each given year. The equation to the regression
line in relation to this correlation being y= 57870 808.9x.
This means that, as the level of the lake continued to rise over the
years, the more the Acacia Forest continued to degrade, confirm-
ing reports by (Mutangah, 1994) that the main cause of stress and
eventual death of the Acacia Xanthophloea was due to the rising
water table in a closed drainage system (i.e. Lake Nakuru Basin).
4. DISCUSSION
The study demonstrates how S1 captured in SLC mode was used
to characterize the degrading Acacia X. Spp along the shores
of the flooded Lake Nakuru Riparian reserve. The bands that
had the highest reflectance were C11, C22 and RVI. They were
used to carry out classifications in an object- based image analy-
sis (OBIA) pipeline using multi-date datasets ranging from 2014
to 2020. We hypothesized that SAR polarimetric bands filtered
through the C2 (2 2) Covariance Matrix could improve charac-
terization of the flooded riparian vegetation. In this study, clas-
sification was carried out on a time series of SLC products, af-
ter which they were imported into eCognition Developer soft-
ware environment for object detection and classification. Rule-
sets where different segmentation types and scales (Nico et al.,
2000, Yang et al., 2017, Kavzoglu and Tonbul, 2017) were ap-
pended on 28 training and testing sites.
The variables that contributed well to the classifications were de-
termined by measuring correlation coefficients. It was worth not-
ing that there was a strong positive correlation between the eclip-
tic fit and rectangular fit (R= 0.91), and between border index
and shape index (R= 0.95). We follow previous studies that
show some improvement of classification accuracy using shape
and geometric variables (Kaplan and Avdan, 2019). Other vari-
ables that contributed to the high-classification accuracies were
the standard deviation of polarimetric bands C11, C22 and RVI,
and their principal components. Strong positive correlations were
noted between the stdev of C11/PCA1 (R= 0.97), C22/PCA2
(R= 0.96) and PCA2/RVI (R= 0.98). The inclusion of the tex-
ture features (here PCA) onto the classification process increased
the K-NN based classification from 15.4%. These results corrob-
orate with previous studies where supervised classification and
textural variables have been used to improve classification accu-
racy (Granata et al., 2020, Osio et al., 2020, Pham et al., 2017).
The difference in the performance of classification algorithm have
been demonstrated using different methods. (Gaˇ
sparovi´
c and Do-
brini´
c, 2020) used McNemar’s χ2test to compare the perfor-
mance of classifier algorithms, including XGB, SVM and Ran-
dom Forest (RF). In this study we also compare the performance
of the three classifiers, namely Naive Bayes, Decision Tree (DT)
and Random Forest (RF) in a total of 84 classifications carried
out across the multi-date images. Let us note that the qualita-
tive and quantitative analysis provided in the previous section
was focusing solely on KNN and RF for the sake of concision.
Results achieved through inter-group comparisons using one way
ANOVA depicted statistically significant difference between the
three classifiers in all the training sites that were used in this study
using both cross-validation and 2/3-1/3 train-test split. Random
Forest was confirmed as the best performing classifier.
5. CONCLUSION
This study demonstrates how Sentinel-1 SLC polarimetric bands
σ(VV) and σ(VH) filtered through covariance matrix could im-
prove wetland classification. We consider an OBIA-based ap-
proach coupled with PCA-derived texture features, as well as ge-
ometric and shape features.
We have conducted numerous experiments on several datasets
and several dates, assessing the effect of all components in the
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2021
XXIV ISPRS Congress (2021 edition)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-347-2021 | © Author(s) 2021. CC BY 4.0 License.
350
Figure 3: Spatio-Temporal Change Detection on Acacia X. Spp on Training and Test sites 1 (left) and 2 (right).
pipeline: input bands for the segmentation, segmentation algo-
rithm and parameters, features to describe the objects, and clas-
sifier settings. We have observed that the optimal selection was
specific to each image (area, date) under study. The average re-
sults achieved by ALOS PALSAR with Random Forest (OA of
93.8%, Kappa of 0.92) makes it a suitable choice as a reference
image in this study. Out of 84 classifications carried out, Ran-
dom Forest (RF) (Zamani Joharestani et al., 2019) achieved the
highest average score (OA of 90.0%, Kappa of 0.88). However in
some instances, Acacia Forest was either confused as class Aca-
cia Undergrowth or Other, hence confirming reports by (Moskal
et al., 2011) on confusion between classes in wetland-based land
cover classification.
Between 2014 and 2020 the area covered by the degraded Acacia
X. Spp trees on the north western side of the lake was 6,236 pix-
els, on the southern part where the largest strands of acacia forest
were existing 22,522 pixels, and on the south western side 8,129
pixels were degraded. There was a strong negative correlation
(R=0.94) between the area covered by the lake against the
area covered by the Acacia Forest remnants on the riparian re-
serve, confirming reports by (Mutangah, 1994) that the flooding
lake could have been the cause of the death of the Acacia Trees
within the closed ecosystem. Future studies should look into the
possibility of using Quad polarized Sentinel 1 SLC products to
characterize flooded riparian vegetation.
ACKNOWLEDGEMENTS
The authors acknowledge: Kenya National Research Fund (K-
NRF) and Campus France through Pamoja PHC programme; Eu-
ropean Space Agency for the provision of Sentinel-1 Products;
Trimble Inc. Germany for the provision of eCognition Developer
Software Licence.
REFERENCES
Amani, M. and Mobasheri, M. R., 2015. A parametric method for
estimation of leaf area index using landsat etm+ data. GIScience
& Remote Sensing 52(4), pp. 478–497.
Anusha, N. and Bharathi, B., 2020. Flood detection and flood
mapping using multi-temporal synthetic aperture radar and op-
tical data. The Egyptian Journal of Remote Sensing and Space
Science 23(2), pp. 207–219.
Figure 4: Distribution of vegetation classes in Lake Nakuru Na-
tional Park (illustration from (Ng’weno et al., 2010)).
Franklin, S. E. and Ahmed, O. S., 2017. Object-based wet-
land characterization using radarsat-2 quad-polarimetric sar data,
landsat-8 oli imagery, and airborne lidar-derived geomorphome-
tric variables. Photogrammetric Engineering & Remote Sensing
83(1), pp. 27–36.
Gaˇ
sparovi´
c, M. and Dobrini´
c, D., 2020. Comparative assessment
of machine learning methods for urban vegetation mapping us-
ing multitemporal sentinel-1 imagery. Remote Sensing 12(12),
pp. 1952.
Granata, F., Gargano, R. and de Marinis, G., 2020. Artificial in-
telligence based approaches to evaluate actual evapotranspiration
in wetlands. Science of The Total Environment 703, pp. 135653.
Kaplan, G. and Avdan, U., 2019. Evaluating the utilization of
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2021
XXIV ISPRS Congress (2021 edition)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-347-2021 | © Author(s) 2021. CC BY 4.0 License.
351
Figure 5: Landcover classification in the study area, before the
onset of flooding.
Figure 6: The state of Acacia forest patches clockwise from upper
left corner: Southern part, North-Eastern part, Southern-Western
part, & North-Western part of the study area where Acacia Forest
patches (in red) existed before the onset of flooding.
the red edge and radar bands from sentinel sensors for wetland
classification. Catena 178, pp. 109–119.
Kavzoglu, T. and Tonbul, H., 2017. A comparative study of seg-
mentation quality for multi-resolution segmentation and water-
shed transform. In: 2017 8th International Conference on Recent
Advances in Space Technologies (RAST), IEEE, pp. 113–117.
Lee, J.-S. and Pottier, E., 2017. Polarimetric radar imaging: from
basics to applications. CRC press.
Mandal, D., Kumar, V., Ratha, D., Dey, S., Bhattacharya, A.,
Lopez-Sanchez, J. M., McNairn, H. and Rao, Y. S., 2020a. Dual
polarimetric radar vegetation index for crop growth monitoring
using sentinel-1 sar data. Remote Sensing of Environment 247,
pp. 111954.
Mandal, D., Ratha, D., Bhattacharya, A., Kumar, V., McNairn,
H., Rao, Y. S. and Frery, A. C., 2020b. A radar vegetation
index for crop monitoring using compact polarimetric sar data.
IEEE Transactions on Geoscience and Remote Sensing 58(9),
pp. 6321–6335.
Mi, L. and Chen, Z., 2020. Superpixel-enhanced deep neural
forest for remote sensing image semantic segmentation. ISPRS
Journal of Photogrammetry and Remote Sensing 159, pp. 140–
152.
Moskal, L. M., Styers, D. M. and Halabisky, M., 2011. Monitor-
ing urban tree cover using object-based image analysis and public
domain remotely sensed data. Remote Sensing 3(10), pp. 2243–
2262.
Mutangah, J. G., 1994. The vegetation of lake nakuru national
park, kenya: a synopsis of the vegetation types with annotated
species list. Journal of East African Natural History 83(1),
pp. 71–96.
Ng’weno, C. C., Mwasi, S. M. and Kairu, J. K., 2010. Distribu-
tion, density and impact of invasive plants in lake nakuru national
park, kenya. African Journal of Ecology 48(4), pp. 905–913.
Nico, G., Pappalepore, M., Pasquariello, G., Refice, A. and
Samarelli, S., 2000. Comparison of sar amplitude vs. coherence
flood detection methods-a gis application. International Journal
of Remote Sensing 21(8), pp. 1619–1631.
Onywere, S., Shisanya, C., Obando, J., Ndubi, A., Masiga, D.,
Irura, Z., Mariita, N. and Maragia, H., 2013. Geospatial extent of
2011–2013 flooding from the eastern african rift valley lakes in
kenya and its implication on the ecosystems. In: Papers, Kenya
Soda Lakes workshop. Kenya Wildlife Service Training Institute,
Naivasha.
Osio, A., Pham, M. and Lef `
evre, S., 2020. Spatial processing
of sentinel imagery for monitoring of acacia forest degradation in
lake nakuru riparian reserve. ISPRS Annals of the Photogramme-
try, Remote Sensing and Spatial Information Sciences 3, pp. 525–
532.
Pellew, R., 1983. The giraffe and its food resource in the
serengeti. i. composition, biomass and production of available
browse. African Journal of Ecology 21(4), pp. 241–267.
Pham, M.-T., Lefevre, S. and Aptoula, E., 2017. Local feature-
based attribute profiles for optical remote sensing image classi-
fication. IEEE Transactions on Geoscience and Remote Sensing
56(2), pp. 1199–1212.
Schumann, G., Di Baldassarre, G., Alsdorf, D. and Bates, P.,
2010. Near real-time flood wave approximation on large rivers
from space: Application to the river po, italy. Water Resources
Research.
Tonbul, H. and Kavzao˘
glu, T., 2019. Application of taguchi op-
timization and anova statistics in optimal parameter setting of
multi-resolution segmentation. In: 2019 9th International Confer-
ence on Recent Advances in Space Technologies (RAST), IEEE,
pp. 387–391.
Yang, J., He, Y. and Caspersen, J., 2017. Region merging using
local spectral angle thresholds: A more accurate method for hy-
brid segmentation of remote sensing images. Remote sensing of
environment 190, pp. 137–148.
Zamani Joharestani, M., Cao, C., Ni, X., Bashir, B. and Talebies-
fandarani, S., 2019. Pm2. 5 prediction based on random for-
est, xgboost, and deep learning using multisource remote sensing
data. Atmosphere 10(7), pp. 373.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2021
XXIV ISPRS Congress (2021 edition)
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-347-2021 | © Author(s) 2021. CC BY 4.0 License.
352
... Models created using SAR data produce uncertainty and variance in modelling accuracy of above ground biomass (Dalponte et al., 2018;Naik et al., 2021). On the contrary, Osio and Lefèvre (2021) confirmed that the use of SAR-C channels captured in Single Look Complex mode in conjunction with machine learning models and object-oriented approach yielded the best results with an Overall Accuracy (OA) of 98.1% and a Kappa of 97.0%, hence improving the above ground biomass classification on the Acacia xanthophloea strands around Lake Nakuru, Kenya. Despite achieving results at local scale, such a model was not suitable to capture individual degraded Acacia xanthophloea target trees that are fallen around the lake. ...
... Underneath the Acacia savanna were the open grasslands thriving on soils and ashes that were well-drained, friable to sandy clay loams (see Fig. 2). Recent studies have shown that the health of Acacia xanthophloea trees have been degrading since the year 2010 due to the persistent flooding around the lake (Osio et al., 2020;Osio and Lefèvre, 2021). ...
Article
Full-text available
Deep-learning-based image classification and object detection has been applied successfully to tree monitoring. However, studies of tree crowns and fallen trees, especially on flood inundated areas, remain largely unexplored. Detection of degraded tree trunks on natural environments such as water, mudflats, and natural vegetated areas is challenging due to the mixed colour image backgrounds. In this paper, Unmanned Aerial Vehicles (UAVs), or drones, with embedded RGB cameras were used to capture the fallen Acacia Xanthophloea trees from six designated plots around Lake Nakuru, Kenya. Motivated by the need to detect fallen trees around the lake, two well-established deep neural networks, i.e. Faster Region-based Convolution Neural Network (Faster R-CNN) and Retina-Net were used for fallen tree detection. A total of 7,590 annotations of three classes on 256×256 image patches were used for this study. Experimental results show the relevance of deep learning in this context, with Retina-Net model achieving 38.9% precision and 57.9% recall.
... Models created using SAR data produce uncertainty and variance in modelling accuracy of above ground biomass (Dalponte et al., 2018;Naik et al., 2021). On the contrary, Osio and Lefèvre (2021) confirmed that the use of SAR-C channels captured in Single Look Complex mode in conjunction with machine learning models and object-oriented approach yielded the best results with an Overall Accuracy (OA) of 98.1% and a Kappa of 97.0%, hence improving the above ground biomass classification on the Acacia xanthophloea strands around Lake Nakuru, Kenya. Despite achieving results at local scale, such a model was not suitable to capture individual degraded Acacia xanthophloea target trees that are fallen around the lake. ...
... Underneath the Acacia savanna were the open grasslands thriving on soils and ashes that were well-drained, friable to sandy clay loams (see Fig. 2). Recent studies have shown that the health of Acacia xanthophloea trees have been degrading since the year 2010 due to the persistent flooding around the lake (Osio et al., 2020;Osio and Lefèvre, 2021). ...
Preprint
Full-text available
Deep-learning-based image classification and object detection has been applied successfully to tree monitoring. However, studies of tree crowns and fallen trees, especially on flood inundated areas, remain largely unexplored. Detection of degraded tree trunks on natural environments such as water, mudflats, and natural vegetated areas is challenging due to the mixed colour image backgrounds. In this paper, Unmanned Aerial Vehicles (UAVs), or drones, with embedded RGB cameras were used to capture the fallen Acacia Xanthophloea trees from six designated plots around Lake Nakuru, Kenya. Motivated by the need to detect fallen trees around the lake, two well-established deep neural networks, i.e. Faster Region-based Convolution Neural Network (Faster R-CNN) and Retina-Net were used for fallen tree detection. A total of 7,590 annotations of three classes on 256 x 256 image patches were used for this study. Experimental results show the relevance of deep learning in this context, with Retina-Net model achieving 38.9% precision and 57.9% recall.
... Models created using SAR data produce uncertainty and variance in modelling accuracy of above ground biomass (Dalponte et al., 2018;Naik et al., 2021). On the contrary, Osio and Lefèvre (2021) confirmed that the use of SAR-C channels captured in Single Look Complex mode in conjunction with machine learning models and object-oriented approach yielded the best results with an Overall Accuracy (OA) of 98.1% and a Kappa of 97.0%, hence improving the above ground biomass classification on the Acacia xanthophloea strands around Lake Nakuru, Kenya. Despite achieving results at local scale, such a model was not suitable to capture individual degraded Acacia xanthophloea target trees that are fallen around the lake. ...
... Underneath the Acacia savanna were the open grasslands thriving on soils and ashes that were well-drained, friable to sandy clay loams (see Fig. 2). Recent studies have shown that the health of Acacia xanthophloea trees have been degrading since the year 2010 due to the persistent flooding around the lake (Osio et al., 2020;Osio and Lefèvre, 2021). ...
Preprint
Full-text available
Deep-learning-based image classification and object detection has been applied successfully to tree monitoring. However, studies of tree crowns and fallen trees, especially on flood inundated areas, remain largely unexplored. Detection of degraded tree trunks on natural environments such as water, mudflats, and natural vegetated areas is challenging due to the mixed colour image backgrounds. In this paper, Unmanned Aerial Vehicles (UAVs), or drones, with embedded RGB cameras were used to capture the fallen Acacia Xanthophloea trees from six designated plots around Lake Nakuru, Kenya. Motivated by the need to detect fallen trees around the lake, two well-established deep neural networks, i.e. Faster Region-based Convolution Neural Network (Faster R-CNN) and Retina-Net were used for fallen tree detection. A total of 7,590 annotations of three classes on 256×256 image patches were used for this study. Experimental results show the relevance of deep learning in this context, with Retina-Net model achieving 38.9% precision and 57.9% recall.
... Models created using SAR data produce uncertainty and variance in modelling accuracy of above ground biomass (Dalponte et al., 2018;Naik et al., 2021). On the contrary, Osio and Lefèvre (2021) confirmed that the use of SAR-C channels captured in Single Look Complex mode in conjunction with machine learning models and object-oriented approach yielded the best results with an Overall Accuracy (OA) of 98.1% and a Kappa of 97.0%, hence improving the above ground biomass classification on the Acacia xanthophloea strands around Lake Nakuru, Kenya. Despite achieving results at local scale, such a model was not suitable to capture individual degraded Acacia xanthophloea target trees that are fallen around the lake. ...
... Underneath the Acacia savanna were the open grasslands thriving on soils and ashes that were well-drained, friable to sandy clay loams (see Fig. 2). Recent studies have shown that the health of Acacia xanthophloea trees have been degrading since the year 2010 due to the persistent flooding around the lake (Osio et al., 2020;Osio and Lefèvre, 2021). ...
Conference Paper
Deep-learning-based image classification and object detection has been applied successfully to tree monitoring. However, studies of tree crowns and fallen trees, especially on flood inundated areas, remain largely unexplored. Detection of degraded tree trunks on natural environments such as water, mudflats, and natural vegetated areas is challenging due to the mixed colour image backgrounds. In this paper, Unmanned Aerial Vehicles (UAVs), or drones, with embedded RGB cameras were used to capture the fallen Acacia Xanthophloea trees from six designated plots around Lake Nakuru, Kenya. Motivated by the need to detect fallen trees around the lake, two well-established deep neural networks, i.e. Faster Region-based Convolution Neural Network (Faster R-CNN) and Retina-Net were used for fallen tree detection. A total of 7,590 annotations of three classes on 256×256 image patches were used for this study. Experimental results show the relevance of deep learning in this context, with Retina-Net model achieving 38.9% precision and 57.9% recall.
Article
Full-text available
Tree degradation in National Parks poses a serious risk to the birds and animals and to a larger extent the general ecosystem. The essence of Forest degradation mapping is to detect the extent of damage on the trees over time, hence providing stakeholders with a basis for forest rehabilitation and intervention. The study proposes a workflow for detection and classification of degrading acacia vegetation along Lake Nakuru riparian reserve. Inspired by previous research on the use of Dual Polarized Sentinel 1 Ground Range Detected (GRD) data for vegetation detection, a set of six Sentinel 1 GRD and Sentinel 2 MSI of corresponding dates (2018–2019) were used. Our study confirms the existing correlation between vegetation indices derived from optical sensors and the backscatter indices from S1 SAR image of the same land cover classes. Factors that were used in validating the results include some comparisons between pixelwise and object-based classification, with a focus on the underlying segmentation and classification algorithms, the polarimetric attributes (VV+VH intensity bands) and the reflectance bands (NIR, SWIR & GREEN), the Haralick features (GLCM) vs. some geometric attributes (area & moment of inertia). Classification carried out on the temporal datasets considering geometric attributes and the Random Forest classifier yielded the highest Overall Accuracy (OA) with 94.25 %, and a Kappa coefficient of 0.90.
Article
Full-text available
Mapping of green vegetation in urban areas using remote sensing techniques can be used as a tool for integrated spatial planning to deal with urban challenges. In this context, multitemporal (MT) synthetic aperture radar (SAR) data have not been equally investigated, as compared to optical satellite data. This research compared various machine learning methods using single-date and MT Sentinel-1 (S1) imagery. The research was focused on vegetation mapping in urban areas across Europe. Urban vegetation was classified using six classifiers-random forests (RF), support vector machine (SVM), extreme gradient boosting (XGB), multi-layer perceptron (MLP), AdaBoost.M1 (AB), and extreme learning machine (ELM). Whereas, SVM showed the best performance in the single-date image analysis, the MLP classifier yielded the highest overall accuracy in the MT classification scenario. Mean overall accuracy (OA) values for all machine learning methods increased from 57% to 77% with speckle filtering. Using MT SAR data, i.e., three and five S1 imagery, an additional increase in the OA of 8.59% and 13.66% occurred, respectively. Additionally, using three and five S1 imagery for classification, the F1 measure for forest and low vegetation land-cover class exceeded 90%. This research allowed us to confirm the possibility of MT C-band SAR imagery for urban vegetation mapping.
Article
Full-text available
This paper discusses the thresholding and unsupervised classification methodologies, in order to find the inundated areas due to incessant rains and rise of water level in Rapti and Ghaghara Rivers during the month of August 2017. Used high resolution multi-temporal Synthetic Aperture Radar (SAR) and Optical images captured during the August 2017 floods in the state of Uttar Pradesh, India. The zonal statistics are calculated to find the inundated area in each district of the selected Area of Study (AOS). District-wise flood water mapping is done by superimposing the extracted water layer. The obtained results are validated against the meteorological observations. From the findings, it is evident that SAR data can effectively be used for flood water mapping and flood monitoring. These findings will therefore help to minimize the flood hazard impact and aid in augmenting the flexibility in flood management. © 2019 National Authority for Remote Sensing and Space Sciences
Article
Full-text available
This paper introduces an extension of morphological attribute profiles (APs) by extracting their local features. The so-called local feature-based APs (LFAPs) are expected to provide a better characterization of each APs' filtered pixel (i.e., APs' sample) within its neighborhood, and hence better deal with local texture information from the image content. In this paper, LFAPs are constructed by extracting some simple first-order statistical features of the local patch around each APs' sample such as mean, standard deviation, and range. Then, the final feature vector characterizing each image pixel is formed by combining all local features extracted from APs of that pixel. In addition, since the self-dual APs (SDAPs) have been proved to outperform the APs in recent years, a similar process will be applied to form the local feature-based SDAPs (LFSDAPs). In order to evaluate the effectiveness of LFAPs and LFSDAPs, supervised classification using both the random forest and the support vector machine classifiers is performed on the very high resolution Reykjavik image as well as the hyperspectral Pavia University data. Experimental results show that LFAPs (respectively, LFSDAPs) can considerably improve the classification accuracy of the standard APs (respectively, SDAPs) and the recently proposed histogram-based APs.
Article
Sentinel-1 Synthetic Aperture Radar (SAR) data have provided an unprecedented opportunity for crop monitoring due to its high revisit frequency and wide spatial coverage. The dual-pol (VV-VH) Sentinel-1 SAR data are being utilized for the European Common Agricultural Policy (CAP) as well as for other national projects, which are providing Sentinel derived information to support crop monitoring networks. Among the Earth observation products identified for agriculture monitoring, indicators of vegetation status are deemed critical by end-user communities. In literature, several experiments usually utilize the backscatter intensities to characterize crops. In this study, we have jointly utilized the scattering information in terms of the degree of polarization and the eigenvalue spectrum to derive a new vegetation index from dual-pol (DpRVI) SAR data. We assess the utility of this index as an indicator of plant growth dynamics for canola, soybean, and wheat, over a test site in Canada. A temporal analysis of DpRVI with crop biophysical variables (viz., Plant Area Index (PAI), Vegetation Water Content (VWC), and dry biomass (DB)) at different phenological stages confirms its trend with plant growth dynamics. For each crop type, the DpRVI is compared with the cross and co-pol ratio (σVH⁰/σVV⁰) and dual-pol Radar Vegetation Index (RVI = 4σVH⁰/(σVV⁰ + σVH⁰)), Polarimetric Radar Vegetation Index (PRVI), and the Dual Polarization SAR Vegetation Index (DPSVI). Statistical analysis with biophysical variables shows that the DpRVI outperformed the other four vegetation indices, yielding significant correlations for all three crops. Correlations between DpRVI and biophysical variables are highest for canola, with coefficients of determination (R²) of 0.79 (PAI), 0.82 (VWC), and 0.75 (DB). DpRVI had a moderate correlation (R²≳ 0.6) with the biophysical parameters of wheat and soybean. Good retrieval accuracies of crop biophysical parameters are also observed for all three crops.
Article
Crop growth monitoring using compact-pol Synthetic Aperture Radar (CP-SAR) data is gaining attention with the rapid advancements toward operational applications. In this study, we propose a vegetation index for compact polarimetric (CP) SAR data (CpRVI). The CpRVI is derived using the concept of a geodesic distance between Kennaugh matrices projected on a unit sphere. This distance is utilized to compute a similarity measure between the observed Kennaugh matrix and the Kennaugh matrix of an ideal depolarizer (a realization of vegetation canopy). The similarity measure is then modulated with a scaled quantity derived from the scattering power ratio of the same and opposite sense polarization with respect to the transmitted circular polarization. In this work, we utilize time series simulated RADARSAT Constellation Mission (RCM) compact-pol SAR data (RH-RV) obtained from the full-pol RADARSAT-2 observations during the Soil Moisture Active Passive (SMAP) Validation Experiment 2016 (SMAPVEX16-MB) campaign in Manitoba, Canada to assess the proposed vegetation index. Among the various crops grown in this region, in particular, we analyze the growth stages of wheat and soybean due to their different canopy structures. A temporal analysis of the proposed CpRVI with crop biophysical parameters (the Plant Area Index-PAI, and Vegetation Water Content-VWC) at different phenological stages confirms the trend of CpRVI with the plant growth. Nevertheless, variations of CpRVI values are apparent with different plant densities for both the crop types. Also, the linear regression analysis confirms that the CpRVI values significantly correlate with PAI (r = 0.72 and 0.85) and VWC (r = 0.62 and 0.75) for both wheat and soybean. We observed good retrieval of PAI and VWC for both wheat and soybean.
Article
Semantic segmentation plays an important role in remote sensing image understanding. Great progress has been made in this area with the development of Deep Convolutional Neural Networks (DCNNs). However, due to the complexity of ground objects’ spectrum, DCNNs with simple classifier have difficulties in distinguishing ground object categories even though they can represent image features effectively. Additionally, DCNN-based semantic segmentation methods learn to accumulate contextual information over large receptive fields that causes blur on object boundaries. In this work, a novel approach named Superpixel-enhanced Deep Neural Forest (SDNF) is proposed to target the aforementioned problems. To improve the classification ability, we introduce Deep Neural Forest (DNF), where the representation learning of deep neural network is conducted by a completely differentiable decision forest. Therefore, better classification accuracy is achieved by combining DCNNs with decision forests in an end-to-end manner. In addition, considering the homogeneity within superpixels and heterogeneity between superpixels, a Superpixel-enhanced Region Module (SRM) is proposed to further alleviate the noises and strengthen edges of ground objects. Experimental results on the ISPRS 2D semantic labeling benchmark demonstrate that our model significantly outperforms state-of-the-art methods thus validate the efficiency of our proposed SDNF.
Article
Wetlands are extraordinary ecosystems and important climate regulators that also contribute to reduce natural disaster risk. Unfortunately, wetlands are declining much faster than forests. The safeguarding of the wetlands also needs knowledge of the dynamics that control the water balance of these environments. Therefore, an accurate estimation of evapotranspiration in wetlands is an essential task. When adequate experimental data are available, some algorithms deriving from Artificial Intelligence research represent a promising alternative to the most common estimation techniques. In this study, starting from daily measurements of climatic variables such as net solar radiation, depth to water, wind speed, mean relative humidity, maximum temperature, minimum temperature, and mean temperature, using the Random Forest, Additive Regression of Decision Stump, Multilayer Perceptron and k-Nearest Neighbors algorithms, 24 estimation models, different in input variables, have been developed and compared. The data have been provided by USGS. They have been obtained from a measuring site in wetlands of Indian River County, Florida using the eddy-covariance technique. The accuracy of these models based on AI algorithms remains good even if the number of input variables is reduced from 7 to 3. Net solar radiation, mean temperature and mean relative humidity or wind speed measurements allow obtaining a sufficiently accurate estimation model. Random Forest and k-Nearest Neighbors provide slightly better performance than Additive Regression of Decision Stump and Multilayer Perceptron. The analyzed models show in most cases the lowest accuracy in the range 2–4 mm/day, while the highest accuracy is obtained in the ranges 0–2 mm/day and 6–8 mm/day, with the exception of the models based on the Additive Regression, which show similar levels of accuracy in the different considered sub-intervals.