=Paper=
{{Paper
|id=Vol-2380/paper_60
|storemode=property
|title=Location-Based Species Recommendation - GeoLifeCLEF 2019 Challenge
|pdfUrl=https://ceur-ws.org/Vol-2380/paper_60.pdf
|volume=Vol-2380
|authors=Costel-Sergiu Atodiresei,Adrian Iftene
|dblpUrl=https://dblp.org/rec/conf/clef/AtodireseiI19
}}
==Location-Based Species Recommendation - GeoLifeCLEF 2019 Challenge==
https://ceur-ws.org/Vol-2380/paper_60.pdf
Location-Based Species Recommendation GeoLifeCLEF
2019 Challenge
Costel-Sergiu Atodiresei, Adrian Iftene
UAIC: Faculty of Computer Science, “Alexandru Ioan Cuza” University, Romania
sergiu.atodiresei@gmail.com, adiftene@info.uaic.ro
Abstract. In this paper, it is presented a system built with the aim to predict plant
species given their location in the context of the GeoLifeCLEF 2019 challenge.
There are developed several prediction models on the basis of a representation in
the environmental space (a feature vector composed of climatic variables and
other variables such as soil type, land cover, distance to water) and using the
spatial occurrences of species (latitude and longitude): Random Forest, K-
Nearest Neighbors, Artificial Neural Networks and XGBoost.
Keywords: GeoLifeCLEF, Random Forest, K-Nearest Neighbors, Artificial
Neural Networks, XGBoost.
1 Introduction
Automatically predicting the list of species that are the most likely to be observed at a
given location is useful for many scenarios in biodiversity informatics. First of all, it
could improve species identification processes and tools by reducing the list of
candidate species that are observable at a given location (be they automated, semi-
automated or based on classical field guides or flora). More generally, it could facilitate
biodiversity inventories through the development of location-based recommendation
services (typically on mobile phones) as well as the involvement of non-expert nature
observers. Last but not least, it might serve educational purposes thanks to biodiversity
discovery applications providing functionalities such as contextualized educational
pathways [1].
The aim of the challenge is to predict the list of species that are the most likely to be
observed at a given location. In the previous edition, in the context of the GeoLifeCLEF
2018 challenge1, several approaches for plant predictions given their location were
used. The winners of the challenge 2018 have developed three kinds of prediction
models, one convolutional neural network on environmental data (CNN), one neural
network on co-occurrences data and two other models only based on the spatial
1 https://www.imageclef.org/GeoLifeCLEF2019
Copyright (c) 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0). CLEF 2019, 9-12 September 2019, Lugano,
Switzerland.
occurrences of species. Results show the effectiveness of the CNN which obtained the
best prediction score of the whole GeoLifeCLEF challenge. The fusion of this model
with the spatial ones only provides slight improvements suggesting that the CNN
already captured most of the spatial information in addition to the environmental
preferences of the plants [2]. Location-based species prediction is very similar with the
problem Species distribution modelling (SDM), also known as environmental (or
ecological) niche modelling (ENM), habitat modelling, or predictive habitat
distribution modelling, which uses computer algorithms to predict the distribution of a
species across geographic space and time using environmental data. The environmental
data are most often climate data (e.g. temperature, precipitation), but can include other
variables such as soil type, water depth, and land cover. SDMs are used in several
research areas in conservation biology, ecology and evolution. Predictions of current
and/or future habitat suitability can be useful for management applications (e.g.
reintroduction or translocation of vulnerable species, reserve placement in anticipation
of climate change) [3].
In this paper are presented 4 machine learning models for predicting plant species in
the context of the GeoLifeCLEF 2019 challenge [4]: (1) K-Nearest Neighbors: (i)
Based on the environmental data (like temperature, precipitation, soil type, water depth,
land cover etc.); (ii) Based on the spatial occurrences of species (a spatial model), (2)
Random Forest: (i) Based on the environmental data; (ii) Based on the spatial
occurrences of species, (3) Artificial Neural Networks based on the environmental data,
(4) XGBoost: (i) Based on the environmental data; (ii) Based on the spatial occurrences
of species. After splitting the train dataset in 90% (for training) and 10% (for testing),
Random Forest and K-NN captured most of the environmental information.
2 Data and Evaluation Methodology
GeoLifeCLEF provides a large training set of species occurrences, each occurrence
being associated to a multi-channel image characterizing the local environment. Indeed,
it is usually not possible to learn a species distribution model directly from spatial
positions because of the limited number of occurrences and the sampling bias.
The train dataset includes 280,945 train and test georeferenced occurrences of plant
species from last year (file GLC_2018.csv). Plus, 2,367,145 plant species occurrences
with uncertain identifications are added (file PL_complete.csv). They come from
automatic species identification of pictures produced in 2017-2018 by the smartphone
application Pl@ntNet, where users are mainly amateur botanists [5]. A trusted
extraction of this dataset is also provided (file PL_trusted.csv), insuring a reasonable
level of identification certainty. Finally, 10,618,839 species occurrences from other
kingdoms (as mammals, birds, amphibians, insects, fungi’s etc.) were selected from the
GBIF database (file noPlant.csv). 33 environmental raster’s (constructed from various
open datasets) covering the French territory are made available, so that each occurrence
may be linked to an environmental tensor via a participant customizable Python code
[6]. The test occurrences data come from independents datasets of the French National
Botanical Conservatories. This TestSet includes 844 plant species. It is a subset of those
found in the train set. The main evaluation criteria will be the accuracy based on the 30
first answers, also called Top 30. It is the mean of the function scoring 1 when the good
species is in the 30 first answers, and 0 otherwise, over all test set occurrences.
3 Proposed Solution
3.1 General Architecture of the System
The general architecture of the system is presented in Fig. 1. There is a Processing
Training Data Module, which extracts and augments (with climatic variables and
other variables such as soil type, land cover, distance to water, etc.) plant species
occurrences with certain and uncertain identifications.
Fig. 1. System architecture
The output of this module consists of 2 .csv files. The first one is made of spatial
occurrences of species (latitude and longitude) and the second one is composed of 29
environmental variables. This files are the input for the Machine Learning Module
where our 4 machine learning models are used to predict 25,000 plant species
occurences from the test data. Finally, the Predictions Collector Module generates 20
run files based on the output (32 .csv files) from the previous module, each run file
containing the prediction of a particular model, or mixed predictions from several/all
models. These runs has been submitted to be evaluated within the GeoLifeCLEF 2019
challenge.
3.2 Implementation Details
The system is based of 3 modules, developed in Python and using several libraries like
Pandas (providing high-performance, easy-to-use data structures and data analysis
tools) [7], NumPy (for scientific computing) [8], Scikit-learn (contains machine
learning algorithms including K-Nearest Neighbors, Random Forest and also
preprocessing) [9] similar to work from [10, 11], Keras (high-level neural networks
API) [12], XGBoost (implements machine learning algorithms under the Gradient
Boosting framework) [13].
3.3 Main Modules
I. Processing Training Data Module
This module processes 2 datasets: (1) PL_complete.csv with 2,367,145 plant species
occurrences with uncertain identifications, (2) PL_trusted.csv - a trusted extraction of
the complete dataset (approx. 10%) insuring a reasonable level of identification
certainty.
The output consists in 4 files: (1, 2) PL_complete_core.csv, PL_trusted_core.csv -
based on the spatial occurrences of species and have 3 columns containing in this order:
latitude, longitude, glc19SpId. The values of latitude and longitude are rounded to 4
decimals to avoid overfitting as much as possible in the machine learning models, (3,
4) PL_complete_env.csv, PL_trusted_env.core - based on the environmental variables
and have 30 columns. Each plant occurrence was linked to environmental tensors via a
customizable Python code provided in the context of the challenge and 29 climatic
variables and other variables were extracted: chbio_1, chbio_3, chbio_4, chbio_5,
chbio_6, chbio_7, chbio_8, chbio_9, chbio_10, chbio_11, chbio_12, chbio_13,
chbio_14, chbio_15, chbio_16, chbio_17, chbio_18, chbio_19, etp, alti, awc_top,
bs_top, cec_top, crusting, dgh, dimp, erodi, oc_top, pd_top. More variables were
available, but those didn’t provide a clear added value. Each occurrence ID in the
submitted run must exist in the testSet file (glc19TestOccId). Each predicted plant
species (glc19SpId) must be one of the species marked as TRUE in the column “test”
of the Table of species Ids and names and identification of test set species [14]. As a
consequence, to predict the test dataset, the .csv files are filtered and contain only the
occurrences of species marked as TRUE in the test set species.
II. Machine Learning Module
A. K-Nearest Neighbors Classifier
One of the machine learning models used is K-Nearest Neighbors Classifier, where the
output is the class membership (e.g. plant species id - glc19SpId) most common among
its k nearest neighbors [15]. The distance between neighbors is determined with the
Euclidean metric.
Multiple K-NN models were developed with 1, 3, and 5 neighbors. No more than 5
neighbors were used because, as can be seen in the section 4, the accuracy decreases as
many trees are used. Finally, the predictions from several/all models were merged, as
stated in section 3.3.III.
This classifier generates 12 different predictions (.csv files), depending to the
number of the nearest neighbors, if the training dataset has certain and uncertain
identifications and based on the spatial occurrences of species or based on the
environmental data. Considering that, the K-NN predicts the following lists of species:
(1) knn_spatial_[1|3|5]nn_complete.csv - 1/3/5 nearest neighbor(s), the complete
training dataset is used and is based on the spatial occurrences of species (longitude,
latitude); (2) knn_spatial_[1|3|5]nn_trusted.csv - 1/3/5 nearest neighbor(s), the trusted
training dataset is used and is based on the spatial occurrences of species (longitude,
latitude); (3) knn_env_[1|3|5]nn_complete.csv - 1/3/5 nearest neighbor(s), the complete
training dataset is used and is based on the environmental data; (4)
knn_env_[1|3|5]nn_trusted.csv - 1/3/5 nearest neighbor(s), the trusted training dataset
is used and is based on the environmental data.
B. Random Forest Classifier
Another machine learning algorithm used is Random Forest Classifier that operates by
constructing a multitude of decision trees at training time and outputting the class that
is the mode (the value that appears most often) of the classes [16]. First, feature scaling
is applied to the train and test data by removing the mean and scaling to unit variance.
Then the Random Forest Classifier is fitted to the training set. The number of trees in
the forest is 10 and the function to measure the quality of a split is “entropy” criteria
for the information gain.
This classifier generates 12 different predictions (.csv files), depending to random
state (the seed used by the random number generator), if the training dataset has certain
and uncertain identifications and based on the spatial occurrences of species or based
on the environmental data. Considering that, the K-NN model predicts the following
lists of species: (1) random_forest_spatial_[1|2|3]_complete.csv - random state = 1/2/3,
the complete training dataset is used and is based on the spatial occurrences of species
(longitude, latitude); (2) random_forest_spatial_[1|2|3]_trusted.csv - random state =
1/2/3, the trusted training dataset is used and is based on the spatial occurrences of
species (longitude, latitude); (3) random_forest_env_[1|2|3]_complete.csv - random
state = 1/2/3, the complete training dataset is used and is based on the environmental
data; (4) random_forest_env_[1|2|3]_trusted.csv - random state = 1/2/3, the trusted
training dataset is used and is based on the environmental data.
A Random Forest Classifier with 8 number or trees was developed, but as stated in
the section 4, the accuracy was smaller than the model with 10 trees.
C. Artificial Neural Networks
An artificial neural network is an interconnected group of nodes called artificial
neurons, which loosely model the neurons in a biological brain. Each connection, like
the synapses in a biological brain, can transmit a signal from one artificial neuron to
another. An artificial neuron that receives a signal can process it and then signal
additional artificial neurons connected to it [17]. First, label binarizer is applied to the
output variable (glc19SpId) to convert multi-class labels to binary labels (belong or
does not belong to the class). Then feature scaling is applied to the train and test data.
A Neural Network model is based on the Sequential model which is a linear stack of
layers. The model needs to know what input shape it should expect. In consequence,
the input layer in a Sequential model (and only the first, because following layers can
do automatic shape inference) is a Dense layer where the input shape is specified via
the argument input_dim. In our case the input shape is equal to a feature vector shape
(composed of 29 climatic variables and other variables such as soil type, land cover,
distance to water) used to train this machine learning model. The Rectified Linear Unit
function (relu) is passed to the input and hidden layers as the activation argument. For
the output layer, the Softmax activation function is used. The system makes use of 3
Sequential models with a different number of hidden layers based on the geometric
pyramid rule (the Masters rule):
a) for one hidden layer the number of neurons in the hidden layer is equal to:
= ×
where nrHidden – the number of neurons in the hidden layer; nrInput – the number
of neurons in the input layer; nrOutput – the number of neurons in the output layer.
b) for two hidden layers:
= 3 ×
1 = ×r
2 = ×
where nrHidden1 – the number of neurons in the first hidden layer; nrHidden2 – the
number of neurons in the second hidden layer.
c) for three hidden layers:
= 4 ×
1 = ×r
2 = ×r
3 = ×
where nrHidden1 – the number of neurons in the first hidden layer; nrHidden2 – the
number of neurons in the second hidden layer; nrHidden3 – the number of neurons in
the third hidden layer.
In our case, we consider 3 models: (1) the first model - trained with observations
from the trusted dataset has 29 neurons in the input layer, one for each environmental
and climatic feature, 197 neurons in the hidden layer, and 1348 neurons in the output
layer, one for each plant species id, then trained with observations from the complete
dataset has 29 neurons in the input layer, one for each environmental and climatic
feature, 336 neurons in the hidden layer, and 3096 neurons in the output layer, one for
each plant species id, (2) the second model - trained with observations from the trusted
dataset has 29 neurons in the input layer, 1089 neurons in the first hidden layer, 33
neurons in the second hidden layer, and 1348 neurons in the output layer, and then
trained with observations from the complete dataset has 29 neurons in the input layer,
2304 neurons in the first hidden layer, 48 neurons in the second hidden layer, and 3096
neurons in the output layer, and (3) the third model - trained with observations from
the trusted dataset has 29 neurons in the input layer, 2744 neurons in the first hidden
layer, 196 neurons in the second hidden layer, 14 neurons in the third hidden layer, and
1348 neurons in the output layer, and then trained with observations from the complete
dataset has 29 neurons in the input layer, 5832 neurons in the first hidden layer, 324
neurons in the second hidden layer, 18 neurons in the third hidden layer, and 3096
neurons in the output layer.
This classifier generates 6 different predictions (.csv files), depending to the model
used and if the training dataset has certain and uncertain identification. Considering
that, the ANN predicts the following lists of species: (1) ann_[1|2|3]_complete_env.csv
- the 1st/2nd/3rd model is used and is trained with the complete dataset; (2)
ann_[1|2|3]_trusted_env.csv - the 1st/2nd/3rd model is used and is trained with the
trusted dataset.
D. XGBoost
XGBoost is an optimized distributed gradient boosting library designed to be highly
efficient, flexible and portable. It implements machine learning algorithms under the
Gradient Boosting framework [18].
First, feature scaling is applied to the train and test data by removing the mean and
scaling to unit variance. Then the XGBClassifer (implementation of the scikit-learn
API for XGBoost classification.) is fitted to the training set. The number of trees to fit
is 100.
This classifier generates 2 different predictions (.csv files), based on the spatial
occurrences of species or based on the environmental data:
● xgboost_env_trusted.csv - the trusted training dataset is used and is based on
the environmental data;
● xgboost_spatial_trusted.csv - the trusted training dataset is used and is based
on the spatial occurrences of species (longitude, latitude);
III. Predictions Collector Module
All the predictions lists (.csv files) are collected and processed to generate 20 run files,
each run file containing the predictions of a particular model, or mixed predictions from
several/all models. A run is a .csv file with 4 columns separated by “;” and containing
in this order: glc19TestOccId ; glc19SpId ; Rank ; Probability. The models who predict
the same plant species (glc19SpId) are combined in a single prediction (row in the .csv)
with the probability being equal with the sum of probabilities associated with each
model. These runs (including the algorithms who produced the predictions list) has
been submitted to be evaluated within the GeoLifeCLEF 2019 challenge:
1) run_1 (Visual retrieval type): K-Nearest Neighbors algorithm (1-NN) - Based on
the environmental data (trusted dataset).
2) run_2 (Visual retrieval type): K-Nearest Neighbors algorithm (1-NN) - Based on
the environmental data (complete dataset).
3) run_3 (Visual retrieval type): K-Nearest Neighbors algorithm (3-NN and 5-NN) -
Based on the environmental data (trusted and complete datasets).
4) run_4 (Textual retrieval type): K-Nearest Neighbors algorithm (1-NN) - Based on
the spatial occurrences of species (trusted dataset).
5) run_5 (Textual retrieval type): K-Nearest Neighbors algorithm (1-NN) - Based on
the spatial occurrences of species (complete dataset).
6) run_6 (Textual retrieval type): K-Nearest Neighbors algorithm (3-NN and 5-NN) -
Based on the spatial occurrences of species (trusted and complete datasets). The
probabilities associated with each model are the same as for run_3.
7) run_7 (Mixed retrieval type): K-Nearest Neighbors algorithm (1-NN) - Based on
the spatial occurrences of species and the environmental data (trusted and complete
datasets).
8) run_8 (Visual retrieval type): Random Forest algorithm - Based on the
environmental data (trusted dataset).
9) run_9 (Visual retrieval type): Random Forest algorithm - Based on the
environmental data (complete dataset).
10) run_10 (Visual retrieval type): Random Forest algorithm - Based on the
environmental data (trusted and complete datasets).
11) run_11 (Textual retrieval type): Random Forest algorithm - Based on the spatial
occurrences of species (trusted dataset).
12) run_12 (Textual retrieval type): Random Forest algorithm - Based on the spatial
occurrences of species (complete dataset).
13) run_13 (Textual retrieval type): Random Forest algorithm - Based on the spatial
occurrences of species (trusted and complete datasets). The probabilities associated
with each model are the same as for run_10.
14) run_14 (Mixed retrieval type): Random Forest algorithm - Based on the spatial
occurrences of species and the environmental data (trusted and complete datasets).
15) run_15 (Mixed retrieval type): K-Nearest Neighbors algorithm (1-NN) and
Random Forest algorithm - Based on the spatial occurrences of species and the
environmental data (trusted dataset). The probability associated with each model is
0.5.
16) run_16 (Mixed retrieval type): K-Nearest Neighbors algorithm (1-NN) and
Random Forest algorithm - Based on the spatial occurrences of species and the
environmental data (complete dataset). The probability associated with each model
is 0.5.
17) run_17 (Mixed retrieval type): K-Nearest Neighbors algorithm (1-NN) and
Random Forest algorithm - Based on the spatial occurrences of species and the
environmental data (trusted and complete datasets).
18) run_18 (Visual retrieval type): Artificial neural networks (ANN) - Based on the
environmental data (trusted and complete datasets). The ANN model with 1 and 2
hidden layers trained with complete dataset aren’t used in this run because the
accuracy obtained in the validation experiments is lower than the model with 3
hidden layers.
19) run_19 (Mixed retrieval type): XGBoost - Based on the spatial occurrences of
species and the environmental data (trusted dataset). The probability associated with
each model is 0.5.
20) run_20 (Visual retrieval type): K-Nearest Neighbors algorithm (1-NN), Random
Forest, Artificial neural networks (ANN), XGBoost - Based on the environmental
data (trusted dataset).
4 Experiments and evaluation
4.1 Experimental results
To estimate the skill of a machine learning model the k-Fold Cross-Validation is used.
Cross-validation is a statistical method used to evaluate machine learning models. It
has a single parameter called k that indicates the number of groups that a given data
sample is going to be split. Because of that, the method is called k-Fold Cross-
Validation. The KFold() scikit-learn class [19] is used to split a given dataset into 10
consecutive folds (k=10). The cross_val_score scikit-learn function [20] is called on
the classifier and the fold. The results of a k-fold cross-validation run are summarized
with the mean of the model skill scores and the standard deviation obtained from the
cross_val_score function [21].
For the K-Nearest Neighbors models the following results were obtained:
Dataset Number of Mean Standard
neighbors deviation
Based on the spatial occurrences of 1 37.85% 0.27%
species (trusted dataset) 3 22.86% 0.22%
5 17.55% 0.21%
Based on the spatial occurrences of 1 26.58% 0.06%
species (complete dataset) 3 17.90% 0.06%
5 14.48% 0.07%
Based on the environmental data 1 31.76% 0.26%
(trusted dataset) 3 20.84% 0.22%
5 16.70% 0.28%
Based on the environmental data 1 15.46% 0.05%
(complete dataset) 3 12.11% 0.05%
5 10.86% 0.03%
The best model is that with 1 nearest neighbor representing run_4. The differences
between the model based on the environmental data and the one based on the spatial
occurrences are quite significant. If the test set contains occurrences in areas without
unnoticed species, the first model (trained with a feature vector composed of climatic
variables and other variables such as soil type, distance to water) will do a better job in
predicting the actual plant species. The complete dataset has three times more plant
species than the trusted dataset so differences between the models trained with these
datasets is normal considering that some species are present more often than others.
For the Random Forest models the following results were obtained:
Dataset Mean Standard deviation
Based on the spatial occurrences of species
36.88% 0.26%
(trusted dataset)
Based on the spatial occurrences of species
25.22% 0.08%
(complete dataset)
Based on the environmental data (trusted
30.20% 0.29%
dataset)
Based on the environmental data (complete
13.89% 0.06%
dataset)
The best model is based on the spatial occurrences of species (trusted dataset)
representing run_11. The conclusion is the same as for the K-NN model. The
differences between the model based on the environmental data and the one based on
the spatial occurrences are quite considerable.
For the Artificial Neural Network models the following results were obtained:
Dataset Number of Mean Standard
hidden layers deviation
Based on the environmental 1 6.46% 0.22%
data (trusted dataset) 2 6.59% 0.27%
3 6.71% 0.24%
Based on the environmental 1 6.28% 0.05%
data (complete dataset) 2 6.41% 0.03%
3 6.65% 0.05%
The differences between models are negligible. The models built with 3 hidden
layers have achieved slightly better results (run_18).
For the XGBoost models (run_19) the following results were obtained:
Dataset Mean Standard deviation
Based on the spatial occurrences
9.38% 0.28%
of species (trusted dataset)
Based on the environmental data
8.24% 0.06%
(trusted dataset)
4.2 Challenge Results
This year two criteria’s where used to evaluate the runs:
● The accuracy based on the 30 first answers, also called Top30. It is the mean
of the function scoring 1 when the good species is in the 30 first answers, and
0 otherwise, over all test set occurrences.
● The Mean Reciprocal Rank was chosen as secondary metric for enabling
comparison with the 2018 edition.
The results of the submitted runs to GeoLifeCLEF 2019 are:
runId top30 runName
26968 0.0205 run_14
26964 0.0191 run_10
26961 0.0190 run_7
26971 0.0184 run_17
26967 0.0180 run_13
26960 0.0168 run_6
27062 0.0159 run_20
26958 0.0146 run_3
26970 0.0102 run_16
26969 0.0099 run_15
26972 0.0089 run_18
26963 0.0079 run_9
26965 0.0068 run_11
26926 0.0067 run_4
26973 0.0064 run_19
26959 0.0063 run_5
26962 0.0062 run_8
26957 0.0061 run_2
26966 0.0058 run_12
26956 0.0058 run_1
The final results presented by the organizers show that the run_14, which is a list
predicted by a Random Forest model based on the spatial occurrences of species and
the environmental data (trusted and complete datasets), has achieved the best accuracy
of our 20 submitted runs. The next one is the run_10, also a list predicted by a Random
Forest model, this time trained with environmental data (trusted and complete datasets).
The third one is the run_7, generated by the 1-NN model based on the spatial
occurrences of species and the environmental data (trusted and complete datasets).
On the opposite side, the last places are occupied by run_1, run_12 and run_2. These
are produced by 1-NN model (based on the environmental data - trusted dataset),
Random Forest model (based on the spatial occurrences of species - complete dataset),
and again 1-NN model based on the environmental data, but this time using the
complete dataset. The predicted plant species from run_18 (ANN) and run_19
(XGBoost) have approximately half and one third, respectively, of the accuracy of the
predicted list from run_14.
It appears that Random Forest and 1-Neareast Neighbors predict the plant species
with the best accuracy, like in experimental results, only if a run contains a mixed
predictions from several models.
5 Challenge Results Analysis
We cannot rely solely on the spatial appearances because of environmental differences
(e.g. two species at 10 km can live in different environments, plain vs. sea coast).
Another issue is that some species are present more often than others (e.g. plants that
have been observed several times, others thousands of times), so we do not have
uniform observations, especially in the complete training dataset as can be seen in the
previous section (4). Last difficulty is that an occurrence means a punctual presence of
that species, and if it is missing, it does not mean that the species does not exist.
Unfortunately this leads to uncertainty in areas with unnoticed species.
An improvement would be to use more than 10 trees for the Random Forest
algorithm to provide a prediction list of 30 species.
6 Conclusions
This paper details the system developed with the aim to predict plant species given their
location for the 2019 edition of GeoLifeCLEF’s challenge. We presented and compared
four kinds of prediction models trained with the spatial occurrences of species and with
the environmental data: Random Forest, K-Nearest Neighbors, Artificial Neural
Networks and XGBoost. The results obtained using the Ranking metric and the Mean
Reciprocal Rank show that the fusion of mixed predictions from several Random Forest
models has the best prediction score of our 20 submitted runs, followed closely by
mixed predictions from several K-Nearest Neighbors models. The Random Forest
models captured most of the environmental information. For further development of
the system the noPlant.csv dataset can be used to extract interesting correlations
between plants and animals. This could lead to a more accurate prediction of plant
species because some animals live only where certain plants are present.
Acknowledgement. This work is partially supported by POC-A1-A1.2.3-G-2015
program, as part of the PrivateSky project (P 40 371/13/01.09.2016).
References
1. GeoLifeCLEF 2019 (Usage scenario section):
https://www.imageclef.org/GeoLifeCLEF2019 accessed last time on April, 2019.
2. Deneu, B., Servajean, M., Botella, C., Joly, A.: Location-based species recommendation
using co-occurrences and environment - GeoLifeCLEF 2018 challenge:
https://www.researchgate.net/publication/327542744_Location-based_species_
recommendation_using_co-occurrences_and_environment-GeoLifeCLEF_2018_challenge
accessed last time on April, 2019.
3. Elith, J., Leathwick, J. R.: Species Distribution Models: Ecological Explanation and
Prediction Across Space and Time. In Annual Review of Ecology, Evolution, and
Systematics, vol. 40 (1): pp. 677–697, (1009)
4. Botella, C., Servajean, M., Bonnet, P., Joly, A.: Overview of GeoLifeCLEF 2019: plant
species prediction using environment and animal occurrences. In CLEF working notes
2019, (2019)
5. Affouard, A., Goeau, H., Bonnet, P., Lombardo, J.-C., Joly, A.: Pl@ntnet app in the era of
deep learning. ICLR 2017-Workshop Track-5th International Conference on Learning
Representations, pp. 1-6, (2017)
6. LifeCLEF 2019 Geo (Challenge description, Data and Evaluation sections):
https://www.crowdai.org/challenges/lifeclef-2019-geo accessed last time on May, 2019.
7. Pandas Data Analysis Library: https://pandas.pydata.org accessed last time on May, 2019.
8. NumPy Package: https://www.numpy.org accessed last time on May, 2019.
9. Scikit-learn Machine Learning Library: https://scikit-learn.org/stable accessed last time on
May, 2019.
10. Cușmaliuc, C. G., Coca, L. G., Iftene, A.: Identifying Fake News on Twitter using Naive
Bayes, SVM and Random Forest Distributed Algorithms. Proceedings of The 13th Edition
of the International Conference on Linguistic Resources and Tools for Processing Romanian
Language (ConsILR-2018). ISSN: 1843-911X, pp. 177-188, (2018)
11. Șerban, C., Sirițeanu, A., Gheorghiu, C., Iftene, A., Alboaie, L., Breabăn, M.: Combining
image retrieval, metadata processing and naive Bayes classification at Plant Identification
2013. Notebook Paper for the CLEF 2013 LABs Workshop - ImageCLEF - Plant
Identification, 23-26 September, Valencia, Spain (2013)
12. Keras Deep Learning Library: https://keras.io accessed last time on May, 2019.
13. XGBoost Gradient Boosting Framework: https://keras.io accessed last time on May, 2019.
14. LifeCLEF 2019 Geo (Submission section): https://www.crowdai.org/challenges/lifeclef-
2019-geo accessed last time on May, 2019.
15. Altman, N. S.: An introduction to kernel and nearest-neighbor nonparametric regression. In
The American Statistician, vol. 46 (3), pp. 175–185, (1992)
16. Ho, T. K.: Random Decision Forests. In Proceedings of the 3rd International Conference on
Document Analysis and Recognition, Montreal, QC, 14–16 August 1995. pp. 278–282,
(1995)
17. Scherer, D., Müller, A. C., Behnke, S.: Evaluation of Pooling Operations in Convolutional
Architectures for Object Recognition. In 20th International Conference Artificial Neural
Networks (ICANN), pp. 92–101, (2010)
18. Chen, T., Guestrin, C.: XGBoost: A Scalable Tree Boosting System. In Krishnapuram,
Balaji; Shah, Mohak; Smola, Alexander J.; Aggarwal, Charu C.; Shen, Dou; Rastogi, Rajeev
(eds.). Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016. ACM. pp. 785–
794, (2016)
19. McLachlan, G. J., Do, K. A., Ambroise, C.: Analyzing microarray gene expression data.
Wiley, (2004)
20. Kohavi, R.: A study of cross-validation and bootstrap for accuracy estimation and model
selection. In Proceedings of the Fourteenth International Joint Conference on Artificial
Intelligence. San Mateo, CA: Morgan Kaufmann, vol. 2 (12), pp. 1137-1143, (1995)
21. Joly, A., Goëau, H., Botella, C., Kahl, S., Servajean, M., Glotin, H., Bonnet, P., Vellinga,
W. P., Planqué, R., Stöter, F. R., Müller, H.: Overview of LifeCLEF 2019: Identification of
Amazonian Plants, South \& North American Birds, and Niche Prediction. In Proceedings
of CLEF 2019, (2019)