=Paper=
{{Paper
|id=Vol-2380/paper_71
|storemode=property
|title=Species Recommendation using Machine Learning - GeoLifeCLEF 2019
|pdfUrl=https://ceur-ws.org/Vol-2380/paper_71.pdf
|volume=Vol-2380
|authors=Nanda H Krishna,Praveen Kumar R,Ram Kaushik R,Mirunalini P,Aravindan Chandrabose,Jaisakthi S M
|dblpUrl=https://dblp.org/rec/conf/clef/KrishnaRRMAJ19
}}
==Species Recommendation using Machine Learning - GeoLifeCLEF 2019==
https://ceur-ws.org/Vol-2380/paper_71.pdf
Species Recommendation using Machine
Learning - GeoLifeCLEF 2019
Nanda H Krishna1? , Praveen Kumar R1? , Ram Kaushik R1? , P Mirunalini1 ,
Chandrabose Aravindan1 , and S M Jaisakthi2
1
Department of Computer Science and Engineering, SSN College of Engineering,
Kalavakkam, Chennai, India
2
School of Computer Science and Engineering, VIT University, Vellore, India
{nanda17093, praveenkumar17114, ramkaushik17125}@cse.ssn.edu.in
{miruna, aravindanc}@ssn.edu.in, jaisakthi.murugaiyan@vit.ac.in
Abstract. Prediction of the species present at a location is useful for
understanding biodiversity and for the purpose of conservation. The ob-
jective of the GeoLifeCLEF 2019 Challenge is to build a species recom-
mendation system based on location and Environmental Variables (EVs).
In this paper, we discuss different approaches to predict the most prob-
able species based on location and EV values, using Machine Learning.
We first developed purely spatial models which took only the spatial
coordinates as inputs. We then built models that took both the spatial
coordinates and EV values as inputs. For our runs, we mainly used Arti-
ficial Neural Networks and the XGBoost framework. Our team achieved
a maximum Top30 score of 0.1342 in the test phase, with an XGBoost-
based model.
Keywords: Species Recommendation · Environmental Variables · Ma-
chine Learning · XGBoost · ANN
1 Introduction
The prediction of the species present at a location based on spatial and en-
vironmental parameters is of great use in understanding biodiversity. It greatly
reduces efforts required to collect and analyse data, and allows for more research
on the effects of climate change, species invasion and other phenomena on the
biodiversity of a region.
With the goal of setting up robust information systems relying on automatic
identification and understanding of living organisms, the LifeCLEF 2019 chal-
lenges [3] were organised. Among these was the GeoLifeCLEF 2019 challenge
[4], the aim of which was to build a species recommendation system using the
Copyright c 2019 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0). CLEF 2019, 9-12 Septem-
ber 2019, Lugano, Switzerland.
?
These authors contributed equally.
given species occurrences and environmental parameters. Environmental Vari-
able (EV) values were given as TIF images, from which the patches for a particu-
lar location (latitude and longitude) could be extracted using a Python script [2].
Multiple datasets of species occurrences were provided for the contest. One
of these was the dataset with trusted occurrences that had an identification
confidence score of greater than 0.98, which was derived from the complete oc-
currences by applying a filter. This dataset, PL trusted, contained over 230,000
occurrences and over 1300 distinct species. We used this dataset as it provided
the most accurate occurrences with a good confidence score.
The test set for the challenge contained 25000 occurrence IDs for which the
species had to be predicted. There were 844 plant species in the test set occur-
rences, which is a subset of those found in the training sets. Thus, some species
present in the training data were non-test species (not present in the test set
occurrences).
The evaluation metric for the challenge was Top30, which is the mean of the
function scoring 1 if the good species is within the top 30 predicted, or 0 oth-
erwise. The metric is ideal as some tens of plant species usually coexist in the
perimeter of the location uncertainty of the occurrences. The Mean Reciprocal
Rank (MRR) was used as a secondary metric to enable comparison with previous
year results.
2 Data Preprocessing
The occurrences dataset PL trusted contained the Latitude, Longitude, Species
ID and some other data. From this, we created three different datasets for our
usage in different runs, based on the different models we had in mind.
2.1 Spatial Data
We extracted the spatial coordinates and Species ID to create a dataset for
training purely spatial models and also a baseline probability-based model.
2.2 Spatial and EV Data
We first created a dataset containing the spatial coordinates and the value of the
central pixel for each EV extracted from the EV image patches. Then we created
another dataset, containing the spatial coordinates, and the average value of the
16 central pixels extracted from each EV image patch. The values from the image
patches were extracted using Python scripts [2], as tensors. A sample generated
from the extractor for a few EVs is shown in Fig. 1. The same preprocessing was
also applied to the test set during prediction.
Fig. 1. Extractor Output Plot
3 Methodology
Our main approaches to this challenge were classifiers based on Artificial Neural
Networks using Keras [6], the Random Forest Classifier from scikit-learn [8] and
the XGBoost library [5].
3.1 Probability-based Model
We created this model for our understanding of the species distribution across
the whole dataset of occurrences, and submitted it as our baseline approach.
We used the occurrences data to obtain the number of occurrences of each in-
dividual species, and thus determined their probabilities. The list of species was
then sorted in descending order of probabilities, and the non-test species were
removed. From this, the top 50 species were chosen. For each test occurrence,
the same list of 50 species was assigned in the submission. This run (26821) had
a Top30 score of 0.0570.
3.2 Purely Spatial Models
The purely spatial models take only the spatial coordinates, that is, the latitude
and longitude of the occurrences as inputs, and output a list of probabilities of
the species. The predictions for each occurrence were sorted in descending order
of probabilities, following which the non-test species were removed. The top
30 species for each occurrence were chosen for the submission. We built purely
spatial models using XGBoost, ANNs and Random Forest Classifiers.
XGBoost: This model used the XGBoost framework, where we set the param-
eter eta to 0.1 and the objective function to XGBoost’s multi : sof tprob which
is used for multiclass classification. The num round parameter in training was
set to 1. This run (26988) had a Top30 score of 0.1063.
ANN: We used an Artificial Neural Network developed using the Keras li-
brary (Tensorflow backend) for this model. The Sequential model had 5 hidden
Dense layers with 256 units and the relu activation function. Two Dropout
layers with rate 0.02 were present, one after the first 2 Dense layers and the
other after the next 2 Dense layers. The final output layer had the number of
units set to the number of species in PL trusted, with sof tmax activation - to
predict class probabilities. The model was compiled with adam optimizer and
categorical crossentropy loss, for 10 epochs and a batch size of 2000. The Top30
score for this run (26875) was 0.0844. The summary of the model is shown in
Fig. 2.
Fig. 2. ANN Model Summary
Random Forest: This model was built using the in-built RandomForestClassi-
fier in the scikit-learn framework, with n estimators set to 10. The Top30 score
of this run (27102) was 0.0834.
3.3 Models Based on Spatial Coordinates and EV Values
These models had the spatial coordinates and the extracted EV values as their
inputs, and were trained to predict species probabilities. We made 8 submissions
based on these data, using XGBoost, a Multiple ANNs model and an ANN taking
selected features as inputs. In each approach, the non-test species were removed
from the list of predictions, and the top 30 species based on probability were
chosen to be submitted for each test occurrence.
XGBoost: We made 4 submissions (26996, 26997, 27012, 27013) using the XG-
Boost library. The differences between these runs was the value of the max depth
parameter of the model, and the dataset used in training. All EV values were
used in three runs (26997, 27012, 27013) while all EV values except the categor-
ical feature clc were used in one run (26996). The value of parameter eta was set
to 0.1, the objective was set to multi : sof tprob and the num round parameter
during training was set to 1 in all these runs. The details of the models can be
found in Table 1. It is to be noted that our top scoring submission was achieved
with this method (26997), with a Top30 score of 0.1342.
Run Extracted EV Values max depth Top30 Score
26996 Single Central None 0.1288
26997 Average of 16 Central None 0.1342
27012 Average of 16 Central 3 0.1263
27013 Single Central 3 0.1273
Table 1. XGBoost Models trained on Spatial Coordinates and EV values
Multiple ANNs: We developed a unique model which consisted of 5 different
ANNs. We split the features - spatial coordinates and EV values - into 5 different
mutually exclusive and exhaustive groups, each of which was the input to an
ANN. The outputs of each ANN were the probabilities of the various species.
The output vectors of each ANN, containing the probabilities, were averaged to
get the final probability for each species. The architecture of all 5 ANNs was the
same as used earlier (refer Fig. 2). The only difference is the input dimension,
which is based on the group of features sent as inputs to the ANNs. Also, the
feature clc was integer encoded before being passed to the ANNs. The features
sent to each ANN can be found in Table 2.
We made 2 submissions using the Multiple ANNs model (27064, 27067). The first
submission (27064) was made based on the dataset with EV values extracted
from the central pixel of the patches, and it obtained a Top30 score of 0.1198.
The second submission (27067) was made based on the dataset with EV values
extracted by averaging the central 16 pixel values, and it obtained a Top30 score
of 0.1135.
Selected Features ANN: Another approach we tried was an ANN with se-
lected important features as inputs. The ANN used has an architecture similar
to that of the ones used earlier (refer Fig. 2) but with different input dimension.
ANN 1 ANN 2 ANN 3 ANN 4 ANN 5
Latitude chbio 10 chbio 17 chbio 6 erodi
Longitude chbio 11 chbio 18 chbio 7 etp
alti chbio 12 chbio 19 chbio 8 oc top
awc top chbio 13 chbio 2 chbio 9 pd top
bs top chbio 14 chbio 3 crusting proxi eau fast
cec top chbio 15 chbio 4 dgh text
chbio 1 chbio 16 chbio 5 dimp clc
Table 2. Groups of Features for each ANN
We selected what we identified as important features based on data observation
and geological knowledge. Thus the features selected as inputs to the ANN were
Latitude, Longitude, alti, awc top, bs top, chbio 1, chbio 10, chbio 11, chbio 17,
chbio 18, chbio 19, chbio 2, chbio 3, erodi and etp.
Of the 2 submissions made using the Selected Features ANN, one (27069) used
the dataset with EV values extracted by averaging the central 16 pixel values
of the patches, while the other used the dataset with EV values extracted from
the central pixel alone. The Top30 scores of these submissions were 0.1227 and
0.1268 respectively.
3.4 Other Unsubmitted Methods
Initially, we had tried to use more advanced methods to approach this problem
such as ResNet and Convolutional Neural Networks. We did this because of
the great reputation of these Networks to problems such as image classification.
However, these runs were highly unsatisfactory and poorer than the rest of our
approaches. Thus, we did not submit these runs for evaluation.
4 Source Code and Computational Resources
We have uploaded our source code in the form of Jupyter Notebooks to a public
GitHub repository3 . Instructions are provided for installing requirements and
using the Notebooks. The resources we used were a 2.6 GHz Intel i7 CPU and
an NVIDIA 940M GPU. Our unsubmitted models were trained using a Google
Cloud VM instance with 8 CPUs.
5 Results
In the GeoLifeCLEF 2019 challenge, our team SSN CSE achieved a top submis-
sion rank of 6, with a best Top30 score of 0.1342. Overall, we were ranked 3rd.
The top rankers were team LIRMM with a best Top30 score of 0.1769 and team
3
https://github.com/nandahkrishna/GeoLifeCLEF2019
SaraSi with 0.1687. The overall results can be seen in Fig. 3 and on the challenge
website [1].
Fig. 3. Graph showing run-wise Top30 scores
6 Conclusion and Future Work
The overall results of the challenge show the difficulty in building species recom-
mendation systems. We approached this problem using various Machine Learning
techniques and evaluated their performance in this task. We thus learnt a great
deal about the uses and advantages of these methods.
Early on, we found that complex models such as ResNet did not perform greatly
in the task. This was shown by the results of last year’s edition of the challenge
[7]. The poor predicting power of our unsubmitted models could be attributed
to the curse of dimensionality, vanishing probabilities due to large number of
classes, and also the fact that the EV image patches are very different from
the traditional photographic images they are generally used for. Even a simple
probability-based model and purely spatial models outperformed these models.
Purely spatial models did not perform very bad, but models using the EV val-
ues greatly outperformed them. The XGBoost models produced good results
in last year’s edition of the challenge and we observe a repeat of that in our
submissions this year, often outperforming ANNs. However, the top submissions
involved species co-occurrence models which would have enhanced the predictive
power and thus performance in the challenge.
In the future, our aim would be to enhance our current models by hyper-
parameter tuning and the incorporation of co-occurrence based data. External
data sources and co-occurrence models could help in enhancing the results. We
also aim to explore different custom designed Neural Network architectures to
improve performance on this task.
7 Acknowledgements
We thank SSN College of Engineering for allowing us to use the High Perfor-
mance Computing Laboratory during our work for this challenge. We thank Dr.
M A Rajamamannan (Government Arts College, Coimbatore) for his help in
identifying the important features for species recommendation.
References
1. GeoLifeCLEF 2019 Challenge, https://www.imageclef.org/GeoLifeCLEF2019
2. GLC19 GitHub Repository, https://github.com/maximiliense/GLC19
3. Alexis Joly, Herv Goau, C.B.S.K.M.S.H.G.P.B.W.P.V.R.P.F.R.S.H.M.: Overview of
LifeCLEF 2019: Identification of Amazonian Plants, South & North American Birds,
and Niche Prediction. In: Proceedings of CLEF 2019 (2019)
4. Botella, C., Servajean, M., Bonnet, P., Joly, A.: Overview of GeoLifeCLEF 2019:
plant species prediction using environment and animal occurrences. In: CLEF work-
ing notes 2019 (2019)
5. Chen, T., Guestrin, C.: XGBoost: A Scalable Tree Boosting System. In: Pro-
ceedings of the 22nd ACM SIGKDD International Conference on Knowledge Dis-
covery and Data Mining. pp. 785–794. KDD ’16, ACM, New York, NY, USA
(2016). https://doi.org/10.1145/2939672.2939785, http://doi.acm.org/10.1145/
2939672.2939785
6. Chollet, F., et al.: Keras. https://keras.io (2015)
7. Christophe Botella, Pierre Bonnet, F.M.P.M.A.J.: Overview of GeoLifeCLEF 2018:
Location-based Species Recommendation (2018), http://ceur-ws.org/Vol-2125/
invited_paper_8.pdf
8. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine
Learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)