Prediction of the species present at a location is useful for understanding biodiversity and for the purpose of conservation. The objective of the GeoLifeCLEF 2019 Challenge is to build a species recommendation system based on location and Environmental Variables (EVs). In this paper, we discuss di erent approaches to predict the most probable species based on location and EV values, using Machine Learning. We rst 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 Articial Neural Networks and the XGBoost framework. Our team achieved a maximum Top30 score of 0.1342 in the test phase, with an XGBoostbased model.
MaThe prediction of the species present at a location based on spatial and environmental parameters is of great use in understanding biodiversity. It greatly reduces e orts required to collect and analyse data, and allows for more research on the e ects 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
identi cation and understanding of living organisms, the LifeCLEF 2019
challenges [
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 occurrences, 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 otherwise. 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
The occurrences dataset PL trusted contained the Latitude, Longitude, Species ID and some other data. From this, we created three di erent datasets for our usage in di erent runs, based on the di erent 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 rst 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 [
ANN: We used an Arti cial Neural Network developed using the Keras library (Tensor ow 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 rst 2 Dense layers and the other after the next 2 Dense layers. The nal 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. Random Forest: This model was built using the in-built RandomForestClassier 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 XGBoost library. The di erences 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 categorical 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 26996 26997 27012 27013
Extracted EV Values
Single Central Average of 16 Central Average of 16 Central
Single Central Multiple ANNs: We developed a unique model which consisted of 5 di erent ANNs. We split the features - spatial coordinates and EV values - into 5 di erent 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 nal probability for each species. The architecture of all 5 ANNs was the same as used earlier (refer Fig. 2). The only di erence 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 rst 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 selected important features as inputs. The ANN used has an architecture similar to that of the ones used earlier (refer Fig. 2) but with di erent input dimension.
ANN 1 Latitude Longitude
alti awc top bs top cec top chbio 1
ANN 2 chbio 10 chbio 11 chbio 12 chbio 13 chbio 14 chbio 15 chbio 16
ANN 3 chbio 17 chbio 18 chbio 19 chbio 2 chbio 3 chbio 4 chbio 5
ANN 4 chbio 6 chbio 7 chbio 8 chbio 9 crusting dgh dimp
ANN 5 erodi etp oc top pd top proxi eau fast text clc We selected what we identi ed 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 classi cation. However, these runs were highly unsatisfactory and poorer than the rest of our approaches. Thus, we did not submit these runs for evaluation. 4
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
In the GeoLifeCLEF 2019 challenge, our team SSN CSE achieved a top
submission 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 [
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
[
In the future, our aim would be to enhance our current models by hyperparameter 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 di erent custom designed Neural Network architectures to improve performance on this task. 7
We thank SSN College of Engineering for allowing us to use the High Performance 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.