Difference between revisions of "Ecological Modeling"

Latest revision as of 14:51, 6 July 2016

Ecological Modeling is a set of functionalities available in gCube for performing data mining operations on biological data. It is available as a library and as a Service (Statistical Manager) in the infrastructure and is able to train models which can be combined with geographical information in order to produce projections on several environmental scenarios or time periods. This system allows for managing complex phenomena, in order, for example, to predict the impact of climate changes on biodiversity, prevent the spread of invasive species, identify geographical and ecological aspects of disease transmission, help in conservation planning, guide field surveys, among many other uses.

Overview

The library is endowed with a set of features which can be resumed as:

• GENERATORS: include probability distributions, classifications, matching or distance measurements etc.
• MODELING: includes models to be trained, e.g. neural networks, species envelopes, support vector machines etc.. The result will be typically a binary file.
• CLUSTERING: involves clustering procedures for grouping together phenomena or multidimensional points.
• TRANSDUCERS: involve algorithms for transforming a dataset into another.
• EVALUATORS: a set of procedures for measuring the quality of a model.

The system is currently able to run processes on the following computational platforms:

• LOCAL MULTICORE MACHINE
• RAINY CLOUD

Generative Algorithms

Currently the following algorithms are supported for projecting probability distributions on geographical maps:

• AQUAMAPS_SUITABLE: Aquamaps Suitable habitat production
• AQUAMAPS_NATIVE: Aquamaps Native habitat production
• AQUAMAPS_NATIVE_2050: Aquamaps Native for 2050 scenario
• AQUAMAPS_SUITABLE_2050: Aquamaps Suitable for 2050 scenario
• REMOTE_AQUAMAPS_SUITABLE: Aquamaps Suitable habitat generated by invoking Rainy Cloud
• REMOTE_AQUAMAPS_SUITABLE_2050: Aquamaps Suitable 2050 habitat generated by invoking Rainy Cloud
• AQUAMAPS_NATIVE_NEURALNETWORK: Aquamaps Native Distribution using a Feed-Forward Neural Network
• AQUAMAPS_SUITABLE_NEURALNETWORK: Aquamaps Suitable Distribution using a feed-Forward Neural Network
• AQUAMAPS_NEURAL_NETWORK_NS: Aquamaps Suitable Distribution using a Feed-Forward Neural Network provided by Neurosolutions (http://www.neurosolutions.com/)

The above algorithms are automatically managed by an underlying library (Ecological Engine) which takes care of the selection of the most proper computational infrastructure for running the generation algorithm.

Modelers

Currently the following models are supported for training purposes:

• HSPEN: Hspen model by Aquamaps
• AQUAMAPSNN: Feed-Forward Neural Network for usage in Aquamaps generations
• AQUAMAPSNNNS: Feed-Forward Neural Network by Neurosolutions (http://www.neurosolutions.com/) for usage in Aquamaps generations

Even in this case, the above algorithms are automatically managed by the Ecological Engine library which takes care of the selection of a computational infrastructure suited for running the modeling algorithm.

Clustering

No clustering algorithms are currently available.

Transducers

No transducers are currently available.

Evaluators

Available evaluation techniques are the following:

• CLASSIFICATION QUALITY ANALYSIS: This evaluation method applies to a probability distribution and a set of occurrences\absence points. Calculation includes the following values
  • TRUE_POSITIVES
  • FALSE_NEGATIVES
  • TRUE_NEGATIVES
  • FALSE_POSITIVES
  • ACCURACY
  • SENSITIVITY
  • SPECIFICITY

• DISCREPANCY ANALYSIS - BETWEEN TWO SPATIAL DISTRIBUTIONS: Evaluates the distance between two spatial probabilities distributions with the same resolution, in terms of
  • ACCURACY
  • MEAN ERROR
  • VARIANCE
  • NUMBER_OF_ERRORS
  • MAXIMUM_ERROR
  • MAXIMUM_ERROR_POINT
  • NUMBER_OF_COMPARISONS

• HABITAT REPRESENTATIVENESS SCORE: A novel concept for objectively assessing the suitability of survey coverage for modelling the distribution of marine species. Described by Colin D. MacLeod in http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=7911400. Produces the following output:
  • HRS OVERALL SCORE
  • HRS VECTOR

Experiments

Manual Review vs. Automatic Reviews

An experiment was performed using the Statistical Service. We tried to compare some Aquamaps distributions, automatically and manually generated, using a test case species (the basking shark): we selected a species for which we had availability of

• a good number of occurrence points
• a manually reviewed map
• a hspec-suitable map generated by the Aquamaps algorithm

The choice fell on the Basking Shark species (Cetorhinus Maximus, Fis-22747). 449 presence data were available for this species. Figure 1 depicts the presence data distribution. Figure 2 depicts the manually reviewed distribution. Figure 3 depicts the original distribution produced by the Aquamaps-Suitable algorithm.

Figure 1.

Figure 2.

Figure 3.

We tried to perform 2 experiments to test if an automatic machine learning system would have been able to extract species environmental preferences from the same parameters used by the Aquamaps algorithm. The machine learning system was trained with both presence and absence data: absence points were extracted from the reviewed map, from places with probability less than 0.1. We chose a feed forward neural network as machine learning tool, and the parameters we used for the training were the same as in the Aquamaps algorithm: depth mean,depth max,depth min,sst mean,sbt mean,salinity mean,salinity b mean, primary production mean,ice concentration,distance from land,ocean area. The first experiment used 449 absence data all coming from the same region where the reviewed map reported probability values less than 0.1. Figure 4 depicts this absence data distribution.

Figure 4.

Figure 5.

We trained the network with all the presence and absence points. The best performing neural network had 1 inner layer with 100 neurons. The map produced by this system is depicted in figure 5 and presents a big spread in the ocean. The map superposes to the reviewed one, but it is quite far from the Aquamaps-Suitable distribution. The holes left by the neural network correspond mainly to low probability points in the reviewed map. Figure 6 depicts this superposition.

Figure 6.

Figure 7.

The second experiment used absence data randomly chosen among the reviewed map points with low probability. Figure 7 depicts the absence data distribution. We trained again the neural network with all these points. This time the best performing presented 1 inner layer with 300 neurons. Figure 8 depicts the resulting distribution. As it can be noticed by the superposition map in 9, this time the distribution is close to the one from the Aquamaps algorithm instead of being similar to the reviewed map.

We tried to make some comments about this result: if we assume that the neural network is working correctly and it is able to learn something about the fish's attitude from the characteristics of the sea associated to the occurrence and absence points, this could indicate that the manually reviewed map could have been build on partial information about the fish. Furthermore this could mean that the reviewer performed the same considerations of the neural network. On the other side, if we are certain that the reviewed map is correct, then we must admit that the information extracted from the sea is not sufficient to understand fish's preferred habitat. Notice that two automatic systems almost agree on a certain distribution for the fish, which is far from the reviewed one and this could indicate the possibility of an evaluation error in the reviewed map. This case could be helpful for implementing an alert for a biologist who wanted to manually revise a map.

Some final Notes:

• for the basking shark species all the maps are very similar either in the native or in the suitable distribution
• the neural network was trained many times with different topologies, in order to use the best configuration in each experiment
• the neural network does not need expert knowledge to produce the map from the inputs, but absence data are necessary, which come essentially from expert knowledge. This is a little paradox as neural networks are declared in literature among the best performing systems for producing distribution maps. Anyway the inputs are dependent on human knowledge.
• the values reported below refer to a training session using 449 presence and 449 absence data. Experiments were made even using 80% of the set for training and 20% for testing and were repeated using 60% for training and 40% for testing. The above considerations still remained valid.
• numeric comparisons were made in order to calculate the performances of the distributions.

Figure 8.

Figure 9.

Numeric details: Experiments were performed considering as "correct positive classifications" probabilities higher than 80% and as "correct negative classifications" probabilities lower than 0.3

Reviewed Map Performances on Occurrence Points

• TRUE_POSITIVES:332
• FALSE_NEGATIVES:117
• TRUE_NEGATIVES:449
• FALSE_POSITIVES:0
• ACCURACY:0.87
• SENSITIVITY:0.74
• SPECIFICITY:1.0

Aquamaps-Suitable Performances on Occurrence Points

• TRUE_POSITIVES:116
• FALSE_NEGATIVES:333
• TRUE_NEGATIVES:444
• FALSE_POSITIVES:5
• ACCURACY:0.62
• SENSITIVITY:0.26
• SPECIFICITY:0.99

Neural Network with 1 inner layers with 100 neurons - Trained on Dense Absence Data

• TRUE_POSITIVES:431
• FALSE_NEGATIVES:18
• TRUE_NEGATIVES:147
• FALSE_POSITIVES:302
• ACCURACY:0.64
• SENSITIVITY:0.96
• SPECIFICITY:0.33

Neural Network with 1 inner layers with 300 neurons - Trained on Sparse Absence Data

• TRUE_POSITIVES:218
• FALSE_NEGATIVES:231
• TRUE_NEGATIVES:428
• FALSE_POSITIVES:21
• ACCURACY:0.72
• SENSITIVITY:0.49
• SPECIFICITY:0.95

Calculation of the distance between distributions by point-to-point differences with tolerance 0.1

Distance of Aquamaps Suitable from Reviewed Map

• ACCURACY:92.04
• MEAN ERROR:0.46
• VARIANCE:0.053
• NUMBER_OF_ERRORS:8059
• MAXIMUM_ERROR:0.9
• MAXIMUM_ERROR_POINT:7301:101:2
• NUMBER_OF_COMPARISONS:101370

Distance of NN Dense Absence Data from Reviewed Map

• ACCURACY:96.80
• MEAN ERROR:0.57
• VARIANCE:0.084
• NUMBER_OF_ERRORS:3241
• MAXIMUM_ERROR:0.999
• MAXIMUM_ERROR_POINT:7506:206:2
• NUMBER_OF_COMPARISONS:101370

Distance of NN Random Absence Data from Reviewed Map

• ACCURACY:66.75
• MEAN ERROR:0.57
• VARIANCE:0.069
• NUMBER_OF_ERRORS:47138
• MAXIMUM_ERROR:0.999
• MAXIMUM_ERROR_POINT:1116:228:2
• NUMBER_OF_COMPARISONS:141762

Distance of NN Dense Absence Data from Aquamaps Suitable

• ACCURACY:82.41
• MEAN ERROR:0.51
• VARIANCE:0.063
• NUMBER_OF_ERRORS:2309
• MAXIMUM_ERROR:0.9
• MAXIMUM_ERROR_POINT:1414:362:3
• NUMBER_OF_COMPARISONS:13127

Distance of NN Random Absence Data from Aquamaps Suitable

• ACCURACY:93.71
• MEAN ERROR:0.56
• VARIANCE:0.055
• NUMBER_OF_ERRORS:8921
• MAXIMUM_ERROR:0.9
• MAXIMUM_ERROR_POINT:7516:485:2
• NUMBER_OF_COMPARISONS:141762

Habitat Representativeness

We applied the HRS technique to the previous experiment. The output of the procedure is currently a score for each input feature (ice concentration, salinity etc.), ranging from 0 to 2, where 0 means that the occurrence\absence points for the species under analysis is rich enough to represent the variability in the projection area (usually all the oceans).

An overall score was even calculated by summing the HRS scores of the features. This is the only little difference respect to the Mac Leod paper, where he suggests to weight each score for the inverse of the eigenvalues for the PCA transformation. We ignored the inverse weighting for two reasons: (i) the eigenvalues highly depend on the ordering of the vectors taken for calculating the PCA (while the HRS don't), (ii) as all the PCA components are used by the HRS algorithm we should consider all the features as independent or at least equally important; all the HRS scores are commensurable and then an inverse weight would always give too much importance to less variable dimensions. Our overall HRS score is almost independent from the ordering of the input vectors and rapidly achieves an asymptotic number when an increasing random number of samples is taken from the projection area.

In the following results, we report an analysis we performed on the basking shark case, discussed in the last experiment. The HRS calculated when using presence data + randomly taken absence data is slightly lower than the one coming from presence data + dense absence data (you can remind the data maps from the attached images). This confirms that the widespread data were more suited to perform the training and the projection, and then the map produced by the Aquamaps Algorithm or the Neural Network trained on such data should be formally more accurate than the one produced by the Neural Network trained on the presence and dense absence data. Furthermore, as the last model was similar to the manually produced map, this could be an alert even to the biologist, maybe having generated the map by using a local information about the species. On the other side, the difference between the HRS scores is not high, the first calculation gets 3.89 while the second gets 4.49. The 0.6 difference could indicate that the scores are not significantly different, and are both too high to be representative for the environment. According to this consideration, both the training datasets would be not useful and we should discard the reliability of all the automatically generated maps. In this case we should rely only on the biologist knowledge about the species.

Overall Habitat Representativeness Score: