EARTHQUAKE PREDICTOR NEURAL NETWORK

(PROJECT ONAMAZU)


| Predictions | Model | Archive |

Neural Network Model Information

Dataset:

Source: United States Geological Survey (USGS)
Training Set: 1973-01-01 to 2015-12-31
Test Set: 2016-01-01 to 2016-04-30
Seed Set: Last 30 Days
Temporal Resolution: 1 Day (24 Hours)
Spatial Resolution: 2 Degrees Latitude and 2 Degrees Longitude
Details: Earthquakes were reorganized into tensors, with each slice representing a single day and consisting of a matrix with the rows and columns representing approximate latitudes and longitudes. Each position in the matrix was filled with the highest magnitude earthquake occurring at that approximate location and time.
Notes: Due to the limited temporal and spatial resolution of the dataset, not all earthquakes from the original dataset were included in the final datasets.

Architecture:

Long Short Term Memory (LSTM) Recurrent Neural Network (RNN)
Hidden Layers: 3
Neurons Per Hidden Layer: 1024 nodes
Timesteps: 30 days
Epochs of Training: 66/100 (Trained to 100, but best performance achieved at 66)
Loss Function: Mean Squared Error
Library: Torch7

Predictive Accuracy (PA = TP / (TP + FP + FN)) = 0.3584421465771 (35.84%)

Predictive Accuracy and other statistical calculations are based on the following criteria:

Statistics:

True Positive (TP) = 4749
True Negative (TN) = 1930751
False Positive (FP) = 2133
False Negative (FN) = 6367
Total Population (TP + TN + FP + FN) = 1944000
Positive (TP + FN) = 11116
Negative (FP + TN) = 1932884
Predict Positive (TP) = 6882
Predict Negative (TP) = 1937118
Prevalence (PRE = positive / total_pop) = 0.0057181069958848 (0.57%)
True Positive Rate OR Sensitivity OR Recall (TPR = TP / positive) = 0.42722202231018 (42.72%)
False Negative Rate OR Miss Rate (FNR = FN / positive) = 0.57277797768982 (57.28%)
False Positive Rate OR Fall-Out (FPR = FP / negative) = 0.0011035323382055 (0.11%)
True Negative Rate OR Specificity (TNR = TN / negative) = 0.99889646766179 (99.89%)
Accuracy (ACC = (TP + TN) / total_pop) = 0.99562757201646 (99.56%)
Positive Predictive Value OR Precision (PPV = TP / predict_positive) = 0.69006102877071 (69.01%)
False Discovery Rate (FDR = FP / predict_positive) = 0.30993897122929 (30.99%)
False Omission Rate (FOR = FN / predict_negative) = 0.0032868415863153 (0.33%)
Negative Predictive Value (NPV = TN / predict_negative) = 0.99671315841368 (99.67%)
Positive Likelihood Ratio (PLR = TPR / FPR) = 387.14046477778
Negative Likelihood Ratio (NLR = FNR / TNR) = 0.57341075500103
Diagnostic Odds Ratio (DOR = PLR / NLR) = 675.15382542324