Figures index

From

Advanced Spatio-Temporal Event Detection System for Groundwater Quality Based on Deep Learning

Divas Karimanzira, Linda Ritzau, Tobias Martin, Thilo Fischer

Applied Ecology and Environmental Sciences. 2023, 11(3), 79-90 doi:10.12691/aees-11-3-2
  • Figure 1. Categorization of anomaly detection algorithms
  • Figure 2. Illustration of the three-dimensional multivariate spatio-temporal data matrix structure
  • a)
  • b)
  • Figure 4. Framework 1: LSTM autoencoder and forecaster as basis for training the encoder which is then used for spatial and temporal anomaly detection by a deep neural network (DNN) classifier
  • Figure 5. An Auto encoder made up of a CNN (3DCNN or Multichannel 2DCNN) decoder and a LSTM or ConvLSTM decoder and forecaster to decode spatial and temporal features for anomaly detection
  • Figure 6. Model test results of the LSTM Composite network und the 3DCNN Auto encoder for the observation wells, identified by their IDs are given. The measured values for the different sensors (features) are plotted in blue
  • Figure 7. Anomaly detection results. Subsequences marked in blue and in the shaded areas indicate anomalous events a) LSTM Composite network und b) 3DCNN Auto encoder. Yellow marked areas are temporary anomalies and green marked represent spatial anomalies of the LSTM composite network
  • Figure 8. Spatial and context anomalies for neighboring wells a) water temperature and b) Potassium, c) point anomalies in nitrate concentration and d) level change in nitrate concetration. In b and c, the red line shows the beginning of the anomalies
  • Figure 9. Anomaly detection results. a) Feature contribution to the anomalies in region 791013210 (i.e. anomaly causes) and c) Feature contribution to the anomalies for two neighboring regions 7910062 and 7910061
  • Figure 10. Model test results of the LSTM Composite network und the 3DCNN Auto encoder for the observation wells, identified by their IDs are given. The measured values for the different sensors (features) are plotted in blue
  • Figure 11. Anomaly detection results. Subsequences marked in blue and in the shaded areas indicate anomalous events. a) LSTM Composite network und b) 3DCNN Auto encoder
  • Figure 12. Anomaly detection results of the modified ST-DBSCAN. a) Spatio-temporal and feature based clusters obtained by running the modified ST-DBSCAN. b) Subsequences marked in blue and in the shaded areas indicate anomalous events
  • Figure 13. a) ε plot b) a k-distance graph