Post on 15-Jan-2015
description
Identification of nesting phase in tortoise populations by neural networks
AISB: 2014 ISAWEL Roberto Barbuti, Stefano Chessa, Alessio Micheli, Rita Pucci
Department of computer science University of Pisa
A project against the tortoises extinction
04 April 2014
Extinction problem
Environment pollution
Habitat loss
New predators
Giant Tortoise endangered
Human influences on the tortoises’ habitat
Protection of hatchlings
To simplify this process it is necessary to use an automatic system able to recognize the animal behavior.
Tortoise@ project: monitoring of tortoises
Tortoise@ project
Device with sensor board
Some sensors available on sensor board: • Light sensor (LDR);
• Temperature sensor (Thermistor);
• Accelerometer sensor.
Communication with base station:
• Radio.
Limits due to hardware: • 8MHz microcontroller; • Equipped with a 8 Kbyte
RAM memory; • 256 Kbyte of flash memory; • Energy capability: two
alkaline batteries.
Tortoise@ project
Phases of procedure
1. Environment monitoring (EM).
2. Movement monitoring (MM).
3. Extended movement monitoring (EMM).
4. Data communication (DC).
We focused on this phase
Tortoise@ project
Movement monitoring (MM): main point of our research
• Data collecting on field • Preprocessing of signals
• Filtering; • Normalization; • Down sampling.
• Recognition algorithms • Identification of a characteristic pattern; • Correlation analysis; • Neural Networks:
• Training; • Validation; • Test.
Protection center for Mediterranean tortoises
Tortoise@ project: Movement monitoring (MM)
Data collecting on field
Eating phase
Nesting phase
Walking phase
Base station device
Tortoise@ project: Movement monitoring (MM)
Axis of accelerometer sensor
• The X axis indicates the movements of the carapace of the tortoise on the short side of carapace;
• The Y axis indicates the inclination of the
carapace on the long side of carapace;
Tortoise@ project: Movement monitoring (MM)
Data collecting on field: samples of recorded signals
Accelerometer signal of eating phase
Accelerometer signal of walking phase
Accelerometer signal of digging phase
Tortoise@ project: Movement monitoring (MM)
Identification of a characteristic pattern
Digging pattern with positive classification (1)
Walking and Eating patterns with negative classification (-1)
Characteristic pattern
Tortoise@ project: Movement monitoring (MM)
Neural Networks: Input Delay Neural Network (1)
Hidden layer
Output layer
Tortoise@ project: Movement monitoring (MM)
Input layer Signal
Data of last window Data left by shifting Data of shifting of window
Neural Networks: Input Delay Neural Network (1)
Hidden layer
Output layer
Tortoise@ project: Movement monitoring (MM)
Input layer Signal
Data left by shifting Data of shifting of window Data of last window
Input Delay Neural Network (2)
Tortoise@ project: Movement monitoring (MM)
Hidden layer
Output layer
Input layer
Data of last window
Data of shifting of window
Accelerometer signal Data left by shifting
IDNN Convolutional Neural Network
• With weight sharing
Output layer
Hidden layer
Input layer
Tortoise@ project: Movement monitoring (MM)
Outputs with digging signal and walking signal
Output of neural network with a digging signal input
Output of neural network with a walking signal input
Tortoise@ project: Movement monitoring (MM)
Accuracy of Neural Networks classification
Structure Error Training (%)
134 patterns
Error Validation Pattern set(%) 20
patterns
Error Validation Set(%)
18 signals
Error Test Set(%) 25 signals
IDNN 10% 12% 0% 4%
IDNN + CNN 10% 5% 0% 4%
IDNN + CNN weight sharing
26% 27% 0% 12%
Tortoise@ project: Movement monitoring (MM)
Memory occupation
Structure Input with float weights Input with int weights
IDNN 2160byte 1980 byte
IDNN + CNN 712byte 536byte
IDNN + CNN weight sharing 536byte 356 byte
Tortoise@ project: Movement monitoring (MM)
Conclusions
• Data collecting on field;
• Analysis of activity signals: • Identification of characteristic pattern.
• Neural network to classify signals: • IDNN;
• IDNN + CNN;
• IDNN + CNN with weight sharing.
The use of IDNN + CNN is a good trade-off between memory occupation and performance obtained with unknown signals.
The system implemented recognizes the digging phase. It is an initial stage to automate the rescue of eggs in order to increase the population growth of tortoises.
Tortoise@ project: Movement monitoring (MM)
Thank you pucci@di.unipi.it
Department of computer science University of Pisa