Research On Bat Species Recognition System Based On Neural Networks

Bat is one of the most widely distributed and most successfully evolutionary mammals groups in the world. Population quantity of bat decreases by years, because of the habitat fragmentation and artificial indiscriminate killing etc. Therefore, the bat population's detection and the species recognition have the important meaning effectively to clear the predation ecological environment and protect the bats.The key to identify the species lies in the feature extraction and the sorter's design. The wavelet analysis decomposes the different frequency components of signal to the non-overlapped frequency bands. It is used for the signal filtering, the signal-noise separation and the feature extraction and is applied to analyze the non-steady signal. The energy values of different frequency channels of sound signal are extracted as feature vector by wavelet packets decomposition.The more the characteristic parameters are, the more complicated the network architecture is. This article uses the principal components analysis method to optimize the feature space. There are two ways to reduce the feature space dimension by the principal components analysis. One way is to delete the redundant variable by PCA. Another way is to transform coordinate of feature space. According to the principal components score, we choose less parameters as new feature space. The paper applies two ways to optimize the network's input space, attaining the ideal result.The artificial neural network has the merits which some traditional technologies don't have, such as good fault-tolerant ability, strong classification ability, parallel processing ability and self-learning capability. Therefore, the neural network is widely applied in the pattern recognition direction. At present some networks just like the BP network, the Hopfield network, the Ko-honen network and so on are commonly used. This article applies the forward feedback network to realize bat's recognition.Finally the recognition results of the test sample are given, and recognition accuracy for every species can simultaneously achieve above 80%. The result shows that this method is feasible to recognize the different species of Rhinolophus in flight.The aim of this research is to realize the bat population's online examination and statistic. Once the detection system based on this theory is presented, it may have great application value in the bat or other community animal's research.