Overview / Usage

To detect changes in migrating bird populations that are usually gradual, regular counts of the flocks should be carried out. This is vital for giving more precise management decisions and taking preventive actions when necessary. Traditional counting methods are widely used. However, these methods can be expensive, time-consuming, and highly dependent on the mental and physical status of the observer and environmental factors. Taking these uncertainties into account, we aimed at taking the advantage of the advances in the artificial intelligence (AI) field for a more standardized counting action. The study has been practically initiated 10 years ago by beginning to take photos on a yearly basis in predefined regions of Turkey. After a large collection of bird photos had been gathered, we predicted the bird counts in photo locations from images by making strong use of AI. Finally, we used these counts to produce several bird distribution maps for further analysis. Our results showed the potential of learning computers in support of real-world bird monitoring applications.

Methodology / Approach

The main process are :

• Data collection

• Automated Bird Detection

• Data augmentation

• Evaluation Criteria

• Birds counting

Technologies Used

** Data Collection** : Photos have been captured at different resolutions and from different viewpoints. In case it was impossible to photograph all birds in an area with a single shot, multiple photos were taken to span the entire flock of birds while keeping as little overlap as possible across those photos. There exist different backgrounds depending on observation environments, such as the vicinity of water surfaces, beaches, forests and farmlands, hosting various bird habitats, bird activities (e.g., flying birds lead to flat sky backgrounds) and weather conditions. Bird poses may also differ according to lateral or frontal directions and actions such as flying, standing straight and bending.

** Automated Bird Detection:** Deep neural networks can be made of several different types of layers, each of which is made up of a fixed number of artificial neurons that jointly act as input to the next layer neurons. Typically, types of hidden layers (i.e., layers except the input and output layers) used for image analysis include

• Convolutional layers each of which convolves the previous layer’s neuron outputs by different convolution filters (i.e., each filter is a linear combination of neighboring neurons inside a fixed-size window) and then apply a non-linear function.

• Max-pooling layers each of which outputs the maximum of input neuron values in each grid when the input neurons are partitioned into non-overlapping rectangular grids.

• Fully connected (FC) layers where each neuron’s output is computed as a non�linear function applied to a linear combination of outputs of all neurons in the previous layer .

Data augmentation : Despite its success in image recognition, deep learning models need lots of images to be trained in many problems, especially when the network size and data variability are large . Limited training data may cause the over-fitting problem in a neural network especially when the network has to generalize across a diverse set of samples such as birds. In that case, significantly more data is required to learn a reliable set of network parameters. To overcome this problem, the feature extraction network can be pre-trained using another available big data set, with possibly different object categories, so that the training can continue from those learned parameters rather than random values using our smaller data set of bird images. This technique for compensating the negative effect of limited training data is called transfer learning . Other solutions that we applied to help improve the performance were to augment the image data set by flipping the images horizontally and drop-out regularization which refers to deactivating randomly chosen neurons during training .

** Evaluation Criteria : **Quantitative evaluation was performed in the test set by comparing the binary detection maps, obtained by applying a uniformly sampled range of thresholds on detection scores, to the validation bounding boxes. Each detection for a particular threshold is considered to be positive if its intersection-over-union (IoU) ratio with a ground-truth annotation is greater than 0.5. By setting each threshold for detection scores, a set of true positives and false positives can be generated to calculate precision and recall that have been commonly used in the literature to measure how well the detected objects correspond to the ground truth objects. Precision is computed as the ratio of the number of correctly detected birds to the number of all detected birds while recall is computed as the ratio of the number of correctly detected birds to the number of all birds in the validation data. We also found the F-measure which combines precision and recall as their harmonic mean.

Bird Counting : In this section, we present a bird counting test in order to see the effectiveness and usability of automated counting over the traditional manual method. We used 45 of our collection of photographs for the experiments. The photographs were selected so that the sample would be as much representative as possible in terms of different bird species, flock size, and environmental conditions of the photo (e.g., presence of fog and obstacles such as trees). We compared the bird counting performances of the following three methods:

• Manual: Two experts, Exp1 and Exp2, that have PhDs in ornithology and at least five-year field experiences on bird counting, separately counted the photographs in random order on a computer monitor. In this method, experts usually count birds by grouping them (i.e., not one by one, but five by five or ten by ten). The maximum allowed duration of effort for each photo was fixed at 3 min.

• Automated: This method corresponds to the Faster R-CNN model output.

• Computer-assisted: We also evaluated the performances of the two experts’ manual countings with the aid of the model outputs. Each expert was shown on the screen the automated detection result image of each photo and asked to improve the overall count by adding and subtracting uncounted and over-counted birds, respectively. Counting processes were held in the same conditions with the manual method two weeks after the manual count. Three different evaluation metrics were recorded for each method and photo: bird count, duration of counting effort and count error. The count error for each photo was calculated as the ratio of the absolute difference of the correct number of birds and the resulting count to the correct number. For instance, if an expert and/or the model counted 80 birds out of 100 correct number of birds, then the error rate would be 20% or 0.20. On the other hand, if 125 birds were counted, the success rate would be 0.25. Q-Q plots were used to check whether the observations of each metric were normally distributed or not. Since the bird count metric was not normally distributed, correlations between the correct number of birds and manual, automated and computer-assisted counts were examined with non-parametric Spearman’s rank test.

Documents and Presentations