Keywords
machine learning, artificial intelligence, radar ornithology, nexrad, bird migration
Start Date
1-7-2008 12:00 AM
Abstract
Radar ornithology using data from the NEXRAD weather radar system has given scientists new tools for studying bird migration in the United States. Unfortunately, the process of identifying echoes from birds in radar data still largely requires that trained technicians spend hours manually scouring radar scans. This paper provides some background for understanding biological echoes in NEXRAD data and then describes our initial investigations of the use of machine learning techniques to help automate the process of echo classification. Doppler data allows researchers to look at bird migration by examining large clusters of birds that could be observed approaching and descending at stopover points where they would rest until they began the next leg of their journey. One of the NEXRAD system’s greatest strengths has also been a significant obstacle for researchers studying bird migration: with 154 radar stations across the United States, each often producing hundreds of volume scans per day, the amount of data to sort through is staggering. The real problem is that classifying birds in radar scans currently requires a skilled technician who has been trained in visually identifying the tell-tale signs that distinguish biological echoes from non-biological echoes. Consequently, the task of plotting a specific migration over any significant amount of space and time quickly becomes a difficult and resource intensive problem. We have begun by using a K-nearest neighbour classifier, a naïve Bayes classifier, and a neural network to classify the echoes. Early validation results using tenfold cross-validation procedures are hopeful and indicate that machine learning techniques could be well suited for this task. Accuracy rates have exceeded 98 percent. Although these early results are encouraging, it is important to keep in mind that each of the training sweeps used in these experiments was selected by an expert because it could be considered a prototypical example of one particular echo type dominating a sweep. Intuitively, these sweeps are the easiest to classify, which may explain the results. The real test will be to apply these methods to more complex data, including ambiguous data as well as mixed sweeps containing both types of echoes. Our next efforts will concentrate on acquiring and experimenting with such data. Our eventual goal is to use machine learning methods to map bird migration pathways.
Identifying Biological Echoes in Radar Scans Using Machine Learning
Radar ornithology using data from the NEXRAD weather radar system has given scientists new tools for studying bird migration in the United States. Unfortunately, the process of identifying echoes from birds in radar data still largely requires that trained technicians spend hours manually scouring radar scans. This paper provides some background for understanding biological echoes in NEXRAD data and then describes our initial investigations of the use of machine learning techniques to help automate the process of echo classification. Doppler data allows researchers to look at bird migration by examining large clusters of birds that could be observed approaching and descending at stopover points where they would rest until they began the next leg of their journey. One of the NEXRAD system’s greatest strengths has also been a significant obstacle for researchers studying bird migration: with 154 radar stations across the United States, each often producing hundreds of volume scans per day, the amount of data to sort through is staggering. The real problem is that classifying birds in radar scans currently requires a skilled technician who has been trained in visually identifying the tell-tale signs that distinguish biological echoes from non-biological echoes. Consequently, the task of plotting a specific migration over any significant amount of space and time quickly becomes a difficult and resource intensive problem. We have begun by using a K-nearest neighbour classifier, a naïve Bayes classifier, and a neural network to classify the echoes. Early validation results using tenfold cross-validation procedures are hopeful and indicate that machine learning techniques could be well suited for this task. Accuracy rates have exceeded 98 percent. Although these early results are encouraging, it is important to keep in mind that each of the training sweeps used in these experiments was selected by an expert because it could be considered a prototypical example of one particular echo type dominating a sweep. Intuitively, these sweeps are the easiest to classify, which may explain the results. The real test will be to apply these methods to more complex data, including ambiguous data as well as mixed sweeps containing both types of echoes. Our next efforts will concentrate on acquiring and experimenting with such data. Our eventual goal is to use machine learning methods to map bird migration pathways.
