Dataset: InsectSound1000-GreenhouseNoiseRecordings
Published by OpenAgrar
About this dataset
The InsectSound1000_GreenhouseNoiseRecordings dataset is a dataset of simultaneous background noise recordings recorded in a greenhouse with the same two microphone arrays and recording settings as used for the insect sound recordings published in the InsectSound1000 dataset (in the “Measurement_Microphone_Array” folder) and the InsectSound1000_MEMS dataset (in the “Respeaker” folder). The original InsectSound1000 dataset was recorded with a four-channel low-noise measurement microphone array and the MEMS-version of the dataset was recorded with a six-channel MEMS microphone array called ReSpeaker Core V2.0. The background noise sounds were recorded in a greenhouse located at the facilities of the Julius Kühn-Institut in Braunschweig, Germany, which is used for research purposes. It had a length of 18.55 m, a width of 9 m and a ridge height of 5 m and was constructed of metal with glass windows. As the greenhouse was made for cultivation in soil, the floor was covered with soil, except for four paths of rubber mats. Further, the structure contained metal constructions for fixing high-growing vegetables and an irrigation system. About a quarter of the area was planted with young pepper plants. Windows were opened and closed automatically by a climate control system. The goal of these environmental sound recordings was to, if possible, not contain insect sounds. Therefore, the recordings were made at the time when the first plants were planted. Temperatures were still relatively low, and so was the number of insects inside the greenhouse. To build a diverse dataset, six different locations inside the greenhouse were recorded, and at most recording locations, two different orientations of the microphones were recorded, to vary the direction in which the constant noise sources would reach the microphones. The microphone arrays were mounted about one meter above the ground using a tripod. The sound events that occurred during the recordings included: a constant low-level humming sound produced by the climate chambers nearby, doors opening and closing, greenhouse windows opening and closing via motors, sounds of construction work (power tools, hammers), air and car traffic noise, bird calls and distant unintelligible speech. The recordings were split into files of 2500 ms, resulting in 1495 files for each microphone array. These two datasets were then randomly split parallel into training (60 %), validation (20 %) and test data (20%) in order to prepare them for use in deep learning model development and evaluation. The parallel split means that the files contained in, e.g. the Measurement_Microphone_Array test folder are recordings of the same time and place as those found in the files of the Respeaker test folder. Each sample is a multichannel wave-file of 2500 ms length, at 16 kHz sample rate and 32-bit resolution. This background noise dataset was used in a publication comparing the influence of microphone quality on the feasibility of acoustic insect detection, comparing recordings from the InsectSound1000_MEMS dataset with the original high-quality InsectSound1000 dataset and the influence of background noise on the behaviour of insect sound recognition models.