Machine learning subfossil midges

This datatset is a basis for a pipeline for an automated identification of the non-biting midges´ subfossil remains using convolutional neural networks (CNNs). It contains relevant image and mode data. Relevant R code is at https://github.com/chironomus/CNN-Chironomidae-subfossil-identification We explored Convolutional Neural Networks (CNN) to automatically identify and measure head capsule length of the most abundant chironomid morphotypes in a number of spatially distributed surface-sediment samples as well as downcore samples from late Pleistocene and Holocene sediment records from Europe. We applied machine learning to identify larval subfossils with varying degrees of taphonomic preservation and no standardized positioning of specimens on the microscope slides, a situation which is typical for slides that were produced for routine identification of chironomid remains from lake sediment records. Additionally, we have developed a pipeline for automatic measurements of microfossils, as size is an important functional trait in Chironomidae that is potentially useful for inferring past changes in the relationship of these organisms with their environment. This preliminary study illustrates the potential of automated, rapid identification and measurements of subfossil Chironomidae, with CNN, although further data are required to increase the performance of our approach. Structure of the dataset model_onnx - contains onnx model, that can be loaded into ParticleTrieur and used for automatic identintification of the images model_tf2 - contains tf2 model, that can be loaded into ParticleTrieur and used for automatic identintification of the images also, contains diagnostic and output files for the ResNet50 models /legend.csv /loss_vs_epoch.pdf /training_parameters.json /tsne.pdf /model_onnx /model_tf2 /accuracy_vs_epoch.pdf /confusion_matrix.pdf /health_summary.txt zip archive "Test" contains final training dataset (images) Also contains image outputs of the scanning of the slides, with images for the following samples (see table below)