Efficient, automated and robust pollen analysis using deep learning

Pollen analysis is an important tool in many fields, including pollination ecology, paleoclimatology, paleoecology, honey quality control, and even medicine forensics. However, labour-intensive manual pollen often constrains the number of samples processed or analysed per sample. Thus, there a desire to develop reliable, high-throughput, automated systems. We present method for analysis, based on deep learning convolutional neural networks (CNN). scanned microscope slides with fuchsine stained, fresh automatically extracted images all individual grains. CNN models were trained reference (122,000 grains, from 347 flowers 83 species 17 families). The used classify different grains series experiments. also propose adjustment reduce overestimation sample diversity cases where are likely contain few species. Accuracy model was 0.98 when each first pooled, then split into training validation set (splitting experiment). accuracy much lower (0.41) kept separate, one such remaining (leave-one-out therefore combined 28 types new leave-one-out experiment revealed overall 0.68, recall rates >0.90 most types. When validating against 63,650 manually identified 370 bumblebee samples, we obtained 0.79, but our procedure increased this 0.85. Validation through splitting experiments may overestimate robustness contexts (samples). Nevertheless, has potential allow large quantities real data be reasonable accuracy. Although compiling libraries time-consuming, simplified by method, can lead widely accessible shareable resources analysis.