Multiscale Maize Tassel Identification Based on Improved RetinaNet Model and UAV Images

The acquisition of maize tassel phenotype information plays a vital role in studying growth and improving yield. Unfortunately, detecting tassels has proven challenging because the complex field environment, including image resolution, varying sunlight conditions, plant varieties, planting density. To address this situation, present study uses unmanned aerial vehicle (UAV) remote sensing technology deep learning algorithm to facilitate identification counting. UAVs are used collect images experimental fields, RetinaNet serves as basic model for tassels. Small accurately identified by optimizing feature pyramid structure introducing attention mechanisms. We also how mapping differences brightness, variety, density affect model. results show that improved is significantly better at than original average precision 0.9717, 0.9802, recall rate 0.9036. Compared with model, improves precision, 1.84%, 1.57%, 4.6%, respectively. mainstream target detection models such Faster R-CNN, YOLOX, SSD, more detects smaller For equal-area differing becomes progressively worse resolution decreases. analyze depends on brightness various models. With increasing worsens, especially small This paper analyzes five varieties. Zhengdan958 prove easiest detect, R2 = 0.9708, 0.9759, 0.9545 5, 9, 20 August 2021, Finally, we use detect under different densities. At 29,985, 44,978, 67,466, 89,955 plants/hm2, mean absolute errors 0.18, 0.26, 0.48, 0.63, Thus, error increases gradually thus provides new method high-precision farmland useful can be high-throughput investigations phenotypic traits.