Vegetation phenology detection of deciduous broad-leaf forest using YOLOv3 from PhenoCam

Abstract

Vegetation phenology identification is significance to the exploration of vegetation growth and is also conducive to the impact of phenology on the ecological environment. Recently, vegetation phenology detection is based on a time series of vegetation phenology to index simulation of vegetation growth time indirectly. In this study, we identify the vegetation phenology of deciduous broad-leaved forest through the deep learning method within a single PhenoCam image. The result of the phenology identification of growing regions, the accuracy MAP of daily identification in daily scales mAP up to 10.2%, which could identify the growing period of most deciduous broad-leaved forests. The identification accuracy mAP in the 8-day scale is up to 69%, and the identification mAP accuracy of vegetation could reach 98.2% when it was divided into four categories. The purpose of this study is to detect the phenological growth period of deciduous broad- leaved forest with rapid development, high precision, and fast deep learning methods. It has a great improvement on the current method of calculating the vegetation phenology period by using the traditional measurement and related mathematical and physical models. While obtaining the phenology period more quickly, it can automatically and accurately obtain the growth area and growth period of the study area, making a certain contribution to the study of vegetation phenology.