تشخيص هوشمند تاسل در تصاوير پهپادي با استفاده از يادگيري عميق براي تعيين تاريخ گلدهي

Automatic Tassel Detection to Estimate Flowering Date in the UAV Images using Deep Learning Techniques

ﺗﺨﻤﯿﻦ ﻋﻤﻠﮑﺮد و ﺑﺮرﺳﯽ روﻧﺪ رﺷﺪ در ﮔﻮﻧﻪﻫﺎي ﻣﺨﺘﻠﻒ از ﯾﮏ ﻣﺤﺼﻮل در ﮐﺸﺎورزي دﻗﯿﻖ ﺑﺮاي ﻣﺤﻘﻘﯿﻦ و ﮐﺎرﺷﻨﺎﺳﺎن ﺣﻮزه ﮐﺸﺎورزي ﺑﺴﯿﺎر ﺣﺎﺋﺰ اﻫﻤﯿﺖ اﺳﺖ. در اﯾﻦ ﻣﻘﺎﻟﻪ روﺷﯽ ﻧﻮﯾﻦ ﻣﺒﺘﻨﯽ ﺑﺮ ﯾﺎدﮔﯿﺮي ﻋﻤﯿﻖ ﺗﮏ ﻣﺮﺣﻠﻪاي ﺑﻪ ﻧﺎم GP-YOLOv5 ﺑﺮاي ﺷﻨﺎﺳﺎﯾﯽ ﺧﻮدﮐﺎر ﺗﺎﺳﻞ در ﺗﺼﺎوﯾﺮ ﭘﻬﭙﺎدي از ﯾﮏ ﻣﺰرﻋﻪ ﺑﺰرگ ذرت در ﺗﺎرﯾﺦﻫﺎي ﻣﺨﺘﻠﻒ رﺷﺪ و ﺗﺨﻤﯿﻦ زﻣﺎن ﮔﻞدﻫﯽ اراﺋﻪ ﺷﺪه اﺳﺖ. در اﯾﻦ راﺳﺘﺎ اﺑﺘﺪا ﺑﻪ دﻟﯿﻞ رﺷﺪ ﺗﻌﺪاد ﮐﻤﯽ از ﺗﺎﺳﻞﻫﺎ در ﻣﺮاﺣﻞ اوﻟﯿﻪ رﺷﺪ ﺑﺮاي داده اﻓﺰاﯾﯽ از ﺷﺒﮑﻪ ﻣﻮﻟﺪ ﻣﺘﺨﺎﺻﻢ GP-GAN اﺳﺘﻔﺎده ﺷﺪ. ﺳﭙﺲ ﺑﺮاي ﺷﻤﺎرش و ﺗﺸﺨﯿﺺ ﺗﺎﺳﻞﻫﺎ ﺳﺎﺧﺘﺎر و ﭘﺎراﻣﺘﺮﻫﺎي آﺷﮑﺎرﺳﺎز YOLOv5 ﺑﺮاي اﻓﺰاﯾﺶ دﻗﺖ ﻣﻄﺎﺑﻖ ﺑﺎ ﭘﺎﯾﮕﺎه داده اﺻﻼح ﺷﺪ. در اداﻣﻪ ﺷﻤﺎرش ﮔﯿﺎﻫﺎن در ﻣﺮاﺣﻞ اوﻟﯿﻪ ﮐﺎﺷﺖ ﺑﻪ ﻋﻨﻮان ﯾﮏ ﭘﺎراﻣﺘﺮ ﻣﻬﻢ در ﺗﻌﯿﯿﻦ ﺗﺎرﯾﺦ ﮔﻞدﻫﯽ در ﻧﻈﺮ ﮔﺮﻓﺘﻪ ﺷﺪ. ﺷﻤﺎرش ﮔﯿﺎﻫﺎن ﺑﺎ اﺳﺘﻔﺎده از آﺷﮑﺎرﺳﺎز CenterNet اﻧﺠﺎم ﺷﺪه اﺳﺖ و از اﻟﮕﻮرﯾﺘﻢﻫﺎي درونﯾﺎﺑﯽ و ﭘﯿﺶ ﺑﯿﻨﯽ ﺑﺮاي ﺗﻌﯿﯿﻦ ﺗﺎرﯾﺦ ﮔﻞدﻫﯽ اﺳﺘﻔﺎده ﺷﺪ. روش ﭘﯿﺸﻨﻬﺎدي ﺑﺎ دو روش ﻣﻌﺘﺒﺮ ﻣﺒﺘﻨﯽ ﺑﺮ ﺗﺸﺨﯿﺺ CenterNet و روش ﻣﺒﺘﻨﯽ ﺑﺮ رﮔﺮﺳﯿﻮن TasselNetv2+ ﺑﺮاي ﺷﻤﺎرش ﺗﺎﺳﻞﻫﺎ ﻣﻘﺎﯾﺴﻪ ﺷﺪ. دﻗﺖ ﻣﯿﺎﻧﮕﯿﻦ در ﺗﺸﺨﯿﺺ ﺻﺤﯿﺢ ﺗﺎﺳﻞﻫﺎ در روش ﭘﯿﺸﻨﻬﺎدي 96/81 و در روش CenterNet، 81/78 درﺻﺪ اﺳﺖ ﮐﻪ ﻧﺸﺎن ﻣﯽدﻫﺪ دﻗﺖ روش ﭘﯿﺸﻨﻬﺎدي ﺑﺎﻻﺗﺮ از روش CenterNet اﺳﺖ.

Estimating crop yields and examining growth trends in different species of a crop in precision agriculture is very important for researchers and agricultural experts. In this article, a new technique based on one-stage objection detection called GP-YOLOv5 for automatic tassel detection in the UAV images of a large maize field at different growing stages and flowering date estimation is presented. Because of the existing small number of tassels in the early stages of growth, GP-GAN is used to augment the training data. After that, the hyperparameters of the YOLOv5 are optimized to increase the tassel detection accuracy. Plant counting using CenterNet in the early stage of growth is calculated to determine the flowering date. Finally, well-known interpolation and prediction algorithms are used to estimate the flowering date. The proposed method is compared with two state-of-the-art methods based on detection “CenterNet” and regression “TasselNetv2+” technique for tassel counting. The average accuracy of GP-YOLOv5 for tassel detection is around 96.81 % and for the CenterNet method, it is around 81.78 %, which indicates that the accuracy of the proposed method is higher than the CenterNet technique.