[U-04] Geospatial Applications for Societal Benefits

[U04-03] Early-Stage Weed Detection In Paddy Field Using Hyperspectral UAV and Machine Learning

*Nik Norasma Che Ya^1,3,4、Muhammad Wafi Alham Mohd Azlee¹、Muhamad Noor Hazwan Abd Manaf²、Abdul Syukor Juraimi²、 Mohd Razi Ismail² (1.Department of Agriculture Technology, Faculty of Agriculture, Universiti Putra Malaysia, 43400 UPM Serdang, Selangor, Malaysia.、2. Department of Crop Science, Faculty of Agriculture, University Putra Malaysia, Serdang 43400, Malaysia、3. Institute of Plantation Studies, Universiti Putra Malaysia (UPM), Serdang 43400, Malaysia、4. Smart Farming Technology Research Centre, Universiti Putra Malaysia (UPM), Serdang 43400, Selangor, Malaysia)

キーワード：rice mapping, unmanned aerial vehicle, weed detection, geographic information system

Weeds, known as undesirable plants, can emerge within agricultural fields and compete with crops for essential resources such as soil nutrients, water, space, canopy, and light. The weedrelated losses in rice production systems must be addressed, as they impede various field activities throughout the crop growth cycle. The objective of this study is to discriminate weeds in the paddy field by using hyperspectral sensors UAV and machine learning techniques at different water level. Three different water levels are field capacity (FC) water level, saturated (S) water level and flooded (FL) water level as treatment. The experimental design used in this experiment is Randomized Complete Block Design (RCBD). Three types of experimental units- namely rice, rice with weeds and weeds with 4 replications of blocking will apply to evaluate the outcome from hyperspectral sensors. This study used of machine learning for image analysis. The findings indicate that hyperspectral imagery can effectively identify weed species within the paddy field. This study showed that Neural Network (NN) followed by Random Forest (RF) and Support Vector Machine (SVM) the accurate machine learning to discriminate weeds from rice in this experiment about 80 % to 100 % accuracy and the best condition for weeds detection on saturated condition water level. The results demonstrated good separation accuracy for weeds and rice at 466, 554, 675, 677, 702, 726, 755, 762, 779 and 797 nm. The outcomes demonstrated that the ML could distinguish between different types of weeds and rice. The saturated water level, in this instance, was the most accurate in identifying weeds. It was determined that every weed could be found at 28 DAS and could be effectively distinguished using hyperspectral data and machine learning.