Deep learning-based temporal change detection of broadleaved weed infestation in rice fields using UAV multispectral imagery

Timely and accurate monitoring of weed infestation is essential for optimizing herbicide application in rice cultivation, particularly within site-specific weed management (SSWM) strategies. Conventional blanket spraying remains widely adopted by farmers, resulting in excessive herbicide usage and increased costs. This study presents a deep learning-based change detection approach to evaluate the temporal dynamics of broadleaved weed infestation in paddy fields. Multispectral imagery was collected using unmanned aerial vehicles (UAVs) over PadiU Putra rice fields, and a Deep Feedforward Neural Network (DFNN) was developed to classify three land cover types: paddy, soil, and broadleaved weeds during the vegetative stage. Post-classification comparison was applied to assess weed infestation rates across multiple Days After Sowing (DAS). The analysis revealed a consistent increase in weed coverage within untreated plots, with infestation rates rising from 40.95% at 34 DAS to 47.43% at 48 DAS, while treated plots remained largely controlled. The change detection maps further enabled estimation of potential herbicide savings through targeted application, indicating a possible reduction of up to 40.95% at 34 DAS. However, continued weed growth reduced this to 37.06%, with an R² of 0.9487, indicating a strong negative correlation between weed coverage and herbicide-saving potential. These findings demonstrate the potential of integrating UAV-based multispectral imaging with deep learning for temporal weed monitoring and precision agriculture applications.