Weed classification in sugarcane fields in Northeast Thailand from multi-temporal Sentinel-1 and Sentinel-2 data together with random forest algorithm

Timely and accurate weed detection is essential for sustainable crop production and management. The integration of multiple satellite data sources with powerful machine learning has transformed precision agriculture by enhancing the accuracy and automation of object classification, enabling large-scale analysis and real-time predictions. However, challenges remain in effectively managing agricultural practices, particularly in weed control. This study employed Sentinel-1 (S1) and Sentinel-2 (S2) satellite data, combined with vegetation indices and random forest (RF) classification algorithm, to map weed presence in sugarcane fields in Northeastern Thailand. The large number of reference data consisting of 744 points was utilized to train and validate weed identification. The combined S1 and S2 dataset significantly enhanced the detection capabilities of the best RF model, achieving an overall classification result of 96 % accuracy and F1 scores exceeding 93 %. While overall weed levels were low, several high-density zones were clearly detected, underscoring the importance of targeted weed management. The combination of S1 and S2 data improved classification performance, addressing challenges posed by mixed pixels in small fields. Stratifying weed density provided deeper insights into field variability over the large scale. Our work presents a scientifically robust and operationally scalable framework for monitoring weed infestations in sugarcane cultivation. The proposed approach demonstrates strong potential for advancing sustainable precision agriculture by facilitating timely and spatially precise interventions. • Weeds pose a significant threat to agricultural productivity. • The Random Forest (RF) well classified weed and non-weed within fields. • The Sentinel datasets achieved high efficiency in mapping weeds. • The spatial distribution patterns of weeds are well demonstrated.