Combining multisource remote sensing images using machine learning methods (RF and SVM) for improved cotton field mapping

Abstract

Large-scale crop mapping serves as a crucial data source for both cropland management and agricultural monitoring. This paper introduces an Improved Cotton White Index (ICWI) specifically developed to enhance the accuracy of cotton identification at the county level. To assess its efficacy, ICWI has been applied in five counties—Pars Abad, Arzuiyeh, Jafarabad, Behshahr in Iran and Moree Plains in Australia. These regions exhibit diverse climates and varying environmental conditions. Utilizing Sentinel 1 and Sentinel 2 time series data, Random Forest (RF) and Support Vector Machine (SVM) models, both with and without incorporating ICWI, were applied for identifying cotton farm in all five study areas. Additionally, the performance of the ICWI-based models was compared with that of models using the White Boll Index (WBI) to evaluate accuracy and robustness across different regions. The ICWI not only improves the accuracy of cotton identification but also contributes to comprehending the crop's phenology. Spectral analysis of the index's output enables the differentiation of various vegetative stages, from initial growth to full flowering. The analysis of results reveals that integrating the ICWI index into SVM and RF models markedly improves cotton identification accuracy across all regions. The ICWI index demonstrates a noteworthy 4 % overall accuracy boost, and an average increase of 9 % in Kappa. Importantly, in all study areas, our method achieved higher accuracy compared to the White Boll Index (WBI). The study's findings indicated that the proposed index has the potential to enhance the accuracy of cotton mapping using satellite time series images.