Analysis on the Impact of Kernel Function Parameter Tuning on Support Vector Machines for Weed Patch Detection

Scientific and practical applications can immensely benefit from an accurate decision-based classifier system to automate camouflage target detection based on feature extraction(s) by incorporating UAV data. One such challenging and practical research problem is weed localisation and detection in sugarcane field. Correctly identifying, localising, and detecting weed is a tricky challenge due to a.) early stage(s) of weed patches which rapidly spread in the entire field and b.) weed’s merging color properties with the sugarcane crop. While machine learning has reached its zenith, support vector machines (SVM) still holds considerable potential for this type of research problem. In this paper, an analysis on the impact of varying support vector machine’s kernel functions and hyperparameters is presented which is implemented on weed infested sugarcane UAV imagery. The study being presented focuses on the 2-class classification (weed and non-weed) of UAV data from sugarcane field afflicted with weeds using tuning of kernel function-based SVM when the weeds are still in their early stages and spread in multiple camouflage patches. Therefore, in this paper a proposed methodology is presented which effectively resolves minimal weed patches (camouflage target) from sugarcane crop (natural background). Emphasis is placed on tuning the SVM classifier to get the most suited kernel function, viz, radial basis function in our study, with an accuracy of 97.5% for 300 px × 300 px weed patch raster.