Robust Optical Flow Estimation for Wilted Plant Leaves Motion: Predicting Water-Stressed and Well-Watered Conditions

Optical flow estimation is essential in computer vision for analyzing and interpreting motion in image sequences. For agriculture, optical flow can be a powerful tool in predicting plant health, particularly in identifying water-stressed versus well-watered conditions in wilted plant leaves. While traditional methods for optical flow estimation have made significant progress, challenges such as ill-defined edges and blurred motion boundaries remain, particularly when dealing with complex motion patterns like those found in natural leaf movements. This paper investigates the robustness of non-local total variation methods in optical flow estimation for wilted plant leaves. By leveraging the strengths of non-local total variation-based approaches, we aim to improve motion boundary preservation in wilted leaf motion sequences, offering a way to differentiate between water-stressed and well-watered leaves. We review existing non-local total variation methods, analyze their performance on wilted plant datasets, and present a new approach integrating weighted guided filtering to handle edge preservation. The findings suggest that non-local total variation with weighted guided filter achieves robust optical flow estimations, offering potential for precision agriculture applications, particularly in detecting early signs of drought stress.