PurposeAccurate information on the spatial distribution of crop yields is essential for the successful implementation of precision agriculture. Achieving high accuracy in yield monitoring necessitates the development of systems that effectively minimize errors caused by uneven field terrain and dynamic harvesting. This study aimed to estimate the real-time volume of radishes using computer vision techniques, applied to images captured under simulated vibration and sloped harvesting field conditions using a laboratory test bench.MethodsAn image acquisition system was developed and installed on a conveyor, which was mounted on a vibration table and a slope platform. Top-view RGB images of radishes were captured under slope levels of 3, 6, and 9 degrees; vibration levels of 0.43, 0.78, and 0.98 m/s2; and combinations of the slope and vibration conditions. The captured images were processed to remove noise induced by the slope and vibration conditions, extract size-related features, and estimate radish volume. Two approaches were used to estimate radish volume: ellipsoidal and multiple linear regression (MLR) models. The volume estimated was compared to the volume measured by the water displacement method (WDM).ResultsThe performance of the noise filtering algorithm was evaluated using peak signal-to-noise ratio (PSNR) and the structural similarity index measurement (SSIM), while the accuracy of the radish volume estimates was assessed through the coefficient of determination (R2). Analysis of variance (ANOVA) was also conducted to determine the statistical significance of the volume estimation results. Results indicated that radish volume was underestimated across all slope and vibration levels, compared to the values by the WDM. The MLR approach performed better than the ellipsoidal approximation approach. Average width demonstrated better performance compared to other geometrical attributes for volume estimation. No significant difference was observed between the mean estimated volumes by the MLR and the WDM method for all the slope and vibration conditions. The highest R2 values were achieved at 0.94, 0.96, and 0.92 for slope levels of 3, 6, and 9 degrees, respectively, when slope was tested independently using the MLR model.ConclusionsThis approach demonstrated potential for providing accurate volume estimates under uneven field conditions and dynamic harvesting.
