A fast histogram equalization and KDE to aid a supervised algorithm to count Eucalyptus seedlings

Abstract

In a previous work, we developed two supervised approaches for automatic seedling counting in a UAV high-definition image from an Eucalyptus plantation stand of approximately 62 acres. We observed that the deployment of accessible and faster algorithms for automatic seedling/tree detection and counting suitable for field laptops are scarce. Moreover, researchers have reported troubles with the variable quality of the acquired aerial imagery, which plays a fundamental role in the algorithms degrading their discriminative power. Analyzing the bimodal BNDVI pixel distribution histograms, we noticed that, regarding the illumination quality, the distance between the two modes for any distribution seems to remain almost invariant. Therefore, the main core of the present work was to devise a fast BNDVI histogram equalization and to deploy a KDE in order to apply to a one-dimensional supervised counting method. From the resulting PDF, we defined a hard threshold to structure the solution as a binarized classification problem. We used the F1-score to predict the amount of seedlings within properly defined testing tiles taken from the large image. For our case of study, since the precision and recall resulted larger than 95%, the FP and FN are very small in comparison to the true number of seedlings.