Validation of an oil-palm detection system based on a logistic regression model

Oil palm plantations cover large areas. One of the main problems is to get an updated census of plants contained in these fields. Nowadays this process is done manually generating subjective results of the number of plants and so economic losses. This paper presents the validation of an oil palm detection and counting system. It uses a logistic regression model to classify images acquired by an unmanned aerial vehicle. First, the photogrammetry software processes acquired images to create orthomosaics. Then, the developed computer vision algorithm analyzes them. It uses a sliding window technique in image pyramids to generate candidates, an LBP descriptor to mathematically model image texture and a logistic regression model to classify windows. Finally, it applies a non-maximum suppression algorithm to enhance the decision. The system was validated using different images to those used in the training process. This process allowed us to determine how each parameter affects the system behavior. Also, it was possible to conclude that the most relevant parameters to improve system performance were the median filter and the size of sliding windows. The final system was assessed with images of real plantations obtaining a detection error of 4.66 percent and an F1 score of 0.97.