MoundCount: A Detection-Based Approach for Automatic Counting of Planting Microsites on UAV Images
Planting by mounding is a commonly used forestry technique that improves soil quality and ensures optimal tree growth conditions. During planting operations, one of the main planning steps is to estimate the number of mechanically created mounds in each planting block. Traditional counting methods involve manual field surveys or human photo-interpretation of UAV images, which are generally subject to errors and time-consuming. In this work, we propose a new approach to count mounds on UAV orthomosaics. Our framework is designed to estimate the required number of seedlings for a given planting block, based on a visual detection approach and a global estimation module. Firstly, a deep local detection model is applied on local patches to recognize and count visible mounds. Then, an estimation model, based on global features is used to predict the final number of plant seedling required for a given plantation block. To evaluate the proposed framework in real-world conditions, we constructed a large UAV dataset, including 18 UAV orthomosaics, comprising 111,000 mounds. We have conducted extensive experiments in our dataset, including a comparison with the state-of-the-art counting methods, as well as an analysis of Human-Level Performance (HLP) in identifying and annotating mounds. The experimental results show that our model reaches the best performance in terms of MAE and MSE, by comparison to state-of-the-art automatic counting mehtods.