A Distance-Transform-Based Adaptive Non-Maximum Suppression Approach for Automated Apple Tree Detection in Intensive Orchards

Automatic tree detection plays a significant role in terrain analysis, planning and decision-making across a range of sectors, including forestry, agriculture and road accessibility analysis. This article focuses on the automatic detection of individual trees from Airborne Laser Scanning (ALS) data. Existing methods have been developed mainly for forests and urban green spaces, where trees are characterised by extensive crowns; in orchards, they are primarily suitable for mature trees. There is therefore a need to refine detection methods dedicated to intensive orchards, taking into account both young and mature trees. The study utilised ALS data with a density of 12 points/m2 on the basis of which a Canopy Height Model (CHM) was generated; the impact of various noise reduction methods on the effectiveness of tree detection was then compared. A new method for detecting treetops was proposed, Distance-Transform-Based Adaptive Non-Maximum Suppression (DTB-ANMS), combining adaptive reduction in local maxima with a radius dependent on the distance between potential vertices. Validation was carried out on 10 test areas comprising 2272 trees in young and mature orchards. The bilateral filter provided the best compromise between noise reduction and the preservation of fine crown structures, whilst DTB-ANMS achieved the highest overall performance (Precision 0.944, F1-score 0.875, Recall of 0.816), outperforming the classic Marker-Controlled Watershed and the local maxima method. The results indicate that publicly available, low-density ALS data can form the basis for an effective, low-cost tree inventory in intensive apple orchards.