Mapping Trees on Agricultural Land Using U-Net Semantic Segmentation from Multitemporal RGBI Orthophotos in Southern Saxony-Anhalt

Trees on agricultural land are key structural components of agroecosystems, contributing to essential ecosystem services like microclimate regulation, erosion control, biodiversity conservation, and the mitigation of climate-induced abiotic stresses, thereby enhancing the resilience of agricultural landscapes. However, existing inventories are often outdated, incomplete, and lack the spatial resolution necessary for in-depth analysis and effective decision-making. 
Therefore, we apply a semantic segmentation approach based on the U-Net architecture, to quantify the current spatial distribution of trees on agricultural lands across southern Saxony-Anhalt (approximately 4,000 km²). The model is based on official digital orthophotos (DOP) with 20 cm spatial resolution and a spectral resolution of four channels (RGBI).             
Given the large study area and the coarse repetition rate of aerial imagery, we further evaluate model performance across different acquisition dates, ranging from the beginning of the 2023 vegetation period (30.04. - spring) to the peak of the 2024 vegetation period (29.08. – late summer).
Training data generation uses a semi-automatic workflow: a normalized surface model is clipped into 512×512-pixel tiles, filtered to retain objects >4m height, and masked to exclude impervious surfaces. This produces 7,894 tiles containing 14,360 annotated features, which are manually verified against true-color imagery. An independent test set is created through manual digitization of agricultural trees, stratified by image acquisition date. Model performance is evaluated using Precision, Recall, F1-score, and Intersection over Union (IoU).
The dataset is split 70/30 for training/validation. Input data includes four channels (RGBI) and the Normalized Difference Vegetation Index (NDVI) as a fifth channel. Data augmentation applies random horizontal/vertical flips and rotations (±15°). The U-Net model is trained using focal Tversky loss (weighted to penalize both false positives and negatives) and the Adam optimizer with default learning rate.
Lowest model errors were reached after 48 epochs. The best-performing model is selected and subsequently applied to each DOP tile intersecting the study area, resulting in predictions for 1182 DOP tiles. First validation results on approximately 8000 reference polygons show an average F1-Score of 0.5 which is comparable to recent studies.          
A total area of 195 km² of trees on agricultural land are mapped. Despite the heterogeneity of acquisition dates, the model produces accurate segmentations and successfully identifies trees on agricultural land in different compositions. The results indicate that semiautomatic training data generation can compensate for seasonal variability in aerial images, which often hinders the application of deep learning models to larger spatial scales.