The ForestSeg Dataset

Our aerial forestry dataset (ForestSeg-T1) is a valuable resource comprising 1,832 high-resolution UAV images captured over the Vietnam National University of Forestry in Chuong My District, Hanoi. The dataset was collected through multiple drone surveys conducted at varying altitudes between 70m and 211m, enabling both detailed imagery and broad coverage of forested areas.

The dataset includes key flight parameters such as altitude, image resolution, and total coverage area. Two drone models were used: DJI Phantom 4 RTK (5472 × 3648 px) and DJI Air 3 (8064 × 6048 px), with each flight covering between 7.1 and 24.5 hectares. Images contain between 1 to 17 annotated tree instances, most commonly 8 to 9 trees per image, reflecting moderate forest density.

ForestSeg-T1 supports a wide range of applications, including forest monitoring, land cover classification, biomass estimation, and environmental change detection. The manually annotated tree crowns ensure high-quality ground truth for instance segmentation tasks, mimicking human perception. Unlike other datasets, ForestSeg‐T1 incorporates multi-seasonal imagery, making it particularly suitable for long-term forest analysis.

The pictures in the dataset are taken in Hanoi, Vietnam in 2023.

Some pictures of our dataset are shown below:

Forest Dataset

Detail

We investigated the impact of temporal variations on instance segmentation performance across four intervals: ForestSeg‐T1 to ForestSeg‐T4. Environmental factors such as lighting, shadow, and seasonal changes were considered to assess model robustness under different conditions.

The segmentation model achieved the best results on ForestSeg‐T4, with an AP of 60.82%, AP50 of 64.71%, and AP70 of 57.13%, aligned with its highest classification accuracy of 71.32%. ForestSeg‐T2 followed with 69.23% classification accuracy and moderate segmentation scores. In contrast, ForestSeg‐T3 yielded the lowest performance across all metrics, suggesting that data quality and timing significantly influence model effectiveness.

These findings demonstrate that selecting the appropriate temporal interval, particularly one with high-resolution imagery and favorable conditions, is crucial for enhancing both classification and segmentation performance. Incorporating temporal diversity into training can improve the model's generalizability and reliability in long-term forest monitoring applications.

The detailed performance results across the four intervals are summarized in the table below:
Table 01: Comparisons of segmentation performance of the proposed method with the state‐of‐the‐art methods.
**The best results for each metric are highlighted in bold.
Method Number of Parameters (M) Inference Time (ms) AP AP50 AP70
Mask-RCNN (ResNet-50) 43.05 8.3 30.63 46.23 26.17
Mask-RCNN (SwinT) 48.55 10.8 56.72 60.12 54.64
Detectree2 62.86 11.4 22.33 49.71 20.22
YOLOV11 22.33 7.5 38.30 52.78 33.51
Ours 10.81 6.2 57.01 62.21 55.32
Table 02: Experiment results in different intervals (The results are conducted without the area feature as described.)
Forest Name Classification Acc. AP AP50 AP70
ForestSeg-T1 65.37 57.01 62.21 55.32
ForestSeg-T2 69.23 59.72 62.11 55.12
ForestSeg-T3 59.01 54.92 57.28 50.01
ForestSeg-T4 71.32 60.82 64.71 57.13

Terms & Conditions of Use

The datasets are released for academic research only, and are free to researchers from educational or research institutes for non-commercial purposes.