This study aims to emphasize the importance of historical land cover maps and introduce a new benchmark dataset derived from very high-resolution historical Hexagon (KH-9) reconnaissance satellite images for use in deep learning-based image segmentation tasks. Our HexaLCSeg dataset comprises high-resolution monochromatic Hexagon images from the 1970s and 1980s covering Turkish and Bulgarian territories, encompassing a large geographic area
The dataset, related source code, and pre-trained models are available below.
Our dataset is inspired by the European Space Agency (ESA) WorldCover project and includes eight LC classes and related RGB codes were set for each class but we adjusted the 0-pixel value as no data and replaced the 0 values with 1 in the ESA RGB code palette. Additionally, a new sub-class for the trees, named Permanent Cropland is defined and its RGB code was set to 1-207-117. This class is important to differentiate permanent fruit trees from other trees, specifically crucial for past agricultural mapping purposes.
The HexaLCSeg dataset comprises eight panchromatic images accompanied by corresponding 3-channel RGB Ground Truth Masks, all with 8-bit radiometric resolution and a spatial resolution of 1 meter. The dataset is organized into a total of 10,000 patches, each sized at 256x256 pixels. We split our dataset into 70% training (7000 patches), 15% validation (1500 patches), and 15% testing (1500 patches). Samples of our dataset are below.
The dataset can be downloaded here.
In our study, we employed the geographic object-based image analysis (GEOBIA) approach to generate accurate land cover (LC) maps, which serve as the ground truth masks for our dataset.
For deep learning-based image segmentation, we employed a total of 9 CNN models, implementing U-Net++ and DeepLabv3+ segmentation architectures with different hyperparameters, paired with SE-ResNeXt50 backbone that pre-trained with weight values from the 2012 ILSVRC ImageNet dataset.
| Model No | Architecture | Loss Function | Augmentation | Loss | Accuracy | IoU | F-1 Score | Precision | Recall |
|---|---|---|---|---|---|---|---|---|---|
| Model 1 | U-Net++ | Focal Loss | No Augmentation | 0.1252 | 0.9734 | 0.8052 | 0.8804 | 0.8805 | 0.8803 |
| Model 2 | U-Net++ | Focal Loss | Horizontal Flip | 0.1253 | 0.9728 | 0.8008 | 0.8776 | 0.8778 | 0.8774 |
| Model 3 | DeepLabv3+ | Focal Loss | No Augmentation | 0.1255 | 0.9720 | 0.7959 | 0.8739 | 0.8744 | 0.8734 |
| Model 4 | U-Net++ | Focal Loss | Random BC | 0.1256 | 0.9717 | 0.7938 | 0.8725 | 0.8727 | 0.8723 |
| Model 5 | DeepLabv3+ | Dice Loss | Horizontal Flip | 0.1292 | 0.9714 | 0.7928 | 0.8714 | 0.8717 | 0.8711 |
| Model 6 | DeepLabv3+ | Dice Loss | No Augmentation | 0.1307 | 0.9711 | 0.7906 | 0.8699 | 0.8702 | 0.8697 |
| Model 7 | DeepLabv3+ | Focal Loss | Horizontal Flip | 0.1257 | 0.9711 | 0.7897 | 0.8698 | 0.8704 | 0.8692 |
| Model 8 | DeepLabv3+ | Focal Loss | Random BC | 0.1259 | 0.9704 | 0.7871 | 0.8667 | 0.8673 | 0.8662 |
| Model 9 | DeepLabv3+ | Dice Loss | Random BC | 0.1401 | 0.9691 | 0.7793 | 0.8608 | 0.8612 | 0.8604 |
The pre-trained models and weights can be found here.
Original images, ground truth masks and prediction masks of our 5 best models are below.
The code was implemented in Python (3.10) Programming Language.
- torch == 2.1.2
- segmentation-models-pytorch == 0.3.3
- Albumentations == 1.4.0
Apart from main data science libraries, RS-specific libraries such as GDAL, rasterio, and tifffile are also required.
Please kindly cite our paper if this code and the dataset used in the study are useful for your research.
E. Sertel, M. E. Kabadayi, G. S. Sengul and I. N. Tumer, "HexaLCSeg: A historical benchmark dataset from Hexagon satellite images for land cover segmentation [Software and Data Sets]," in IEEE Geoscience and Remote Sensing Magazine, vol. 12, no. 3, pp. 197-206, Sept. 2024, doi: 10.1109/MGRS.2024.3394248.