TPSet-4C

Fine-grained annotation was performed by jointly considering spectral characteristics, texture, shape, and spatial relationships in the original imagery. Particular attention was paid to distinguishing dry beaches, reservoir water, and dams, as well as to reviewing ambiguous regions such as internal class boundaries, narrow dam structures, and areas with complex background interference. The original image content was not modified during the construction of TPSet-4C.

Therefore, the acquisition dates, spatial resolution, sensor information, and geographic coverage are inherited from the original TPSet dataset. Users requiring detailed acquisition metadata should refer to the corresponding metadata or documentation of the original TPSet.The current version of TPSet-4C contains 1,389 training samples and 3,564 independent test samples, with each sample having a spatial size of 512 × 512 pixels.

In the training set, the pixel proportions of Background, Dry Beach, Reservoir Water, and Dam are approximately 92.82%, 4.60%, 0.66%, and 1.91%, respectively. In the test set, the corresponding proportions are approximately 99.757%, 0.140%, 0.036%, and 0.067%. A large number of test tiles contain no tailings-pond objects, resulting in strong class imbalance and sparse foreground targets.

These characteristics make TPSet-4C suitable for evaluating fine-grained recognition, false-positive suppression, and cross-region generalization under complex background conditions. In particular, Reservoir Water and Dam account for only a small fraction of all pixels, while dams often exhibit narrow and irregular spatial structures, posing additional challenges for small-object recognition and boundary delineation.The semantic annotations are spatially aligned with the corresponding remote sensing images on a pixel-by-pixel basis, and no synthetic interpolation is used to generate additional samples.

Real-world complexities in the source imagery, including shadows, bare land, natural water bodies, and other background objects with spectral or textural characteristics similar to tailings-pond components, are intentionally retained. These regions may constitute major sources of confusion during model training and evaluation. No systematic missing-data issue affecting normal use has been identified in the current release.

Potential annotation uncertainty mainly occurs at ambiguous class boundaries or in regions where limited spatial resolution, small object size, or spectral and textural similarity makes visual discrimination difficult. Users are therefore encouraged to consider the strong class imbalance when evaluating models and to report both class-wise and boundary-related performance metrics where appropriate.The dataset package mainly contains the original remote sensing images and their corresponding TPSet-4C semantic labels, with image–label pairs organized on a one-to-one basis.

The four semantic classes are Background, Dry Beach, Reservoir Water, and Dam.