UAVLitchi: A UAV-based Benchmark Dataset for Litchi Fruit Segmentation and Detection in Natural Environment

The dataset is specifically designed for litchi detection and segmentation, with the objective of advancing research on fruit detection and analysis in complex orchard environments. Currently, there is a lack of unified data to address challenges such as occlusion, overlapping, uneven coloring, and blending with surrounding foliage, which hinders deep learning models from achieving reliable comparisons and results. UAVLitchi aims to fill this gap by providing a diverse set of high-resolution images to reduce these complexities in natural environments. The UAVLitchi dataset includes a total of 5000 images, each sized 256 × 256 pixels, with 4500 RGB images and 500 multispectral images. The dataset consists of 83,250 instances of litchis, with an average object size covering 0.62% of the image area, making it best suitable for small object detection applications. Detailed statistical analyses of the dataset's segmentation performance across SOTA models are provided, demonstrating its suitability for a wide range of precision agriculture applications, including crop yield estimation and autonomous fruit counting.