Sentinel-2 Multispectral Dataset for Coffee Crop Semantic Segmentation in the IGCV Region, Brazil

This dataset provides a collection of multispectral image patches and corresponding binary masks designed for the semantic segmentation of coffee plantations within the Campo das Vertentes Geographical Indication (IGCV), Brazil. Derived from harmonized Sentinel-2 Level-2A surface reflectance imagery, the dataset includes samples in 64×64, 128×128, and 256×256 pixel dimensions, maintaining a 10 meter spatial resolution. Each patch comprises five channels: Red (B4), Green (B3), Blue (B2), Near Infrared (B8), and NDVI. The corresponding ground truth masks are based on official vector data from EMATER MG. To ensure a stable spectral representation of the crop canopies, the imagery was processed using annual median composites synchronized with the reference year of each mask to minimize cloud and shadow interference. This collection is particularly suited for evaluating deep learning architectures in fragmented agricultural landscapes dealing with class imbalance. Finally, the dataset also includes model artifacts generated during the experiments, such as predicted masks, weights of the best performing models, and TensorBoard training logs. These resources enable the inspection of training dynamics, reproducibility of the results, and further benchmarking of segmentation models.本数据集收录了多光谱图像块与对应的二值掩码，旨在用于巴西Campo das Vertentes地理标志（IGCV）区域内咖啡种植园的语义分割任务。该数据集源自经协调校正的哨兵二号（Sentinel-2）L2A级地表反射率影像，包含64×64、128×128及256×256三种像素尺寸的样本，空间分辨率均为10米。每个图像块包含五个通道：红色波段（B4）、绿色波段（B3）、蓝色波段（B2）、近红外波段（B8）与归一化差异植被指数（NDVI）。对应的真值掩码基于巴西米纳斯吉拉斯州农业技术推广局（EMATER MG）的官方矢量数据生成。为确保作物冠层的光谱表征稳定性，所有影像均采用与各掩码对应参考年份同步的年度中值合成法进行预处理，以最大限度降低云层与阴影带来的干扰。该数据集尤其适用于在存在类别不平衡问题的碎片化农业景观中评估深度学习架构的性能。此外，本数据集还包含实验过程中生成的模型相关产物，例如预测掩码、最优模型权重以及张量板（TensorBoard）训练日志。这些资源可用于检视训练动态过程、复现实验结果，以及进一步开展分割模型的基准测试工作。