Multimodal Cotton Seedling Dataset with RGB, RealSense Infrared, Depth and Thermal Images for YOLO-Compatible Agricultural Row Perception

This dataset contains a multimodal agricultural monitoring package focused on cotton seedlings and agricultural row perception for low-cost robotic monitoring systems. The dataset was prepared for research in precision agriculture, agricultural robotics, computer vision, multimodal sensing, crop-row perception and YOLO-compatible annotation workflows. The dataset was acquired in the Mexicali Valley, Baja California, Mexico, using a non-motorized experimental data-acquisition platform. The package includes visual data from multiple modalities: multiview RGB images, RealSense color images, RealSense infrared images, RealSense depth or depth-visualization products, and thermal images or thermal visualizations. The dataset is distributed as three trajectory-level ZIP files, each corresponding to one acquisition trajectory. The validated trajectory-balanced pilot annotation subset contains 240 selected multimodal images, corresponding to 80 distinct reviewed images per trajectory. For each trajectory, the subset includes 20 RGB images, 20 RealSense infrared images, 20 RealSense depth/depth-visualization images and 20 thermal images or thermal visualizations. Across the three trajectories, this corresponds to 60 RGB images, 60 RealSense infrared images, 60 RealSense depth/depth-visualization images and 60 thermal images or thermal visualizations. The YOLO-compatible annotation package was prepared manually in CVAT and reviewed manually. The reviewed task-level YOLO inventory contains 339 non-empty label files and 1,678 parsed annotated instances across the three trajectories. No missing labels or malformed label lines were detected in the reviewed trajectory-level scans. The label inventory should not be interpreted as 339 unique images, because the same selected image may appear in different task-specific YOLO exports, including segmentation, oriented bounding box and pose-style workflows when available. The annotation folders are compatible with YOLO-style workflows and include segmentation, oriented bounding box and pose-style tasks when available. The annotations are provided as a trajectory-balanced pilot multimodal annotation subset and should not be interpreted as exhaustive ground truth for the complete dataset. The package also includes dataset documentation, file-level indexes, checksums, quality-control reports, processed tables and article-oriented figures. Important files include README.md, master_index.csv, checksums_sha256.csv, dataset_summary.csv, variable_dictionary.csv, quality_report.csv, annotation_summary.csv, yolo_total_image_inventory.csv, yolo_image_label_match_report.csv, processing_method_summary.csv and publication_readiness_validation files. The dataset is organized into folders such as raw/, organized/, processed/, derived/, annotations/, article_tables/, article_figures/, docs/ and code/. The current deposited structure includes the main scripts used for the acquisition GUI, post-processing GUI and interpretive dashboard in the code/ folder of the second trajectory package. These scripts are provided as reproducibility and inspection resources, not as independently validated model-performance software. The package is intended to support reproducible reuse, dataset inspection, multimodal agricultural perception experiments, preliminary YOLO-compatible benchmarking, image-quality analysis, sensor-related post-processing and Data in Brief-style documentation. Known limitations should be considered before reuse. RGB vegetation products, if included, should be interpreted as RGB-derived proxies and not as calibrated multispectral indices. Thermal images should not be used alone to infer water stress, disease, irrigation condition or crop physiological status. RealSense infrared data should not be interpreted as calibrated agronomic near-infrared data. RealSense depth visualizations may represent encoded depth products rather than metric depth unless raw depth and scale metadata are available. Multimodal synchronization should be considered approximate unless external hardware synchronization is explicitly documented. GPS, IMU and MTF-01 records should be treated as field-acquisition and quality-control records, not as surveyed navigation, localization or topographic ground truth. The YOLO annotations are a pilot subset and not complete annotation coverage of the full dataset.