Researchers have released a dataset of 14,642 images that tracks maize leaf damage from four major pest species across their larval development, offering a chronological benchmark for training computer-vision models to spot early herbivory.
The collection, published November 21 in Scientific Data, records feeding patterns of fall armyworm, tobacco cutworm, oriental armyworm, and black cutworm, with each image annotated for pest species, estimated larval instar, and chronological age.
The dataset addresses a gap in existing computer-vision resources. The authors say those resources focus on late-stage, macroscopic defoliation and miss the temporal progression of early-stage herbivory. Standardized rearing combined with high-resolution flatbed scanning captured the micro-textural boundaries of bite marks, producing tripartite annotations that link visual phenotypes to developmental timelines.
In a separate study published November 2 in Frontiers in Plant Science, a different team introduced a stage-wise collaborative feature enhancement framework for detecting maize insect pests in complex field scenes.
Built on the RT-DETR-R18 detector, the model integrates three modules: a Similarity-aware Decoupling Module to strengthen fine-grained category discrimination during backbone feature extraction, Adaptive Sparse Self-Attention to regulate spatial interactions under complex backgrounds, and an Efficient Up-Convolution Block to recover spatial detail for small targets during top-down feature fusion.
Two Papers Pursue Complementary Research Directions
The two publications reflect complementary directions. One establishes a controlled-environment benchmark for fine-grained visual classification and developmental domain adaptation. The other targets deployment-ready detection in complex agricultural scenes. The Scientific Data authors note their dataset provides a foundation for future early-warning research. The Frontiers in Plant Science authors caution that further field and deployment evaluation is required to establish practical applicability.
The Scientific Data paper, titled “A Spatio-temporal Image Dataset of Maize Leaf Herbivory by Four Major Pests,” was received on May 22, 2026, accepted on September 28, 2026, and published on November 21, 2026, under DOI 10.1038/s41597-026-08441-3.
The author group includes Qin, Z., Xu, X., Zhang, Q. and colleagues. The authors declare no competing interests. The publisher notes that Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The dataset comprises 14,642 unaugmented images collected in a controlled environment. Feeding patterns were recorded from four primary pest species: Spodoptera frugiperda (fall armyworm), Spodoptera litura (tobacco cutworm), Mythimna separata (oriental armyworm), and Agrotis ipsilon (black cutworm).
Baseline multi-task learning and cross-instar evaluations were performed to quantify morphological distribution shifts occurring across developmental stages. The authors state that by linking visual phenotypes to developmental timelines, the dataset establishes a controlled-environment benchmark for evaluating fine-grained visual classification and developmental domain adaptation, providing a foundation for future early warning research.
The Frontiers in Plant Science paper, titled “Stage-wise collaborative feature enhancement for accurate maize insect pest detection in complex agricultural scenes,” appears in Volume 17, Article 2011789, published November 2, 2026, under DOI 10.3389/fpls.2026.2011789.
The authors are Jiayi Zuo and Cong Zhang from the School of Electrical and Electronic Engineering, Wuhan Polytechnic University, Wuhan, China; Donghan Liu from the School of Information and Electronics, Beijing Institute of Technology, Beijing, China; and Zheng Wang from the School of Computer Science, Wuhan University, Wuhan, China.
Dataset Accessible via Roboflow Universe
The dataset is accessible at https://universe.roboflow.com/526/maize-pest (accessed September 8, 2026). During model training and evaluation, all images were resized to 640 × 640 pixels.
The separately constructed IP102-Maize dataset was derived from the IP102 v1.1 Detection benchmark (Wu et al., 2019). Based on the crop-level taxonomy provided by IP102, 13 maize-associated pest categories were selected, yielding 4,533 images.
The distribution of images and annotated instances in IP102-Maize is reported as 2,900 training images with 3,516 instances, 726 validation images with 918 instances, and 907 test images with 1,109 instances, for a total of 4,533 images and 5,543 instances. IP102-Maize was used solely for additional validation in a separately trained setting.