Search for a command to run...
Dataset Description This dataset consists of 1,390 corn leaf images collected from Pokhara-16, Gandaki Province, Nepal. The images capture pest-induced damage under real field conditions and are intended for object detection and classification tasks. All images were pre-processed by removing blurred or non-informative samples and cropping images containing multiple leaves to ensure a single-leaf focus. Annotation Image annotation was performed using the LabelImg tool in collaboration with an agricultural expert. This dataset includes the following pest damage categories: Grasshopper damage, Fall armyworm damage. Some images contain both types of damage within a single leaf. Data Splitting and Cross-Validation The dataset is divided into five mutually exclusive subsets: Split1, Split2, Split3, Split4, Split5. Each split contains approximately 20% of the total dataset. A stratified sampling strategy was applied based on: Grasshopper-only damage, Fall armyworm-only damage, combined damage. A 5-fold cross-validation protocol can be adopted. In each iteration: Training: Validation: Test = 3splits: 1split: 1split. For this the folds should be rotated such that each split serves: as training set three times, as validation set once, as test set once. Artifact Augmentation (Foreign Attached Substances-FAS) To improve model performance against false detections, additional artifact-based augmentation was performed. A total of 40 FAS patches were collected and categorized as: 8 water spots, 12 soil splashes, 2 pest bodies, 18 other external artifacts. Each FAS patch was embedded into healthy leaf images by sampling random positions within the central 50% of the leaf region ensuring placement within the leaf mask. An alpha blending technique based on distance transform was applied. The interior region of the patch remains unchanged while the boundary regions are smoothly blended with surrounding pixels. This process produces 40 artifact-augmented images. The artifact-augmented images with empty annotation can be included only in training sets as negative sample enabling the model to learn to ignore non-pest artifacts reduce false positives caused by foreign substances.