DEEP LEARNING-BASED MULTI-CLASS WEED DETECTION FOR KAZAKHSTAN AGRICULTURE: A REGION-SPECIFIC DATASET AND HIERARCHICAL FIELD ANALYSIS SYSTEM

Abstract

Weeds reduce global crop yields by 30–50% through competition for water, nutrients, and sunlight, causing annual productivity losses exceeding $100 billion worldwide. In Kazakhstan, nearly half of the total cultivated area is infested with weeds, including 2.5 million hectares affected by black oats alone. Current broad-spectrum herbicide application creates critical problems including environmental contamination, 40–60% higher production costs, and accelerating development of herbicide-resistant weed strains. This study presents a comprehensive machine learning-based information system for multi-class weed species identification and classification in agricultural field conditions. We compiled a region-specific weed dataset containing 5,588 images (3,461 original + 2,127 augmented) with 7,659 bounding box annotations across 10 plant species (9 weeds and 1 crop), refined from an initial 15-class corpus through systematic class filtering and targeted augmentation to address class imbalance and inter-class confusion. Five classes with insufficient samples, high visual overlap, or poor model performance were excluded, reducing class imbalance from approximately 9:1 to 1.8:1. Three YOLO architectures were evaluated on this optimized dataset: YOLOv8l achieved 97.17% mAP@0.5, YOLOv26l achieved the highest accuracy at 97.70% mAP@0.5, and YOLOv26n achieved 96.73% mAP@0.5 while maintaining a lightweight 2.4M-parameter architecture suitable for edge deployment. Preliminary experiments on the original 5-class (90.0% mAP@0.5) and unfiltered 15-class (83.3% mAP@0.5) datasets established baselines that motivated the dataset refinement strategy. Furthermore, a novel hierarchical field analysis system was proposed that reconstructs 100-hectare field imagery from phone-captured samples using a structured zone-block framework (9 zones × 9 blocks × 4 images per block), enabling practical weed mapping without specialized drones or agricultural robots. The trained models were deployed in a full-stack web application (WeedScan KZ) built on FastAPI and React, providing single-image analysis, full-field reconstruction, an analytics dashboard with weed infestation percentages, zone-based heatmap visualization, species distribution analysis, and class-specific treatment recommendations.