Multistage AI Crop Inspection and Segmentation Framework (MACISF)
- 1UIET, Panjab University
Abstract
Crop quality assessment is one of the most critical components of modern agriculture, which directly influences yield optimization, post-harvest management, and food security. Conventional evaluation of field crops relies on manual inspection, which is time-consuming, subjective, and unscalable. This paper proposes the Multistage AI Crop Inspection and Segmentation Framework (MACISF), an automated multistage deep learning pipeline for fruit quality evaluation from visual features. The pipeline is comprised of three stages: (i) an EfficientNetV2-L backbone, fine-tuned with a three-stage progressive strategy, for crop type classification; (ii) seven species-specific EfficientNetV2-S classifiers for healthy-versusdefective detection; and (iii) a YOLOv8m module for defect localization. Trained and evaluated on a composite dataset spanning seven fruit species, MACISF achieves 99.23% crop type accuracy (macro F1-score 99.23%, macro AUC-ROC 0.993), 99.39% mean health state accuracy with 99.51% mean defective-class recall, and a projected grading [email protected] of 92.3%. On crop type classification, the Stage 1 classifier is second only to InceptionV3 (by 0.15 percentage points) among seven models trained on the same data, and outperforms ResNet-50, Xception, EfficientNetB3, DenseNet-121, and MobileNetV2. Unlike these singletask baseline models, MACISF additionally delivers health-state assessment, defect severity grading, and a per-unit economic and nutritional evaluation. Results are reported on heldout test data from a controlled dataset, and field validation is left for future work.
Related
- FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection
- Automatic bridge crack detections: A lightweight segmentation model for UAV inspection
- From segmentation to morphometry: A post-processed U-Net framework for corneal endothelial cell analysis
- Deep Learning-Meets-Active Surfaces: A Unified Framework for Multi-Object Segmentation and Shape Analysis in 3D Fluorescence Microscopy
- A Computer Vision-Based Framework for Multilevel Egg Fertility Classification Using L*a*b* Color Segmentation, Wavelet Texture Features, and Support Vector Machine
BibTeX
@misc{joshi2026multistage,
title = {{Multistage AI Crop Inspection and Segmentation Framework (MACISF)}},
author = {P. Joshi and A. Nagrath},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.11.758275},
url = {https://www.biorxiv.org/content/10.64898/2026.10.11.758275v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Multistage AI Crop Inspection and Segmentation Framework (MACISF)
AU - P. Joshi
AU - A. Nagrath
PY - 2026
DA - 2026/10/11
PB - bioRxiv
DO - 10.64898/2026.10.11.758275
UR - https://www.biorxiv.org/content/10.64898/2026.10.11.758275v1
AB - Crop quality assessment is one of the most critical components of modern agriculture, which directly influences yield optimization, post-harvest management, and food security. Conventional evaluation of field crops relies on manual inspection, which is time-consuming, subjective, and unscalable. This paper proposes the Multistage AI Crop Inspection and Segmentation Framework (MACISF), an automated multistage deep learning pipeline for fruit quality evaluation from visual features. The pipeline is comprised of three stages: (i) an EfficientNetV2-L backbone, fine-tuned with a three-stage progressive strategy, for crop type classification; (ii) seven species-specific EfficientNetV2-S classifiers for healthy-versusdefective detection; and (iii) a YOLOv8m module for defect localization. Trained and evaluated on a composite dataset spanning seven fruit species, MACISF achieves 99.23% crop type accuracy (macro F1-score 99.23%, macro AUC-ROC 0.993), 99.39% mean health state accuracy with 99.51% mean defective-class recall, and a projected grading [email protected] of 92.3%. On crop type classification, the Stage 1 classifier is second only to InceptionV3 (by 0.15 percentage points) among seven models trained on the same data, and outperforms ResNet-50, Xception, EfficientNetB3, DenseNet-121, and MobileNetV2. Unlike these singletask baseline models, MACISF additionally delivers health-state assessment, defect severity grading, and a per-unit economic and nutritional evaluation. Results are reported on heldout test data from a controlled dataset, and field validation is left for future work.
ER -