Multispectral images of apples for ripeness, sweetness and variety grading [Data set]

This dataset originates from an extensive research effort focused on the development and application of a cost-effective, custom-built multi-spectral imaging chamber designed for evaluating the quality of apples. The imaging setup is capable of capturing spectral reflectance across a range of wavelength bands, enabling detailed, non-invasive analysis of fruit characteristics such as ripeness, sweetness, and varietal differences. In this study, a total of 8 wavelengths were considered. All images were collected under uniform and controlled lighting to minimize environmental variability and enhance data consistency.

For Grading by Ripeness, three classes were considered: Under-Ripe, Ripe and Over-Ripe. For Grading by Ripeness, the following five types of apples were considered, each of four in quantity.
1. Red Delicious (USA)
2. Royal Gala
3. Red Delicious (New Zealand)
4. Washington
5. Kinnaur

For Grading by Variety, three classes- Red Delicious USA, Alpita and Royal Gala were considered. For Grading by Variety for all 3 types, each of the seven in quantity was taken.

For Grading by Sweetness, four classes were considered according to sugar content in % Brix in apples: 10, 12, 13, and 15 classes. For Grading by Sweetness, five varieties were considered, each of four in quantity. The following types of apples were considered for Grading by Sweetness.
1. Red Delicious (USA)
2. Royal Gala
3. Red Delicious (New Zealand)
4. Washington
5. Kinnaur
The images are concatenated with the help of MATLAB code and the concatenated dataset is created for grading by sweetness, ripeness and variety. For this study, APPLENET, a CNN-based architecture, was used to process the concatenated images, and the accuracy achieved was 87 %,65 % and 92 % for grading by ripeness, sweetness and variety, respectively.
The dataset is labelled and structured to support a wide range of applications, particularly in the domains of agricultural technology and food quality monitoring. It offers possible use cases for developing classification and regression models for fruit grading, maturity evaluation, and early detection of surface-level defects. Researchers working on agricultural automation, deep learning in food quality inspection, or horticulture may find this dataset particularly valuable.

Beyond apples, the methodology used for image acquisition and data annotation can be adapted for other fruits, offering scalability for broader agricultural research. The detailed documentation and consistent imaging protocol enhance reproducibility, making this dataset a useful benchmark for relative studies.

This data collection contributes meaningfully to ongoing efforts in computer vision and AI-powered agriculture by providing a reliable, annotated source of multi-spectral fruit images for non-destructive quality evaluation.

本数据集源自一项针对开发与应用低成本定制多光谱成像舱（multi-spectral imaging chamber）的系统性研究工作，该成像装置可捕获多个波段的光谱反射率，从而实现对果实特性的详细无创分析，涵盖成熟度、甜度以及品种差异等维度。本研究共设置8个采集波长。所有图像均在均匀可控的光照条件下采集，以最小化环境变量干扰并提升数据一致性。

在按成熟度分级任务中，共设置未成熟（Under-Ripe）、成熟（Ripe）、过成熟（Over-Ripe）三个类别。本次研究选用了以下5种苹果，每种各4个样本：
1. 蛇果（Red Delicious (USA)）
2. 皇家嘎啦果（Royal Gala）
3. 蛇果（Red Delicious (New Zealand)）
4. 华盛顿苹果（Washington）
5. 金瑙尔苹果（Kinnaur）

在按品种分级任务中，共设置美国蛇果（Red Delicious USA）、阿尔皮塔（Alpita）、皇家嘎啦果（Royal Gala）三个类别。针对该分级任务，共选用3类苹果，每类各7个样本。

在按甜度分级任务中，根据苹果的可溶性固形物含量（以°Brix为单位），共设置4个等级：10、12、13和15°Brix。本次研究共选用5种苹果，每种各4个样本，分别为：
1. 蛇果（Red Delicious (USA)）
2. 皇家嘎啦果（Royal Gala）
3. 蛇果（Red Delicious (New Zealand)）
4. 华盛顿苹果（Washington）
5. 金瑙尔苹果（Kinnaur）

研究人员通过MATLAB代码对图像进行拼接，分别构建了用于成熟度、甜度和品种分级的拼接数据集。本研究采用基于卷积神经网络（Convolutional Neural Network, CNN）的APPLENET架构处理拼接后的图像，最终在成熟度、甜度和品种分级任务上的准确率分别为87%、65%和92%。

本数据集已完成标注与结构化处理，可支持多种应用场景，尤其在农业技术与食品品质监测领域。其可用于开发果实分级、成熟度评估以及表面缺陷早期检测的分类与回归模型。从事农业自动化、食品品质检测深度学习以及园艺学研究的人员均可从本数据集获益。

除苹果之外，本研究采用的图像采集与数据标注方法可适配其他水果，为更广泛的农业研究提供可扩展的解决方案。详尽的文档与统一的成像流程提升了实验可重复性，使本数据集可作为相关对比研究的有效基准。

本数据集为计算机视觉与人工智能赋能农业的现有研究提供了可靠的标注多光谱果实图像资源，可用于无损品质评估，具有重要的研究价值。

搜集汇总

背景与挑战

背景概述

该数据集是一个用于苹果成熟度、甜度和品种分级的标注多光谱图像集合，来自一个低成本定制成像室的研究，包含8个波长的图像，在受控环境下采集以确保数据质量。数据集针对不同分级任务定义了具体类别（如成熟度分3类、甜度分4类、品种分3类），覆盖多个苹果品种，并已用于训练CNN模型，在成熟度、甜度和品种分级任务中分别达到87%、65%和92%的准确率。它支持农业自动化和食品质量监测应用，具有可扩展性和可重复性，适用于开发分类和回归模型。