Abstract: During the blueberry harvesting season, the pre-identification of fruit ripeness is of great importance for harvest management and yield assessment. The hyperspectral (HS) imaging technology based on remote sensing which contains abundance of spectral and spatial information, provide great potential for development in target detection from complex background. In this study, HS images of three representative southern highbush blueberry varieties were collected in Citra and Waldo, FL, USA in the 2013 harvesting season. Three maturity stages (mature, intermediate and immature) of blueberry fruit were identified from the images containing different background objects such as leaves, branches, sky, and ground. A joint algorithm (named as ‘SVDD+K’) of blueberry fruit maturity recognition model by support vector data description (SVDD) and K-means clustering was used to discriminate mixed blueberry fruit based on pixels and appearance of different objects. The results showed that the detection accuracies were 96.1% for mature fruit, 94.7% for intermediate fruit and 91.2% for immature fruit. To evaluate the performance of the proposed algorithm, other classification methods, i.e., K-nearest neighbor and spectral angle mapping, were used to compare the results. With the highest detection accuracy, the newly developed SVDD+K algorithm was more adaptable to complex backgrounds, especially for small-size objects, which were more accurate in recognition of small-size objects.

Key words: remote sensing, blueberry, harvesting, smart agriculture, target detection

摘要： 在蓝莓收获期，果实成熟度识别的预先识别对果实收获管理及产量评估具有重要意义。基于遥感的高光谱成像技术，以其包含丰富的光谱及空间信息，在复杂背景对象识别中具有极大的开发潜力。通过采集美国佛罗里达州三种典型南部高丛蓝莓的高光谱图像，研究了基于高光谱图像处理技术在复杂背景中进行蓝莓果实成熟度（成熟、近成熟以及未成熟）识别的数据处理算法。通过支持向量数据描述（SVDD)和K-means聚类算法，构建了蓝莓果实成熟度识别模型，从像素层面和外观层面对蓝莓果实进行混合识别。实验结果表明：对蓝莓成熟果实的识别正确率达到96.1%，近成熟果实识别率为94.7%，青果识别率为91.2%。为了评估算法性能，使用另外两种对象识别算法（KNN和SAM）作为比较，试验表明新提出的算法对复杂背景适应性更好，尤其是对小尺寸对象识别准确率更高。

关键词: 遥感, 蓝莓, 收获, 智慧农业, 对象识别