Close range spectra imaging of agricultural plants is widely performed to support digital plant phenotyping, a task where physicochemical changes in plants are monitored in a non-destructive way. A major step before analyzing the spectral images of plants is to distinguish the plant from the background. Usually, this is an easy task and can be performed using mathematical operations on the combinations of selected spectral bands, such as estimating the normalized difference vegetative index (NDVI). However, when the background of plants contains objects with similar spectral properties as plant then the segmentation based on the threshold of NDVI images can suffer. Another common approach is to train pixel classifiers on spectra extracted from selected locations in the spectral image, but such an approach does not take the spatial information about the plant structure into account. From a technical perspective, plant spectral imaging for digital phenotyping applications usually involves imaging several plants together for a comparative purpose, hence, the imaging scene is relatively big in terms of memory. To solve the challenge of plant segmentation and handling the memory challenge, this study proposes a novel approach, which combines chemometrics with advanced deep learning (DL) based semantic segmentation. The approach has four key steps. As a first step, the spectral image is pre-processed to reduce illumination effects present in the close-range spectral images of plants resulting from the interaction of light with complex plant geometry. Different chemometric pre-processing methods were explored to find possible improvements in the segmentation performance of the DL model. The second step was to perform a principal components analysis (PCA) to reduce the dimensionality of the images, thus drastically reducing their size so that they can be handled more easily using the available computer memory during the training of the DL model. As the third step, small random images (128 × 128) were subsampled from the tall and wide image matrices to generate the training and validation sets for training the DL models. In the last step, a U-net based deep semantic segmentation model was trained and validated on the sub-sampled spectral images. The results showed that the proposed approach allowed efficient handling and training of the DL segmentation model. The intersection over union (IoU) scores for the segmentation was 0.96 for the independent test set image. The segmentation based on variable sorting for normalization and standard normal variate pre-processed data achieved the highest IoU scores. A combination of chemometrics and DL led to an efficient segmentation of tall and wide spectral images which otherwise would have given out-of-memory errors. The developed method can facilitate digital phenotyping tasks where close-range spectral imaging is used to estimate the physicochemical properties of plants.

农业植物的近距离光谱成像被广泛用于支持数字植物表型分析，这是一项以非破坏性方式监测植物物理化学变化的任务。分析植物光谱图像之前的一个主要步骤是将植物与背景区分开来。通常，这是一项简单的任务，可以对所选光谱带的组合使用数学运算来执行，例如估计归一化差异植物指数 (NDVI)。然而，当植物的背景包含与植物具有相似光谱特性的对象时，基于 NDVI 图像阈值的分割可能会受到影响。另一种常见的方法是对从光谱图像中选定位置提取的光谱训练像素分类器，但这种方法没有考虑植物结构的空间信息。从技术角度来看，用于数字表型应用的植物光谱成像通常涉及将多个植物一起成像以进行比较，因此，就记忆而言，成像场景相对较大。为了解决植物分割的挑战和处理记忆挑战，本研究提出了一种新方法，将化学计量学与基于高级深度学习 (DL) 的语义分割相结合。该方法有四个关键步骤。作为第一步，光谱图像经过预处理，以减少由于光与复杂植物几何形状的相互作用而导致植物近距离光谱图像中存在的照明效果。探索了不同的化学计量学预处理方法，以发现 DL 模型分割性能的可能改进。第二步是执行主成分分析 (PCA) 以减少图像的维数，从而大大减少它们的大小，以便在 DL 模型训练期间使用可用的计算机内存更容易地处理它们。作为第三步，从高和宽图像矩阵中对小的随机图像（128 × 128）进行子采样，以生成用于训练 DL 模型的训练和验证集。在最后一步中，基于 U-net 的深度语义分割模型在子采样光谱图像上进行了训练和验证。结果表明，所提出的方法允许有效处理和训练 DL 分割模型。对于独立测试集图像，分割的联合交集 (IoU) 分数为 0.96。基于标准化变量排序和标准正态变量预处理数据的分割获得了最高的 IoU 分数。化学计量学和深度学习的结合导致高光谱图像和宽光谱图像的有效分割，否则会产生内存不足错误。所开发的方法可以促进数字表型分析任务，其中使用近距离光谱成像来估计植物的理化特性。化学计量学和深度学习的结合导致高光谱图像和宽光谱图像的有效分割，否则会产生内存不足错误。所开发的方法可以促进数字表型分析任务，其中使用近距离光谱成像来估计植物的理化特性。化学计量学和深度学习的结合导致高光谱图像和宽光谱图像的有效分割，否则会产生内存不足错误。所开发的方法可以促进数字表型分析任务，其中使用近距离光谱成像来估计植物的理化特性。