Data for PhenoCam Images Raspberry Pi Models for Corn Growth Stage Classification

This dataset on “Data for PhenoCam images Raspberry Pi models for corn growth stage classification” includes the original data, intermediate model outputs, and final results obtained (graphical and tabular forms) from the research study. The original data (images) consists of 10 different corn growing PhenoCam sites in US, which were annotated for growth stage labeling manually based on visual appearance. The intermediate model output includes learning curves (training and validation log), and predictions of test images, for respective deep learning (DL) models (4) across individual PhenoCam sites (10) and image clipping levels (5). The final results after model development and testing comprise confusion matrices and classification reports. The lightweight DL models developed for this study are ELiteCrop0, ELiteCrop1, ELiteCrop4, and MobNetCropV2 using transfer learning techniques. The results are based on site-wise and five PhenoCam image clipping levels (0-40%) across individual models. The dataset provides model performance evaluation results based on intrasite (training and testing on the same PhenoCam site) and intersite (testing on a new PhenoCam site) methods. Initially, models were developed using a supercomputer (CCAST, NDSU), and they were finally deployed on Raspberry Pi (single board computer). The dataset documents the implementation and performance of the finalized model (ELiteCrop0) deployed on Raspberry Pi, including inference time, computational efficiency, and hardware utilization metrics. The contents of the dataset include: 1. Abstract, 2. PhenoCam data annotation visual class labels, 3. Training and validation accuracy and loss curves, 4. Confusion matrix plots, 5. Sample prediction plots, 6. Classification report, 7. CPU timing, 8. Sample prediction plots for Raspberry Pi, 9. Intersite evaluation with ELiteCrop0, and 10. Intersite sample prediction plots for Raspberry Pi with ELiteCrop0.本数据集为面向玉米生育期分类的PhenoCam图像树莓派模型数据集，包含本研究中获取的原始数据、模型中间输出结果及最终成果（含图形与表格形式）。原始图像数据涵盖美国境内10个不同玉米种植物候相机（PhenoCam）监测站点，所有图像均基于视觉外观完成人工标注，以标定对应生育期标签。 模型中间输出结果包括针对4种深度学习（Deep Learning, DL）模型，在10个物候相机站点及5种图像裁剪级别下的学习曲线（训练与验证日志）以及测试集图像预测结果。模型开发与测试后的最终成果包含混淆矩阵与分类报告。 本研究开发的轻量级深度学习模型包括ELiteCrop0、ELiteCrop1、ELiteCrop4与MobNetCropV2，均采用迁移学习技术构建。实验结果基于各模型在站点级以及5种（0-40%）物候相机图像裁剪级别下的表现。本数据集提供了基于站点内（在同一物候相机站点开展训练与测试）以及站点间（在全新物候相机站点开展测试）两种评估方法的模型性能评价结果。 模型最初依托超级计算机（CCAST, NDSU）开发，最终部署于树莓派（Raspberry Pi）单板计算机。数据集还记录了最终部署于树莓派的ELiteCrop0模型的实际运行情况与性能表现，包括推理时长、计算效率及硬件利用率指标。 数据集内容如下：1. 摘要；2. 物候相机数据标注视觉类别标签；3. 训练与验证准确率及损失曲线；4. 混淆矩阵图；5. 样本预测图；6. 分类报告；7. CPU计时数据；8. 树莓派端样本预测图；9. ELiteCrop0模型的站点间评估结果；10. 采用ELiteCrop0模型的树莓派端站点间样本预测图。