Increasing efficiency and productivity in the field of agriculture is important to provide sufficient food to the world’s increasing population. It is important to monitor crops using image processing in order to realize these increases in efficiency and productivity. In order to monitor crops with high quality and accuracy, high resolution images are needed. In this research, a crop monitoring method for pecan nut trees was developed using high-resolution video taken from the side of a vehicle. First, trees were extracted by applying an object detection model to the video data. Second, the extracted trees were divided into canopy and trunk areas. Finally, using labels made by experts and the canopy image as input, the convolutional neural network (CNN) model was trained to classify unhealthy and healthy trees. The model achieved an area under the curve for classification over 0.95. Gradient-weighted Class Activation Mapping (Grad-CAM) was also applied to the model for the purpose of evaluation, and it clarified that the model is focusing on the hollow features of the canopy when performing its classification.

提高农业领域的效率和生产力对于为世界不断增长的人口提供足够的食物非常重要。为了实现效率和生产力的这些提高，使用图像处理来监控作物非常重要。为了高质量和准确地监测作物，需要高分辨率图像。在这项研究中，使用从车辆侧面拍摄的高分辨率视频开发了一种山核桃树的作物监测方法。首先，通过将对象检测模型应用于视频数据来提取树木。其次，提取的树木分为树冠区和树干区。最后，使用专家制作的标签和冠层图像作为输入，训练卷积神经网络 (CNN) 模型对不健康和健康的树木进行分类。该模型实现了超过 0.95 的分类曲线下面积。梯度加权类激活映射（Grad-CAM）也被应用于模型进行评估，并阐明模型在执行分类时关注冠层的空心特征。