In order to realize rice spike layer height detection, a research on rice spike layer height calculation method based on depth camera was carried out to provide technical support for the height control of harvester header. In order to select a color space suitable for rice spike layer segmentation, this paper visualized mature rice images in RGB, HSV, LAB and i1i2i3 color spaces in 3D space respectively, and analyzed the separability of rice spike layer by observing pixels distribution. The rice spike layer was extracted based on the multithreshold segmentation method, small joint region removing and image erosion. A spike layer heights calculation method was proposed based on a depth camera and the method was validated in the laboratory and in the field. The results showed that the rice spike layer had the best separability in HSV color space, and the adopted image segmentation and processing methods could effectively segment the rice spike layer. The laboratory test of the developed calculation method showed that the calculation errors of both Hhs and Hls were less than 1.5%, and the field test showed that the calculation errors of Hhs and Hls were 1.6% and 18.2%, respectively.

This paper aims to solve the problems of low recognition rate of potato leaf diseases and difficulty in localization of late blight spots in the natural environment based on potato leaf images collected in the field environment. Firstly, the potato leaf disease was identified using five neural network models: ALEXNET, VGG16, InceptionV3, RESNET50, and MobileNet. The results show that the InceptionV3 model has the highest recognition accuracy, which can reach 98.00%. Secondly, the late blight spots of potato leaves were detected, and an improved CenterNet-SPP model was proposed. The model obtains the center point of the object through the feature extraction network and then obtains the image information such as the offset of the center point and the size of the target through the regression of the center point. The mAP of the trained model under the verification set was up to 90.03%.Using F ₁ as the evaluation value analysis, compared with other objective detection models, the CenterNet-SPP model has the best effect, whose accuracy is 94.93%, recall rate is 90.34%, F ₁ value is 92.58%, and the average detection time of an image is 0.10 s.This paper provides a relatively comprehensive deep learning algorithm and model research basis for potato disease recognition and detection in the natural environment.