Ground straw mulching level classification based on a terral grid system and deep learning

•A terral grid system is developed to obtain local sliced images at grid intersections.•AC-MobileNet is proposed to obtain the truth of straw mulching levels.•ResNet101 network is well-trained for straw mulching classification automatically.•Quick and accurate detection of straw mulching levels is achieved instead of human inspection. The detection of ground straw mulching levels plays a crucial role in implementing conservation tillage efficiently. To address the need for quick and accurate determination of straw mulching levels, a rapid detection method based on deep learning was proposed, which consists of a terral grid system, semi-automatic straw mulching grading based on ASPP-CBAM MobileNet (AC-MobileNet), and a deep residual network ResNet101. The terral grid system was applied to obtain local sliced images at the grid intersections from straw mulching images. Then, the AC-MobileNet model was used to acquire the type of local sliced images automatically, making it obtain the true value of each straw mulching level and establish the dataset quickly. Subsequently, the deep residual network ResNet101 was utilized for straw mulching level detection. The test results showed that the AC-MobileNet model achieved a sliced image classification accuracy of 96.3%, surpassing the MobileNetv3 network model by 2.1%. In comparison with classification networks such as AlexNet, ShuffleNetv1, and EfficientNetv2, the AC-MobileNet model exhibited the highest accuracy. The accuracy of ResNet101 in straw mulching level classification was 98.3%, outperforming DenseNet161, EfficientNetv2, and VGG16 networks. The proposed method demonstrated a detection speed of 750 times higher than that of manual detection in the field, keeping the straw mulching level classification results aligned with the manual results. The developed methodology paves the road for accurate and rapid detection of ground straw mulching levels.