Objectives The compost maturity is mainly judged by complex chemical and biological experiments, which is difficult to operate and inefficient. Convolutional neural networks simulate human vision, which can retain the color information of compost images, and extract representative features such as contour, lines and granularity at the same time, thus avoiding the influence of different illumination conditions on the prediction effect of compost maturity. This paper proposed and verified a prediction model combining images of compost and convolution neural network.

Methods Composting samples were collected from Jiangsu, Shandong and Zhejiang provinces, and the composting materials in three provinces were straw, cauliflower residues and livestock manure, respectively, and the composting cycle was 50 days, 45 days and 60 days in turn. In the factory shed, HIKVISION video camera (model C3W) was used to take composting images of different maturing stages in JPEG format under automatic light and compensation at night. The used focal length was 2.8 mm, resolution was 1080 p, and the distance was 1 m from the compost surface. Except for the image data sets from the three composts, the fourth image data set was taken in the mixture of them three in ratio of 1∶1∶1. 80% of the images from each data set was used to train CNN model and to establish prediction model parameters, and the remaining 20% to testify the prediction accuracy of the model.

Results The compost maturity prediction model was composed of one input layer, three convolution layers, three pool layers, two full connection layers and one output layer. The accuracy of the predicted maturity from the compost images was averaged 98.7%, 98.7%, 98.8% and 98.2% for the vegetables residues, straws, livestock manure and the mix of above, respectively. Comparing with the most optimal result of classical algorithm on each data set, the average accuracy of this method in image feature extraction and classification were improved by 3 to 14 percentage points. Texture feature was more effective than color feature in judging compost maturity by CNN method.

Conclusions As the priority of convolutional neural network in extracting the appearance features of compost images, the compost maturity prediction model based on it could accurately and rapidly identify the compost maturity directly through compost images under natural light conditions.