In order to quantitatively analyze the light radiation environment inside Chinese solar greenhouse (CSG) and select reasonable building design parameters, a CSG solar radiation environment model reflecting various factors such as geographical location, outside solar radiation, orientation and building parameters, front roof shape, and covering materials was studied. The model considered the impact of both cloudy and sunny weather conditions on the inside solar radiation environment, and established a simulation calculation method for inside direct radiation and scattered radiation. When calculating solar scattered radiation, the ground reflected radiation and atmospheric longwave radiation were considered. When calculating the transmittance of covering material, a structural shading loss and dust film model was introduced to calculate its impact on the transmittance. The model was validated experimentally in a CSG at Yongqing in Hebei Province, China. The results showed that the model can effectively simulate the solar radiation of various points, such as the ground and wall in the greenhouse, at any time, with an average relative error of 8.19% between the simulated and measured values. Based on the established model, the impact of the geographical location, azimuth angle, and building parameters of CSG on inside solar radiation was analyzed. The research results can provide theoretical references and relevant data for the wall and soil heat storage, crop planting, and energy balance of enclosure structures in CSG.

An accurate prediction model of greenhouse air temperature can be required for the greenhouse environment control using control algorithm, such as model predictive control. Long short-term memory neural networks (LSTM) have been widely used to predict time series data, such as air temperature. However, the prediction accuracy of LSTM can be reduced for the long time series data, due to data forgetting. In this study, the LSTM model was combined with the attention mechanism to construct the LSTM-AT model. The query vector, key vector, and value vector were calculated, according to attention mechanism and output states of LSTM’s hidden layer. The similarity between the query and key vector was calculated to obtain the similarity score. Softmax function was used to obtain attention distribution for the normalization processing. The larger the attention value was, the higher the relevance of input information to the task objective was. The dot product operation was carried out with the normalized weight and value vector to obtain the output of the attention mechanism. The local information integration and data dimension transformation were carried out through the full connection layer. Finally, the output data was obtained in the output layer of the LSTM-AT model. The weights were assigned to the output states of LSTM’s hidden layer, according to the degree of importance. The forgetting of long time series data was effectively solved to improve the prediction accuracy of indoor air temperature. The prediction performances were verified and compared on the LSTM-AT, LSTM, recurrent neural network (RNN), gated recurrent unit (GRU), and bi-directional long short-term memory (BiLSTM) model in the different prediction horizons (12, 24 and 48 h). The results showed that the prediction accuracy of the five models shared a decreasing trend with the increase in prediction time. The maximum and minimum RMSE for the LSTM-AT model were 1.34 and 0.59 ℃, respectively. The maximum and minimum RMSE for the rest four models were 3.37 and 0.66 ℃, respectively. The maximum and minimum MAPE for the LSTM-AT model were 8.14% and 2.48%, respectively. The maximum and minimum MAPE for the rest four models were 38.7% and 2.90%, respectively. The average RMSE for LSTM-AT, LSTM, GRU, RNN, and BiLSTM were 0.89, 1.42, 1.89, 2.10, and 1.51 ℃, respectively. The average MAPE for LSTM-AT, LSTM, GRU, RNN, and BiLSTM were 4.26%, 8.96%, 13.57%, 17.70%, and 10.67%, respectively. The sort data of the prediction model was ranked in descending order of the LSTM-AT, LSTM, BiLSTM, GRU, and RNN. Therefore, the prediction accuracy of the LSTM-AT model was higher than that of rest four models. The prediction performances of the LSTM-AT and LSTM model were compared under different weather conditions (sunny, cloudy, and rainy), in order to further explore the universality of the LSTM-AT model. The minimum and maximum RMSE for LSTM-AT were 0.26 and 0.70 ℃, respectively. The minimum and maximum RMSE for LSTM were 0.68 and 1.57 ℃, respectively. The minimum and maximum MAPE for LSTM-AT were 1.61% and 10.51%, respectively. The minimum and maximum MAPE for LSTM were 4.27% and 25.07%, respectively. The prediction accuracy of the LSTM-AT model was higher than LSTM in all weather conditions. The LSTM-AT model has a higher prediction accuracy to accurately predict the indoor air temperature.

Chrysanthemum tea has been one of the most popular food products, due to the health benefits and high commercial value. The medicinal and economic chrysanthemum can greatly vary in the different flowering stages. Among them, the flowering stages of chrysanthemum can be categorized into the Jumi (flower buds), Taiju (flower buds just before blooming), and Duohua (fully bloomed flowers). At the same time, the chrysanthemum is required for the best picking time. However, manual picking cannot fully meet the requirement of large-scale production at the early flowering stage, due mainly to the labor-intensive and time-consuming. Untimely picking or picking errors at the different flowering stages can also lead to the waste of chrysanthemum, even the serious economic losses. Therefore, the picking robot can be expected to realize the accurate and rapid recognition of chrysanthemum in different flowering stages using lightweight model. In this study, an improved YOLOv8s model (YOLOv8s-RDL) was proposed for the object detection of chrysanthemum. Firstly, the C2f (faster implementation of CSP bottleneck with 2 convolutions) in Neck network was replaced by RCS-OSA (one-shot aggregation of reparameterized convolution using channel shuffle). The features were extracted to more efficiently fuse by the Backbone layer; Secondly, the decoupled head was replaced with the Dyhead (dynamic head), and then integrated into the DCNv3 (deformable convolutional networks v3). The multi-head self-attention mechanism was combined to strengthen the expression of the target detection head. Finally, the LAMP (layer-adaptive magnitude-based pruning) was used to reduce the number of parameters and the complexity of the model network. The amount of calculation was significantly reduced to maintain a high mean average precision level of the improved model. A comparison was also made to explore the effect of RCS-OSA in the different positions of the network. The performance of the model was depended mainly on the different pruning conditions and rates in the same pruning direction. The best improved model was obtained after network structure improvement and pruning. The best solution was attributed to the replacement of the C2f with the RCS-OSA only in the Neck part, particularly in the pruning condition of the adaptation model. The ablation experiments show that the mean average accuracies of the improved model for Duohua, Taiju and Jumi were 99.0%, 97.7% and 96.3%, respectively, in the classification and detection of flowering stage, which were 0.3, 1.9 and 3.8 percentage points higher than that of YOLOv8s. The mean average precision, precision and recall of the YOLOv8s-RDL for the Jumi, Taiju, and Duohua were 97.7%, 96.5%, and 95.2%, respectively, which were 1.9, 5.2 and 6.4 percentage points higher than YOLOv8s baseline model, respectively. The size of model weight also decreased by 6 MB. The improved model was greatly reduced the number of parameters and weight size, indicating the high detection accuracy. The mean average accuracies of YOLOv8S-RDL were 35.3, 2.9, 3.4, 1.9, 3.8 and 1.7 percentage points higher than those of SSD, YOLOv5s, YOLOv6s, YOLOv8s, Ginger-YOLOv5s and MSC-YOLOv8, respectively. At the same time, there were the smallest weight size and parameters of the improved model. In addition, the superiority of the algorithm was verified to introduce the TIDE (toolkit for identifying detection and segmentation errors) indicator. The detection errors of classification and background were reduced by 0.55 and 1.26, respectively, in the YOLOv8s-RDL, compared with the YOLOv8s. The better performance of detection and classification was achieved to reduce the influence of background and interference factors. The improved model was also fully met the requirements of accurately and rapidly distinguish the chrysanthemum flowering stages. This finding can also provide the theoretical reference and technical support to realize the automatic picking of chrysanthemum in the various flowering period.