Insect identification is the basis of insect research and disaster control and is of great importance for the design of pest control strategies and the protection of beneficial insects. Due to human subjective limitations and the small size and uneven distribution of pests, traditional methods of distinguishing and counting pest types based on experience cannot quickly and accurately detect and identify pests. Therefore, this paper proposes an object detection algorithm based on the improved Mask R‐CNN model, aiming to improve the accuracy and efficiency in pest identification and counting. The algorithm improves the FPN structure in the feature extraction network and increases the weight coefficient when fusing feature layers of different scales. Based on the task of target detection and recognition, weight coefficient is adjusted to a proper parameter so that the semantic information and positioning information can be made full use to achieve more accurate recognition and positioning. The results of the experimental analysis of 1000 sample images show that the improved Mask R‐CNN model has a recognition and detection accuracy of 99.4%, which is 2.7% higher than that of the unimproved Mask R‐CNN model. The main contribution of this method is to improve the detection speed, and at the same time, the recognition accuracy has been significantly improved. This algorithm provides technical support for pest detection in the agricultural field and makes a contribution to the intellectualization of agricultural management.

Objectives: The ERα biological activity prediction model is constructed by the compound molecular data of the anti-breast cancer therapeutic target ERα and its biological activity data, which improves the screening efficiency of anti-breast cancer drug candidates and saves the time and cost of drug development. Methods: In this paper, Ridge model is used to screen out molecular descriptors with a high degree of influence on the biological activity of Erα and divide datasets with different numbers of the molecular descriptors by screening results. Random Forest (RF) is trained by Root Mean Square Error (RMSE) and Coefficient of determination (R2) to determine the parameter range of RF optimized by Improved Sparrow Search Algorithm (ISSA-RF) which adds adaptive weights compared with the ordinary Sparrow Search Algorithm (SSA). Then the divided datasets were put into the ISSA-RF with defined parameter ranges to construct a regression prediction model for the biological activity of compounds on Erα, and compared with Genetic Algorithm Optimized Support Vector Machine (GA-SVM), Back Propagation Neural Network (BP), Extreme Gradient Boosting (XGBoost) for analysis and comparison. Results: We have tried a variety of combinations of molecular descriptors with different numbers and the above four models all achieve the best accuracy model on the dataset constructed when using 100 molecular descriptors. The ISSA-RF model proposed in this paper has a high degree of agreement between the predicted biological value of ERα and the actual value and prediction accuracy (RMSE) is 0.6876389. Conclusions: In the training model, ISSA-RF is proposed and it is proved that adding adaptive weights can greatly optimize the fitness accuracy of the sparrow algorithm. In the experimental part, this paper uses a variety of molecular descriptors for training, which reduces the chance of model training accuracy caused by the number of different molecular descriptors, and limits the search range of the ISSA-RF model to avoid the local optimization of the model. Secondly, the parameter optimization time is greatly reduced. In conclusion, the prediction model of drug compound molecules that antagonize ERα gene activity (ISSA-RF) proposed in this paper improves the accuracy and efficiency of anti-breast cancer drug candidates, and provides a new idea for building a quantitative structure-activity relationship model.