An Automated Pyramid Attention Driven Pest Detection System with Spatial and Texture-Aware Feature Learning

Abstract

Pest recognition is a significant feature in the current agricultural system since the level of insect infestation defines the health of crops, their level of yield and general productivity, particularly in vegetable production where plants are more vulnerable at any stage of their development. The identification of the pests must be done as early as possible and properly in order to minimize wastage of crop and unwarranted application of pesticides. Nevertheless, the current approaches of detecting pests have a number of limitations, such as the use of manual examination, the inability to represent features, their low generalization in the absence of different environmental conditions, and the inability to differentiate between similar symptoms of pests. The traditional and standalone approaches do not tend to reflect the complex spatial and textual division of the real field images. In order to overcome these constraints, the present paper presents a new feature extraction and classification model, SPARTA-Net (Spatial Pyramid Attention Representation Texture Aggregation) to effectively detect pests in vegetable crops. The proposed method is a mixture of multi-scale spatial pyramid analysis, feature refining by attention and texture-based representation to depict global and local pest features. The novelty of the method is that it has a hybrid feature aggregation strategy in which the spatial, structural and texture features are combined into a single feature, which improves strength and classification accuracy. Experimental results demonstrate that the suggested model works well with an accuracy of 98%, precision of 97%, recall of 97%, F1-score of 97% and AUC of 99% at a low loss. These findings are indicative of its capability to identify pests in precision agriculture with accuracy and in real-time.