RiPa-Net: Recognition of Rice Paddy Diseases with Duo-Layers of CNNs Fostered by Feature Transformation and Selection

Rice paddy diseases significantly reduce the quantity and quality of crops, so it is essential to recognize them quickly and accurately for prevention and control. Deep learning (DL)-based computer-assisted expert systems are encouraging approaches to solving this issue and dealing with the dearth of subject-matter specialists in this area. Nonetheless, a major generalization obstacle is posed by the existence of small discrepancies between various classes of paddy diseases. Numerous studies have used features taken from a single deep layer of an individual complex DL construction with many deep layers and parameters. All of them have relied on spatial knowledge only to learn their recognition models trained with a large number of features. This study suggests a pipeline called “RiPa-Net” based on three lightweight CNNs that can identify and categorize nine paddy diseases as well as healthy paddy. The suggested pipeline gathers features from two different layers of each of the CNNs. Moreover, the suggested method additionally applies the dual-tree complex wavelet transform (DTCWT) to the deep features of the first layer to obtain spectral–temporal information. Additionally, it incorporates the deep features of the first layer of the three CNNs using principal component analysis (PCA) and discrete cosine transform (DCT) transformation methods, which reduce the dimension of the first layer features. The second layer’s spatial deep features are then combined with these fused time-frequency deep features. After that, a feature selection process is introduced to reduce the size of the feature vector and choose only those features that have a significant impact on the recognition process, thereby further reducing recognition complexity. According to the results, combining deep features from two layers of different lightweight CNNs can improve recognition accuracy. Performance also improves as a result of the acquired spatial–spectral–temporal information used to learn models. Using 300 features, the cubic support vector machine (SVM) achieves an outstanding accuracy of 97.5%. The competitive ability of the suggested pipeline is confirmed by a comparison of the experimental results with findings from previously conducted research on the recognition of paddy diseases.