Deep neural networks (DNNs) have become the de facto standard for image recognition tasks, and their applications with respect to plant diseases have also obtained remarkable results. However, the large number of parameters and high computational complexities of these network models make them difficult to deploy on farms in remote areas. In this paper, focusing on the problems of resource constraints and plant diseases, we propose a DNN-based compression method. In order to reduce computational burden, this method uses lightweight fully connected layers to accelerate reasoning, pruning to remove redundant parameters and reduce multiply–accumulate operations, knowledge distillation instead of retraining to restore the lost accuracy, and then quantization to compress the size of the model further. After compressing the mainstream VGGNet and AlexNet models, the compressed versions are applied to the Plant Village dataset of plant disease images, and a performance comparison of the models before and after compression is obtained to verify the proposed method. The results show that the model can be compressed to 0.04 Mb with an accuracy of 97.09%. This experiment also proves the effectiveness of knowledge distillation during the pruning process, and compressed models are more efficient than prevalent lightweight models.

This paper proposes a data anomaly detection and correction algorithm for the tea plantation IoT system based on deep learning, aiming at the multi-cause and multi-feature characteristics of abnormal data. The algorithm is based on the Z-score standardization of the original data and the determination of sliding window size according to the sampling frequency. First, we construct a convolutional neural network (CNN) model to extract abnormal data. Second, based on the support vector machine (SVM) algorithm, the Gaussian radial basis function (RBF) and one-to-one (OVO) multiclassification method are used to classify the abnormal data. Then, after extracting the time points of abnormal data, a long short-term memory network is established for prediction with multifactor historical data. The predicted values are used to replace and correct the abnormal data. When multiple consecutive abnormal values are detected, a faulty sensor judgment is given, and the specific faulty sensor location is output. The results show that the accuracy rate and micro-specificity of abnormal data detection for the CNN-SVM model are 3–4% and 20–30% higher than those of the traditional CNN model, respectively. The anomaly detection and correction algorithm for tea plantation data established in this paper provides accurate performance.

Two-photon polymerization (TPP) based on laser direct writing is currently one of the most prevailing 3D micro/nano fabrication techniques. Nanomaterials can be doped in resins and assembled by TPP for developing advanced 3D functional devices. However, there lacks an effective visualization tool to determine the distribution and orientation of the nanomaterials as-doped in the composite resins. Herein, we present a nondestructive, in situ, and rapid characterization method to determine the orientation and distribution of the nanomaterials within cured resins using polarized second-harmonic generation (p-SHG). The directional assembly of the ZnO nanowires within micro/nanostructures fabricated by TPP is, for the first time to the best of our knowledge, characterized by p-SHG optical microscopy with a fast imaging speed by two orders of magnitude higher than that of the Raman mapping technique. Our method opens a window for nondestructive, rapid, in situ, and polarization-resolved characterization of functional devices made by TPP micro/nanofabrication.

Knowledge distillation, which intends to transfer the expertise from a complex teacher model to a concise student model, has achieved impressive success in object detection. However, many existing distillation methods are designed for object detection tasks on natural images and perform poorly on more challenging remote sensing images. In this study, we first attribute the deficiencies of knowledge distillation in remote sensing object detection to two core reasons: (1) lack of relation distillation among different instances and pixels, and (2) significant differences in feature magnitude between the teacher-student pair. Then, we propose a dual prediction-guided knowledge distillation framework, which includes relation distillation and output distillation to address two issues, respectively. Prediction-guided relation distillation is proposed to capture the knowledge of global relation at both instance-wise and pixel-wise¹, then allowing students to better understand and depict the features distribution across different categories. Prediction-guided output distillation is proposed to mitigate the impact of feature magnitude inconsistencies on distillation with classification and location knowledge, then allowing students to directly capture task-relevant information. Finally, experimental results have indicated the consistent effectiveness of our method across anchor-based two-stage, one-stage, and anchor-free detectors with eleven comparison knowledge distillation methods on three remote sensing detection datasets. Source codes are available at https://github.com/RQ-W/DPGD.git.

Knowledge distillation provides an effective paradigm for developing lightweight remote sensing object detectors. However, existing methods predominantly concentrate on the design of foreground feature masks while neglecting the suppression of small-scale targets by dominant large-scale counterparts in cluttered backgrounds during the distillation, which severely degrades detection robustness. To address this, we propose scale-conscious knowledge distillation. Multiscale feature distillation (MSFD) in this framework decouples the coupled features in conventional methods into hierarchical multiscale representation spaces, enabling the student model to capture cross-granularity features from micro-details to macro-structures through differentiated receptive fields. Scale-adaptive output distillation (SAOD) innovatively introduces a dynamic weighting mechanism based on the normalized object area, effectively alleviating the gradient vanishing issue for small objects. Experimental results show that our method consistently improves the performance of student models with different architectures on both the DOTA and DIOR datasets. Source codes are available at https://github.com/RQ-W/SCD.git