Detection Of Agricultural Pest Insects Based On Imaging And Spectral Feature Analysis

Insect pests are one of the main factors causing frequent agricultural disasters.To improve the crop production and protection ability and to apply timely targeted pesticides capable of reducing input costs and benefiting the environment,an accurate early-stage approach on detection of pests and quantification of damages caused by pests is required.Most current pest detection approaches are costly and time-consuming as they require agricultural professionals to manually collect and classify specimens in the field,thus difficult to meet the requirement of lean production in modern agriculture.More inexpensive methods are therefore required,and automated systems using the spectroscopy detection technique and computer vision technique are investigated in this paper.The main contributions in this paper are as follows:(1)Since it is difficult to directly detect the larvae of pests that causes damages within crops,healthy and infected crop stems are both studied under help of the hyperspectral imaging technique.Based on features of high dimension and redundancy of hyperspectral imaging data,a stacked sparse auto-encoder(SSAE)is constructed for the effective representation of sparsity information at different infestation stages by pests' larvae.Combined with SSAE,we proposed a Softmax function-based algorithm,called adaptive loss-sensitive adaptive algorithm,to further enhance the performance of our approach on distinguishing normal and early infected samples.We also defined an index,called Relative Scatter Value(RSV),to quantify the differences about distribution between-classes and within-class of spectral data before and after introducing adaptive loss-sensitive adaptive algorithm.The results show that the excellent linear separation was introduced to the data by this approach.(2)The visual localization of adult pests in paddies was studied by adopting the computer vision technique,and we propose an automated localization method that combines image saliency analysis and GrabCut segmentation.Based on differences about overall color and space between pests-targeted regions and paddies,a saliency map is constructed capable of indicating potential regions of pests.The initial regions,with GrabCut algorithm offered by the thresholding Saliency Map,achieves the automatic positioning of pest targets.Besides,the method was further optimized in terms of localization accuracy and running time,and after optimization the localization accuracy rises to 0.9 above with running one image less than 150 ms.(3)The visual recognition of adult pests is investigated by using convolutional neural network(CNN),and it is optimized based on critical parameters and training strategies of CNN.Various types of low-level feature extraction,middle-level feature scale,high-level information organization and classifiers are tested for seeking the best configuration for the agricultural pest recognition task,directed at features of significant intra-class differences,large inter-species similarity,wide position variations and frequent occlusion by pest targets.In this test,the paddy images including 14 species of typical pests with achieving the recognition accuracy of 0.883,the memory requirement of 6.0 MB and the running time of 0.7 ms,significantly improving the availability for practical applications.(4)Since the visual recognition models are usually subject to the small scale and imbalanced distribution of pest image data set,the feature transferring strategy is introduced to improve the performance of visual recognition models.The CNN is selected as a baseline framework.The image feature pre-trained on a large-scale image data set is transferred to our pest recognition tasks and is set as initial parameters of CNNs.And,the results show a significant improvement on relatively conventional random initialization methods.By designing control experiments in which CNN features are transferred layer-by-layer with fixed or adjustable parameters,the transferring strategy that CNN low-level features can be transferred directly,successive middle-level features should not be transferred separately and its high-level features are not suitable to be transferred is verified.By randomly changing the size and distribution of training data set,the relative insensitivity to the data scale and class imbalance of the feature transferring is proved.By visualizing the model training curves,low-level features,high-level representations and represented data distributions,essential reasons that cause the advantages of feature transferring over random initialization are stated.