Here we show that combining Iso-Seq with Illumina sequencing at high coverage enables expression profiling and sequence error correction of IsoSeq reads, particularly those derived from low-expression genes. The clustering analysis of the SMRT link pipeline discarded  18.5% of the FLNC reads, likely caused by low sequence accuracy. To overcome this technical issue, we applied a hybrid error correction pipeline consisting in performing the error correction of the unclustered FLNC reads, followed by an additional clustering step of both to resolve redundancies. Error correction with Illumina reads recovered a significant amount of Iso-Seq reads that would have otherwise been removed by the standard Iso-Seq pipeline, highlighting the importance of integrating multiple sequencing technologies with complementary features . Transcriptome reconstruction has been widely used to develop references for genome-wide expression profiling in the absence of an annotated genome assembly . Though a genome reference is available for grape, transcriptome reconstruction overcomes the limitations of a cultivar-specific reference that lacks the gene content of other cultivars. Although cultivar-specific genes appear nonessential for berry development, those private genes could contribute to cultivar characteristics. For example, the wine grape Tannat accumulates unusually high levels of polyphenols in the berry; its cultivar specific genes account for more than 80% of the expression of phenolic and polyphenolic compound biosynthetic enzymes . De novo transcriptome assembly from short RNA-seq reads has been used to explore the gene content diversity in Tannat , Corvina , and Nebbiolo . Iso-Seq identified 1,501 Cabernet Sauvignon transcripts expressed during berry development that were found in neither the genome of PN40024 nor the transcriptomes of Tannat, Nebbiolo and Corvina. Some private Cabernet Sauvignon transcripts have functions potentially associated with traits characteristic of Cabernet Sauvignon grapes and wines like their color and sugar content.

These transcripts included three sugar transporter-coding genes, big plastic pots which could be involved in the accumulation of glucose and fructose during berry ripening , and a chalcone synthase, a flavanone 3-hydroxylase, and a flavonoid 39-hydroxylase, all involved in the flavonoid pathway. Chalcone synthases catalyze the first committed step of the flavonoid biosynthesis pathway , which produces different classes of metabolites in grape berry, including flavonols , flavan-3-ols and proanthocyanidins , and anthocyanins . In addition, products of the flavonoid 39-hydroxylase can lead to the synthesis of cyanidin-3-glucoside, a red anthocyanin . The analysis of the gene space in the genome assembly showed that private Cabernet Sauvignon genes identified using Iso-Seq are only a fraction of the private Cabernet Sauvignon transcriptome. As in other transcriptome reconstruction methods, Iso-Seq can only identify transcripts that are expressed in the organs and developmental stages used for RNA sequencing. Obtaining the full set of private transcripts without genome assembly would require sequencing additional organs and developmental stages. In addition, it is challenging to differentiate isoforms derived from close paralogous genes, alleles of the same gene, and alternative splicing variants, in any transcriptome obtained by RNA sequencing ; this potentially leads to an overestimation of the genes in the final transcriptome reference. This study could not resolve isoform redundancy in the final transcriptome for about 37% of the gene loci in the Cabernet Sauvignon genome. This is a limitation of Iso-Seq as well as of all transcriptome references that cannot be overcome without a complete genome assembly. In this study, we tested whether the transcriptome reconstructed using Iso-Seq can be used for expression profiling. Only an approximately 3% difference in read alignment between ISNT and the genome reference was observed, implying that at high coverage, ISNT detects almost all genes expressed during berry development.

The slight difference in mapping rate between the two references can be explained by either the absence of some low-expression transcripts in the ISNT or the residual error rate in isoform sequences. Gene expression analysis using the ISNT as reference showed similar results compared to the Cabernet Sauvignon genome assembly, with a very high correlation of expression level and differential gene expression, and with similar global transcriptomic changes. However, we observed differences in the number of expressed and differentially expressed features that depend on the reference used. Those differences could be explained by the diploid phasing of the Cabernet Sauvignon genome assembly and that multiple ISNT transcripts might correspond to a single gene locus. Nonetheless, similar relative amounts of Biological Process GO terms were found among the differentially expressed genes, confirming that the transcriptome obtained using Iso-Seq captured the transcriptional reprogramming underlying the main physiological and biochemical changes during grape berry development. In addition, gene expression analysis revealed that some private isoforms are significantly modulated during berry development, indicating that in addition to identifying the private gene space, the ISNT reference makesit possible to observe its expression. In conclusion, this study demonstrates that Iso-Seq data can be used to create and refine a comprehensive reference transcriptome that represents most genes expressed in a tissue undergoing extensive transcriptional reprogramming during development. In grapes, this approach can aid developing transcriptome references and is particularly valuable given diverse cultivars with private transcripts and accessions that are genetically distant from available genome references, like the non-vinifera Vitis species used as rootstocks or for breeding. The pipeline described here can be useful in efforts to reconstruct the gene space in plant species with large and complex genomes still unresolved.

Agriculture plays an important role in economic growth, and improving crop yield is of great concern in Vietnam. On the one hand, insect pesticides can affect the metabolic processes of crops to degrade crop yield and quality. On the other hand, fruit flies are known to cause 50 to 100% crop loss unless timely interventions are implemented. There are just a small number of fruit fly species that have been discovered, namely Bactrocera dorsalis, B. correcta, B. cucurbitae, B. tau, B. latifrons, B. zonata, B. tuberculata, B. moroides and B. albistriga, while some species remain unidentified. The species which are harmful to fruits are of the common fruit fly species, namely B. cucurbitae and B. tau. To optimize crop yields, agricultural workers tend to use a pesticide scheduler rather than consider the likelihood of pests’ presence in the crop. Thus, this not only causes many pesticide residues in agricultural commodities but also brings great pressure to the ecological environment. The overuse of pesticides is partly because information about pest species and densities cannot be provided in a timely and accurate way. In contrast, if the information is provided in atimely fashion, it could be possible to take proper prevention steps and adopt suitable pest management strategies including the rational use of pesticides . Traditionally, the information about the environment and pest species is acquired mainly through handcrafted feature engineering such that workers manually use sensors and compare a pest’s shape, color, texture, and other characteristics with justification from the domain experts. Likewise, counting is typically time-consuming, labor intensive, and error-prone. Therefore, it is urgent and significant to establish an autonomous and accurate pest identification system. There is a growing tendency of utilizing machine vision technology to solve these problems with promising performance in the agriculture research field. In this work, growing berries in containers we focus on developing a solution to detect oriental yellow flies which usually harm citrus fruits such as oranges and grapefruits. We implement and evaluate the object detection models by applying the models with test sets simulating potential disturbances occurring in real scenario. Additionally, the work presented in this paper will not only focus on the use of different types of object detection algorithms but also apply the TFLITE format of the models compatible to edge device system such as TPU processors. This direction of study is to develop real-time detection application with the emerging edge computing technology to enhance the performance of the system in terms of detection accuracy, power efficiency, and latency reduction with the purpose of detecting the living fruit flies beside the stuck and dead ones on the trap. Moreover, the article will describe the hardware implementation so that the work can be reproduced and further developed. Our contributions are: We constructed, developed, and provided a more in-depth discussion of the end-to-end camera-equipped trap, named AlertTrap with installation of a Lynfield-inspired sticky trap, to instantly detect fruit flies and the solar-energy powering system controlled by a separate Raspberry Pi. We evaluate three different compact and fast object detection deep learning models, namely SSDMobileNetV1, SSD-MobileNetV2, and the Yolov4-tiny. Nevertheless, we introduce artificial disturbances imitating inference effects which may compromise the detection performance in real-time scenario. Moreover, we also evaluate the SSD-MobileNetV1 and SSD-MobileNetV2 models with their TFLITE format versions on a TPU device. With the results, we compare not only their ability to accurately detect and localize the fruit flies which we had trained them to predict, but also the increase in processing speed as well as the power saving factor.Insect detection techniques can be classified into three system types, namely manual, automatic, or semi-automatic systems.

Manual insect detection techniques are known as a process in which trained workers count the trapped flies on a daily basis. These turn out to be error-prone, time consuming, and labor-intensive, while semi-automatic and automatic systems can address the disadvantages with the replacement of highly accurate and autonomous emerging technological software and hardware. Specifically, the remaining two types of insect detection systems are often called e-traps as they are fueled by electronic components with extensive computer algorithms such as a center-controlled unit connecting with a camera and the trap actuators. Thus, they are also known as vision based insect traps. As suggested in the names, the automatic insect detection systems are fully autonomous, whereas the semi-automatic ones involve human interaction in the loop. For example, in [24], the images of insect body parts are classified to aid humans to better categorize the insects. Generally, the e-traps are equipped with a wide range of post-processing techniques to detect and classify trapped insects. These techniques are recognized by the sensor type that is used to capture the existence of insects in the trap. Particularly, they are image based, spectroscopy-based, and optoacoustic techniques, which correspond respectively to the visible-light camera, the near-infrared camera, and the ultrasound sensor. The image-based techniques consist of three sub-domain techniques, namely deep learning and shallow learning, which both are sub-domains in the machine learning field, and image processing techniques. Shallow learning-wise, Kaya et al. created a machine learning-based classifier that can differentiate between 14 butterfly species. The texture and color characteristics are extracted by the writers. A three-layer neural network is used to process the extracted features. The categorization accuracy achieved is 92.85 percent. The detection approach is based on image processing as described in. While image-processing techniques are simpler than deep learning techniques, their accuracy is reasonable and the system is wired with the illumination environment. However, extensive feature engineering must take place prior to the classification. By using auto-brightness adjustment, the algorithm first reduces the effect of changing lighting and weather conditions. Then, using a coordinate logic filter improves the edges by amplifying the difference between the dark bug and the bright background. Finally, the technique uses a circular Hough transform followed by a noise reduction filter to identify the trap’s limits. The achieved accuracy rate is 75%. In [14], it was reported that a Wireless Sensor Network was created for detecting pests in greenhouses. The image processing technique first removes the effect of light changes from the photos, then denoises them, and finally recognizes the blobs. In [15], it was suggested that insect image processing, segmentation, and sorting algorithms could be used as insect “soup” images. In insect “soup” photos, the insects float on the liquid surface. The method was evaluated on 19 soup images by the authors, and it worked well for many of them. Using McPhail traps, a WSN was developed to detect the olive fruit fly and medfly in the field. WSNs are sensor networks that gather data and may be built to process information and transfer it to humans. WSNs may also have actuators that respond to specific events. The template comparison algorithm is the detection algorithm. The identification is based on the detection of specific anatomical, patterning, and color characteristics. Near Infrared Spectroscopy was used to identify infested olives in harvested crops. The Genetic Algorithm extracts the features from the collected full spectral data. The retrieved features serve as the input for the classifier.