The pretrained models of these architectures were fine tuned with our proposed insect dataset so that they can be used for the yellow fly detection application, as fine tuning is also one of the common solutions for data scarcity problem in object detection. Because the models had been trained with COCO dataset, which is a large dataset having over 200,000 labeled images with 1.5 million object instances for 80 object categories, and, hence, contains common features for object detection problem, fine tuning the models with 200 yellow fly images helped the models perform the yellow fly detection task.To solve the real-time object detection in the yellow fly detection problem, variants of Single-Shot Multibox Detector are used. The SSD method was first proposed in [36] by Wei Liu et al. and described as a one stage object detection method that completely omits the region proposal and pixel/feature resampling stages used in region proposal-based techniques such as Faster-RCNN. The SSD network is based on a feed-forward network that uses default bounding boxes with different shapes, ratios, and scales to produce a fixed-size collection of bounding boxes with corresponding shape offsets and confidence scores. In addition, the early layers of the network are based on a standard image classification without classification layers, blueberry container size which is called the base network. In this work, MobileNetV1 and MobileNetV2 are used as base networks for the SSD detection models.
The elimination of region proposal and pixel/feature resampling stages helps to improve the processing speed of the model compared to two-stage techniques such as Faster-RCNN with a small trade-off in the model’s accuracy, which enables the implementation of real-time object detection with high accuracy on embedded system for yellow fly detection problem.The approach was first proposed in [37] by A. Howard et al. and was described as a lightweight deep neural network for mobile and embedded system applications with an efficient trade-off between latency and accuracy. The model is based on depth wise separable convolution including depth wise convolution layer which is used to apply a single filter per input channel, and pointwise convolution layer, which creates a linear combination of the output of the depth wise layer. In addition, to construct the model further less computationally expensive, width multiplier, which is used to thin the network uniformly at each layer, and resolution multiplier, which is applied to input images and the internal representation of each layer, were introduced as a hyperparameter to tune, and choose the size of the model.The MobileNetV2 approach was first presented in by M. Sandler et al. The approach is built based on MobileNetV1; therefore, it also makes use of the depth wise separable convolution architecture which consists of depth wise convolution layer and 1×1 point wise convolution layer. In addition, the approach also utilizes linear bottleneck layers in convolutional blocks to optimize the neural architecture. Moreover, inverted residual design is also used in the model to implement shortcuts between bottlenecks with the purpose of improving the ability of gradient propagation across the multiplier layers. Nevertheless, the implementation of the inverted design also showed better performance and significantly more memory efficiency in the work. The training and evaluation of the SSD with MobileNetV1 and MobileNetV2 base networks is based on the pre-trained models provided in the Object Detection API in TensorFlow Model Garden.
The models were also trained on the Google Colab Pro environment to utilize the provided GPUs for the training purpose.The assessments in this research are dedicated to search for the most appropriate object detection method among the current state-of-the-art algorithms which have been implemented for insect and fly recognition under our hardware constraints and problem definition. As we only target one type of fruit flies that particularly causes harm to the citrus fruits, we have replaced the yellow sticky paper with a white disc containing the special attractant as a hard refinement to pick up only the flies we are interested in. The object detection problem is then simplified to only one-class object detection, which eases the need for exhausting feature extraction. However, the general constraints, such as correctness and fastness, for an object detection task on an edge-device still hold since early detection and separation of the infected areas are extremely important to the fruit yield.Ultimately, SSD-MobileNetV1, SSD-MobileNetV2 and YOLOv4-tiny are the best candidates for these requirements because they utilize extracted features from a backbone classification model to automatically propose object-related regions instead of using a region-proposal module to pool the related regions before classifying them as many two stage object detection models, such as Fast-RCNN and Faster-RCNN.Regarding the correctness, YOLOv4-tiny clearly outperforms the two SSD models over all the evaluations on four different types of testset with very high and stable results. This could make YOLOv4-tiny become the most probable candidate, because YOLOv4-tiny demonstrates a robust testing performance towards citrus fruit fly detection although it has been fine-tuned only on a training dataset without augmentation effects. SSD-MobileNetV2 shows appropriate robustness given its small number of trainable parameters by yielding good results in two over four testsets, while SSD-MobileNetV1 only works with the original testset. Nevertheless, SSD-MobileNetv2 fails dramatically with the Blurry testset, which simulates a very frequent event that could happen in a fruit field. YOLOv4-tiny is no doubt the chosen one among the three methods if we would not have taken other aspects into account.
Conventionally, highly accurate object detection methods trade their processing speed for its better performance due to the employment of more parameters in their architecture. While missing a fraction of time could lead to undetectable events in which the flies appear, our second choice, which is the SSD-MobileNetV2 model, should be considered. To realize this choice after extensive performance analysis with four different testsets, SSD-MobileNetV2 must have been fine tuned with more augmented versions of the original training dataset before going to production to leverage its robustness to the level of YOLOv4-tiny while retaining its processing speed. Moreover, TFLITE version of SSD-models are also tested on a cloud TPU Google engine, TPUv2, for the feasibility of edge-device deployment. The overall assessment table for YOLOv4-tiny and SSD-MobileNetV2 is shown in Table 2 in terms of F1-Score and inference time.While San Joaquin Valley vineyards are currently fertilized with boron through the soil and foliage , some growers have expressed interest in applying boron via drip irrigation or “fertigation.” Fertigation is a relatively simple, cost-effective and efficient way to apply nutrients. However, irrigation water with more than 1 part per million boron can lead to vine toxicity, so the safety of boron fertigation is also a concern. Our research evaluates the safety and efficacy of boron fertigation in grapevines using drip irrigation. Boron is unique among the micronutrients due to the narrow range between deficiency and toxicity in soil and plant tissues. For grapevines, this range is 0.15 ppm to 1 ppm in saturated soil extracts, and 30 ppm to 80 ppm in leaf tissue. The goal of boron fertilization of grapevines is to keep tissue levels within this narrow range, since both deficiency and toxicity can have serious negative effects on vine growth and production. Fertilization amounts must be precise to avoid toxicity while providing adequate boron to satisfy grapevine requirements . On the east side of the San Joaquin Valley, boron deficiency of grapevines occurs on soils formed from igneous rocks of the Sierra Nevada. This parent material is low in total boron, growing raspberries in containers which is crystallized in borosilicate minerals that are highly resistant to weathering. Boron deficiency is often associated with sandy soils and vineyard areas with excessive leaching, such as in low spots or near leaky irrigation valves. Vine symptoms of boron deficiency are more widespread and pronounced following high rainfall years, when greater amounts of soluble boron are leached from the root zone. In addition, snowmelt water has very low levels of boron, and vineyards irrigated primarily with this water have a greater risk of deficiency. Boron is required for the germination and growth of pollen during flowering, and vines that are deficient in this micronutrient will have clusters that set numerous shot berries, small berries with a distinctive pumpkin shape. When boron deficiency is severe, vines produce almost no crop. Foliar symptoms appear in the spring: shoots have shortened, swollen internodes and their tips sometimes die, and leaves have irregular, yellowish mottling between the veins. Grapevines are also sensitive to too much boron. Toxicity is common on the west side of the San Joaquin Valley, where most soils are derived from marine sedimentary and metasedimentary parent material that is rich in easily weathered boron minerals. Symptoms of boron toxicity include leaves that are cupped downward in the spring and that develop brown spots adjacent to the leaf margin in middle or late summer, intensifying and leading to necrosis as boron accumulates.
Yields are reduced, the result of diminished vine vigor and canopy development. When foliar boron sprays are applied in excess in the spring, juvenile leaves become cupped within 2 weeks; however, vines quickly recover and yields are usually unaffected. Toxicity also occurs when boron fertilizer is applied in excess, regardless of the soil type, and this can lead to yield loss. Over-fertilization is the sole reason for boron toxicity on the east side of the San Joaquin Valley, so it is critical to establish how much boron fertilizer can be applied safely and effectively. Our research investigated the uptake of boron by grapevines when fertilizer was applied with a drip-irrigation system.Research was conducted from 1998 to 2001 in a mature ‘Thompson Seedless’ raisin vineyard near Woodlake in Tulare County. The vineyard was planted in Cajon sandy loam on a recent alluvial fan associated with the Kaweah River. This soil is derived from granitic parent materials, and the surface soil is highly micaceous with a slight to moderate amount of lime. The underlying soil has a coarse, sandy texture. At the onset of this study, the vineyard’s boron status was in the questionable range for deficiency. The vine’s leaf petioles and blades contained about 30 ppm boron. While the foliage had no symptoms of boron deficiency, in the past the grower had observed sticking caps and pumpkin-shaped shot berries, which are indicative of boron deficiency. During the course of the research, the vineyard was drip-irrigated from April through October. The vineyard canopy covered 60% of the land surface during summer months and about 20 inches of water was applied during the season. Boron treatments consisted of applying fertilizer in varying amounts 3 weeks prior to bloom on May 18, 1998, and then again 3 weeks prior to bloom the following year, on May 3, 1999. Growers who fertigate grapevines with a drip system generally inject material into the irrigation water over a 45-to- 60-minute period at the beginning of an irrigation set. We simulated fertigation by applying Solubor, a soluble boron product , to a shovel-sized hole beneath drippers during the first hour of the irrigation set. By doing this, precise amounts of boron could be applied to each plot and plot size could be reduced. This technique has been used successfully in previous research with other nutrients . The experiment was designed as a randomized complete block with five treatments, five blocks and five vine plots . To evaluate the rate of boron uptake and accumulation in tissue with consecutive years of fertilization, grape tissue samples were collected in 1998 and 1999 at bloom and then again about 6 weeks later during veraison. Veraison is the stage of development where berries begin to soften and/or color. To evaluate carryover, leaf tissue samples were also collected at this Tulare County site at both bloom and veraison in 2001, 2 years after the fertilization was discontinued. In each case, 100 petioles and 50 blades were sampled per plot from the center three vines. Petioles and blades were taken opposite inflorescences during bloom, and recently matured leaves were sampled at veraison. Samples were oven-dried, ground in a Wiley mill and sent to the UC Davis DANR Analytical Laboratory for analysis of total boron. Statistical analysis was by ANOVA using least significance difference to separate treatment means. A second experiment was conducted in 1998 in Fresno County near Selma, in a mature Thompson Seedless raisin vineyard planted on Pollaski sandy loam and drip-irrigated.