Wireless multimedia sensor network for rape disease detections

Agriculture is now changing from modern farming to modern agriculture. Agricultural marketing (IOT) will play a major role in promoting agricultural farming, including agricultural infrastructure development, the development of modern agricultural technology and agricultural use. The use of modern technology, computer technology, and computers is a computer-simultaneous work of coordinating the development of smart farming and productive farming [1].

Crop rotation as a major activity for agricultural material is of great importance [2]. Because of climate change, it is not only necessary to do a thorough investigation and research on the growth of the crop, as well as to understand the small crop yields. The growth of the crop and the growth of crop yields provides scientific and agricultural support. In order to maximize the profitability of farming, various species of biodiversity and biodiversity were established. The crop rotation system also confirms two types of information and re-engineering. The room for environmental creation produces heat and soil, such as heat, humidity, wind, wind, rain, pH prices and so on [3]. Panel drawing photo gets a picture of the size of the plant. The size and size of the crop can be shown directly. Many numbers make up a moth-free area, and they can use the Internet.

The whole story is being prepared in this way. The second phase creates the design of the hardware platform. The third stage, the provision of scalar sensor node is asked. Then in the fourth stage, the formation of a photo snapshot picture is described. The fifth stage, the program is launched. The sixth stage produces test results. Finally, summarize this pamphlet.

Tomato trees are protected against many diseases and damage that cause disease and pests. Several reasons may be due to the harvest: (I) abiotic complications due to nature, such as heat, humidity, nutrients, fertilizers, bright colors and plants; (2) pathogens from plants to plants such as whitefly, vegetable gardens, worms, and so on. (3) Many diseases, including bacteria, viruses, and fungi. The disease is a pest, as well as a plant that has a variety of shapes, such as shape, color, shape, etc. Therefore, because of similarities, these changes are difficult to distinguish, which makes their awareness difficult, recognition and preparation can prevent many of the damage. Based on the information we have discussed, we consider the following characteristics:

Figure 4 showed the characteristic of different conditions and varying diseases and pests identified in our method. A detailed research of the symptoms of each class and pest is described in [18].

At each step, the window will sliding to a new location, it will predict the current region proposal at the same time, so the size of the reg layer will be 4 times larger than the number of the windows, compute the positions number of k boxes. Based on the research results of others, we found that using a variety of different feet the feature map of degree is predicted, and better results will be obtained. Therefore, we choose to use the output of 6 convolutional layers as the prediction layer of our network, which are res3b3 [19], res5c/conv1–2 [20], res5c/conv2–2 [21], res5c. /conv3–2, pool6. There will be two branches after each layer, one branch is responsible for predicting candidate frames, and one convolutional layer prediction candidate frame can be directly followed by the feature map. The other branch is responsible for predicting which category the candidate box belongs to [22]. Of course, each The candidate box for a layer of prediction and the number of categories it belongs to are fixed. Therefore, we can get 6 sets of candidate boxes, then we only need to combine the 6 sets of candidate boxes together and calculate the loss function, then update the parameters of the network. The detailed structure of the network is described below.

For a complete target process, the first thing we need to do is to build a data set for our algorithm to learn. This data set should include the original image, as well as the location of each object in the original image. The detailed information we have Introduced in 2.2. Next we need to do a simple processing of the data set. Due to the algorithm used in this paper, we need to adjust the image to a fixed size. The fixed size used in this paper is 300 × 300. Next, we need to make an enhancement to the data, mainly because the training data included in this article is relatively small, so we need to increase the amount of data. Our loss function for the image is defined as:

Here, i is the index of a small number of anchors, and p_i is the prediction probability of the anchor i as an object. If the anchor is positive, the ground live label \( {p}_i^{\ast } \) is 1, and if the anchor is negative, it is 0. t_i is the vector representing the four parameterized coordinates of the predictive boundary frame, and \( {t}_i^{\ast } \) is the vector of the ground live frame associated with the positive anchor. Classification loss L_cls is the logarithmic loss of two classes (object and non-object). For regression loss, we use \( {L}_{reg}\left({t}_i,{t}_i^{\ast}\right)=R\left({t}_i-{t}_i^{\ast}\right) \), where R is a robust loss function (smooth L1) defined in [23].

For one of the feature maps [24] in each of the above layers, a solid number of candidate boxes. For example, if the size of the feature map is m × n and the number of candidate frames generated by each position is k, then for a detection task of a class c object, (c + 4) × k × m × n data. We call him the prior box. Here’s how k is generated.

It is necessary to generate k differences in the same position of the feature map.The candidate box is because it can handle the problem of objects of different scales. The aspect ratio used in this paper is {2,3}. In the feature map of res3b3, we only use the aspect ratio of 2, that is, in the layer of res3b3, each point on the feature map is generated. 3 candidate boxes, each of the other feature maps produces 4 candidate boxes. Thus we get m × n × k candidate boxes. The six candidate boxes consist of two squares, min_size×min_size, max_size×max_size, and the other two are calculated based on the aspect ratio and min_size. This way we have a lot of candidate boxes. Next we need to match the candidate box to the real box we have calibrated in our dataset to get the real box of our training. First we need to calculate the IOU [25] of the candidate box and the real box we generated in the current layer, and then select the positive sample with IOU greater than 0.5, and the negative sample with less than 0.5. However, there is a big problem in doing this. Because we generate a candidate box at every position, there will be a large number of samples that are negative samples, so that we will have sample imbalance when training. The final training results have a relatively large impact. Therefore, we need to deal with the phenomenon of sample imbalance here, using focal loss to deal with such a situation, mainly to balance the negative and positive samples on loss. The method used in this paper is to use the difficult sample mining method. In this paper, the value of the loss function of each candidate box is calculated in the forward calculation, and then several candidate boxes with the largest loss function value are selected to be retained. Other candidates The box is directly discarded to ensure that the ratio of positive and negative samples in the training process is 3:1. In this way, we have obtained all the training data.

We tried the difficulties in our Cole and pathogens, for example [17]. It includes about 5000 images collected from many Cole fields in Korea. Pathogens and diseases can be carried out in different areas such as weather, space and moisture. Thus, the images are collected in different ways according to time (such as lighting), weather (such as heat, humidity), and sweet surfaces (using a greenhouse). Additionally, our dataset includes photographs and variations, samples from the beginning, middle and end endings, images contained with different pages of plants (eg, stems, leaves, fruits, etc.), different plants, green make-up plants, etc. components Examples for each class are illustrated in Table 1. After the additional supplementary techniques, the number of annotation figures is similar to the set limits for the image in the image: geometric change (change, crop, fluctuation, fluctuation) and radical change (contrast to lighting, color, sound). The previous class is a horizontal class with a more detailed picture. When creating a CNN storage library, its box is used as an unpleasant example.

However, the alternate page view of Alternaria is a very similar type of rust, which determines their number. As shown in Fig. 5, an example of GoogLeNet Inception is used to remove the original statues of the image, making it more likely to deteriorate in different places. Therefore, the requirements used for CNN are a clear indication of the rape of the leaves.

In addition to the databases used [18], we have a new class with “yellow curl” images. As mentioned earlier, one of the major problems that we have found to reduce energy generated by the system is the balance between classrooms because of the limited resources and data available. This can be illustrated by the number of individual pictures, as shown in Table 1 and Fig. 6.

The size of the data cells affects the correctness of the apple’s awareness. In this paper, two experimental teams are being conducted to show that the damage is real. The model was already taught and after data development. From the results shown in Fig. 7, it may be clear that the teaching method is a non-disclosure process, and eventually results in a total of 86.79%. However, the additional requirements and policies make up 97.62% accurately, making this total by 10,83% compared to those that were not lost. It can be seen from the results which are known as, as shown in Fig. 7, this phenomenon is due to the following reasons: (1) Systematic arrangements that are made up of various specialized photography techniques provide many opportunities for learning how to apply models on CNN; (2) the spread of different types of photographs to the teaching of weight gain in the example from CNN, but the photographic figure of the image is a wide range of variations and will cause serious problems; (3) The picture begins with the image of an apple photograph, so the CNN model can better appreciate the apple’s character made from apple fruit. The display results show that data expansion may result in increased power levels.

Table 2 shows the final results of our refined system. The relative value proves that all categories have a satisfactory improvement compared to our previous results. The average accuracy has increased by about 13%. In fact, this is due to the implementation of a secondary diagnostic unit (CNN filter bank) that enables the system to filter misclassified samples, with a primary focus on reducing false positives.

Sample size and change are another key point influencing the end result. For example, in the case of gray mold, the number of samples is less than that of leaf mold. In addition, the gray model class shows high intra-class variability, which may confuse the system with other classes (see Fig. 7).

On this paper, a new way to get an UAV with wireless clocks (WSN) is run to look different from the plots. The following approach can overcome the problems associated with the distribution of interest-oriented areas, such as how long it is not possible to use WSN, related and continue to look at different issues. These problems have been supported by using a basic WSN mobile device that can always remember the subnetworks that are available in all regions. Because the movement organs are the ones that exist, it is easy to deal with problems related to soil conditions. We have been asking for a strong - known knowledge of mysterious learning known by pathogens. This process creates a cure for disease and the location of tomato trees, which represents a significant difference in other ways of fighting disease. Our review uses pictures that are carried out in situ with various camera tools and uses them using the latest technology and software programs using the GPU, rather than using a collection of natural resources (vegetation, vegetation) and research in the laboratory. In addition, our data on the diseases of tomato and insect pests have different functions, such as light, size, and so on. Our goal is to gain a deeper understanding of our work experience. Therefore, a comparison between the test results and the various depths of construction and the demonstration of how our detective detection system can find eight types of diseases and pests, including various and varied challenges. Additionally, we find that using advanced and more advanced technology can achieve better performance. We hope that our prepared program will play a key role in the research process. Future work will improve the future results and further work will spread the knowledge of disease and insects to other plants.

The main purpose of the project is to use the UAV as a link to airline and to fix WSN change when the UAV is near your target area. Ad-hoc communications and communication systems have been used to solve these problems. In addition, the establishment of a user interface, can deal with data and monitor data. We need to emphasize that the cheap and reliable UAV platforms are now available and can be easily used for such purposes. The project provides an opportunity to try using four automobile vehicles in real-life situations. The UAV can stop the project and fly over a different distance.