For the disease occurrence prediction modeling, NCTU_Formosa built a model predicting the diseases relating to everyday weather information that was based on the neural network. There were two kinds of data in our predict system, which were the recorded diseases happening data that were collected from the government agency and the weather data from the Central Meteorological Bureau's websiteCentral Meteorological Bureau's website that corresponded to the data forward. Then, we combined these two, and then deleted the recorded data that had no weather data to match with.

     In reality, the spore germination needs some specific weather conditions. Also, the plants may be influenced by the change of weather and become weak. However, this relationship is hard to be recognized by humans. Therefore, we used Convolutional Neural Network(CNN) as our method to catch the weather patterns.

Convolutional Neural Network(CNN)

     Convolutional Neural Network(CNN) is a powerful network that can automatically recognize the patterns of the features in a period of time, just like the favorite weather change to the fungal diseases in the life cycle. In other words, it means it can recognize the pathogen life cycle in a specific time and mark them out because life cycle’s patterns can happen everywhere over a period of time. For example, our model used weather data for tje past two weeks as the model input, and the weather patterns can be found in this 14-day data.

     After the CNN layer, the weather features then converted to weather change features, then a max pooling layer was added to filter some noises after the CNN layer. Weather patterns that caused diseases wouldn’t change in a short time, so the function of max pooling was to only return the maximum values in the filter.

一個公式

     For example: if the input array is [2,5,1,7,0,4] and the max pooling filter size is 2. When the filter step(distance the filter moves) is 1, the first max pooling output will be max(2,5)=5, and the second output is max(5,1) =5 and so on.

一張圖

     Finally, after using the full connection layer to convert the max pooling output into high dim space, the model then classified them into two classes, negative(disease will not happen) and positive(disease will happen).

     However, the network output was (a number that) hard to be realized by humans , so we used the softmax function to transform it into the diseases happening probability. Here is the formula of the softmax function:

一個公式

     So on, the model output can be easily realized by machines and humans, and next let’s talk about how we defined our model cost function and the optimizer.

     We chose the cross-entropy as the network cost function because it performed well that the exclusion classification mission. Here this the formula of cross-entropy.

     Suppose His a cross-entropy function, y'i is the real label, and the yiis the network prediction output, the cost is

一個公式

     Parameters of the Neural Network were being optimized by Adam optimizer, which is a moat commonly used ways to optimize the network.

     After training, the model will be tested by independence test data, and the result was shown below.

一張圖片

     For the use in reality, the network would connect to our IoTtalk system, giving the user daily warning from the diseases happening.