INGENIEUR
processor ( see Figure 1 ) which connects IoT devices or applications to 2 ) processing center on the cloud ( Internet ).
Object Processor The object processor is set up at the base station . It is dedicated to connecting sensing modules , processing raw data from IoT devices and transmitting them to the processing center . More importantly , the object processor can receive feedback from the processing center to carry out / perform an intelligent action . Object processor is also known as fog computing — an architecture that uses edge devices to carry out computation , storage and communication locally .
Processing Center The processing center continuously and automatically processes massive amounts of data online . Multiple intelligent computing systems are placed at the processing center as a computing hub . It is responsible for intelligently processing IoT data from the object processor and sending instructions to the object processor or data storage . The most common open-source library for data analytics on the processing center ( cloud ) is TensorFlow , which is originally developed by Google for deep learning applications . Another important characteristic of a processing center is its capability to adopt reinforcement learning to optimise communication networks as network load and time consumption increases .
In our work at the Smart Farming Technology Research Center , University Putra Malaysia ( UPM ), in collaboration with Malaysian Agricultural Research Institute ( MARDI ), we utilise deep learning technique for insect pest recognition in paddy fields . Mobile-based RGB imaging systems were integrated with the conventional insectpest light traps to collect insect-pest images simultaneously in a paddy field . Fog computing using Python as an object processor was used to pre-process the RGB insect pest images , including image resizing , labelling and recording conversion ( see Figure 2 ).
A total of 1,142 insect-pests were labelled , where Zig Zag Brown Plant Hopper ( BPH ) and Green Leaf Hopper ( GLH ) were identified as the main population during that particular season of the test experiment . A pre-trained deep learning model called Faster Recurrent Convolutional Neural Network ( R-CNN ) was utilised under the Tensorflow framework at the processing center in the cloud . The Faster R-CNN model was able to detect the insect-pests , classify and count the pests population in the paddy field with an accuracy of 93 %. Such a recognition model can be exploited as an I-IoT application for farmers to identify fruits , vegetable or plant diseases based on images captured by their mobile devices . Furthermore , with complementary data and enhancement to the model , farmers may be able to select remedies or pesticides to improve crop production and / or quality .
Deep Learning has also been used in our previous work on oil palm fresh fruit bunches ( FFB ) quality classification at the oil palm mill for automatic sorting of fruit bunches . In this work , an optical-based imaging system was developed at the sorting conveyer to collect FFB images as the fruits travel on the conveyor in real-time . We used a deep convolutional network implemented using the Tensorflow framework at the processing center on the cloud . In this model , the FBB can be classified into five ripeness or quality categories : empty bunch , under ripe , unripe , ripe and overripe with approximately 80 % accuracy . The images were subjected to noise such as very busy image backgrounds due to foreign objects like soil dirt on the conveyer during data acquisition . The accuracy can be improved when object processer is implemented locally at the edge to pre-process the image data .
CHALLENGES AND FUTURE TRENDS
There are several challenges associated with the deployment and application of I-IoT in agriculture .
Firstly , the challenge of interference , which will likely arise due to the massive amount of IoT devices for agricultural and other purposes deployed together using unlicensed spectrums such as Zig Bee , Wi-Fi , Sigfox and LoRa . This can lead to loss of data and reduced reliability of the IoT ecosystem . On the other hand , for applications that utilise image or vision data , operating IoT devices on licensed spectrum will add cost to the capital expenditure of the agricultural IoT .
14 VOL 84 OCTOBER-DECEMBER 2020