CNN

CNN is broadly used in agriculture since it has strong capacity for image processing. Major applications of deep learning in agriculture can be classified as plant or crop classification, pest and yield prediction, robot harvesting, monitoring of disaster etc. Mainly disease recognition model for plants can be applied by leaf image and pattern classification. Berkley Vision and Learning Centre has developed a novel deep learning framework to build a disease detection model. This system is capable of identifying around 10–​15 cases of ailment leaves from healthy leaves, it also has a capability to separate plant leaves from the surroundings. In 2007, to control and identify weeds an approach was designed which is a combination of CNN and K-​mean features learning. Manual model for weed detection leads to false recognition and weak extraction skill in feature extraction. One of the pretrained CNN architects broadly used in classification of plants is Alexnet. Based on experimental results of Alexnet, we can say that CNN architect outperforms the machine learning algorithm, that is, hand-​crafted structures for the unfairness of phonological phases. For optical image division and subsequent restoration of missing information in a time series of satellite imagery, self-​organizing kohonen maps are used. In this method for post-​processing setup, geospatial analysis and several filtering algorithms are used. Although CNN has several uses, it faces many challenges that have slowed down its application in plant classification. For example, each pixel of space borne SAR imagery is characterized by backscatter phase and intensity in multiple polarizations. For yield prediction and robot harvesting, fruit counting is one of the important factors. We cannot produce satisfactory results through traditional counting or video or camera image counting and also these processes are time consuming. Preprocessing of these types of images is challenging because of occlusion and illumination. Hansen et al. (2018) introduced a technique to identify the livestock animal such as pig using the face recognition feature of CNNs. Conventionally, radio frequency identification tags were used for detecting the animals which was earlier a cumbersome job.

To accompany a fully convolution network, a method known as blob detection was proposed. The first step is to gather the human formed labels from a set of fruit images and then this model is trained for an image segmentation performance. Then CNN is used to count the bifurcated pictures and give an approximation of number of fruits. The last stage of the work involves applying a regression equation to map intermediate fruit count estimation to final human generated label count. Accuracy as well as efficiency is increased by combining deep learning with blob detection.

Land classification method is used to identify the land as land use and cover, for disaster assessment of risk and for food and agriculture. Overall idea of deep learning method is to integrate information developed by multiple heterogeneous sources using machine learning techniques to provide information processing and picturing capability. This process includes four steps: (i) noise filtration and data clustering, (ii) clearing land cover, (iii) map post-​processing, (iv) geospatial analysis. Kussul et al. (2017) introduced a multilevel deep learning approach for land cover and crop types classification using multitemporal multisource satellite imaginary.

Nowadays, new technologies are emerging in agriculture sector; for example, for high-​quality image processing unmanned aerial vehicles are used. For farmers, operating highly autonomous machines is quite difficult. It is not easy to operate these machines without supervision. So, detection of real time risk with these machines automatically with high reliability becomes a requirement. For sustainable land use certain conditions need to be considered, they are planning regarding reducing CO₂ emission, diminishing land degradation and improving economic returns using valuable data from satellites. For decision-​making in precision agriculture and agro-​industry, CNN and genetic algorithm have become convenient methods with the use of translating satellite images. Weather forecasting is one of the main factors for farmers that can be predicted using CNN. Similarly, crop yield estimation is one of the main factors for farmers, consumers as well as for the government which needs to be predicted before the harvest of the crop. CNN is not only used for crop estimation or agriculture purpose but also is used in classifying the animal behaviour.