Corn is one of major crops in Indonesia. Diseases outbreak could significantly reduce the maize production, causing millions of rupiah in damages. To reduce the risks of crop failure due to diseases outbreak, machine learning methods can be implemented. Naked eyes inspection for plant diseases usually based on the changes in color or the existence of spots or rotten area in the leaves. Based on these observations, In this paper, we investigate several image processing based features for diseases detection of corn. Various image processing features to detect color such as RGB, local features on images such as scale-invariant feature transform (SIFT), speeded up robust features (SURF), and Oriented FAST and rotated BRIEF (ORB), and object detector such as histogram of oriented gradients (HOG). We evaluate the performance of these features on several machine learning algorithms. They are support vector machines (SVM), Decision Tree (DT), Random forest (RF), and Naive Bayes (NB). Our experimental evaluations indicate that the color may be the most informative features for this task. We find that RGB is the feature with the best accuracy for most classifiers we evaluate.

Plant diseases outbreaks can cause significant threat to food security. Early detection of the diseases using machine learning could avoid such disaster. Currently, deep learning, which is a recent technology in machine learning, gained much popularity for object recognition tasks. Convolutional neural network (CNN) is one major techniques for object identification in deep learning. In this paper, we evaluate the effect of different depth of CNN architectures on the detection accuracies of the plant diseases detection. Various CNN architectures with different depth are investigated. They are simple CNN baseline (with two layer of convolutional layers), AlexNet (with five convolutional layers), and VGGNet (with 13 convolutional layers). We also evaluate GoogleNet architectures. Unlike previously mentioned architectures, GoogleNet use convolutional layers with various resolutions to be concantenated with each other, emphasizing the effect on not only the deep architecture but also a wide one. The experimental results suggest that CNN with deeper architecture, i.e. VGGNet, outperforms others, indicating that having deeper architectures may be more benefit for this task.

Developing an automatic detector of plant diseases is one of application fields in machine learning. Ground-truth diagnoses of plant diseases which are conducted by experts in laboratory tests are often inapplicable for fast and cheap implementations. Using machine learning approaches, the images of leaves or fruits are used as input data. From the data, we design discriminative features that are good for diseases classification. However, finding suitable features from the images are often challenging due to high intra-variability and inter-variability of the data. In this paper, we present an unsupervised feature learning algorithm using the convolutional autoencoder for detection of plant diseases. The use of convolutional autoencoder has two main advantages. First, the use of handcrafted features is not necessary as the network itself may learn to produce discriminative features. Secondly, the procedure is conducted in an unsupervised manner and hence, no labeling of the data are required. Here, we use the output of the autoencoder as inputs to SVM-based classifiers for automatic detection of plant diseases. The method indicates to be better than conventional autoencoder with more hidden layers.

In this paper, we propose a novel deep learning-based feature learning architecture for object classification. Conventionally, deep learning methods are trained with supervised learning for object classification. But, this would require large amount of training data. Currently there are increasing trends to employ unsupervised learning for deep learning. By doing so, dependency on the availability of large training data could be reduced. One implementation of unsupervised deep learning is for feature learning where the network is designed to “learn” features automatically from data to obtain good representation that then could be used for classification. Autoencoder and generative adversarial networks (GAN) are examples of unsupervised deep learning methods. For GAN however, the trajectories of feature learning may go to unpredicted directions due to random initialization, making it unsuitable for feature learning. To overcome this, a hybrid of encoder and deep convolutional generative adversarial network (DCGAN) architectures, a variant of GAN, are proposed. Encoder is put on top of the Generator networks of GAN to avoid random initialisation. We called our method as EGAN. The output of EGAN is used as features for two deep convolutional neural networks (DCNNs): AlexNet and DenseNet. We evaluate the proposed methods on three types of dataset and the results indicate that better performances are achieved by our proposed method compared to using autoencoder and GAN.

Forest areas in Indonesia covered about 2/3 of total land areas which has about 4000 wood species. Wood identification plays a key role in wood utilization not only for determining appropriate use but also for supporting legal timber trade. However, the identification process requires high expertise and complex method which can be done in the laboratory. In order to simplify the identification process, we develop wood identification using computer vision by using Histogram of Oriented Gradient (HOG) to extract the species of wood and Support Vector Machines (SVM) to classify wood species. These methods combination will improve the accuracy of wood identification process. The result showed that the HOG method can extract the texture of woods and SVM classifier can generate the boundary decision after executing the training process. By doing the testing process of SVM classifier, the result showed that the accuracy from the identification is 70.5% for using positive testing image and 77.5% for using negative testing image. This accuracy value can be reached because the texture for each training image has different texture pattern especially the number and location of vessels.

At the moment, there are increasing trends of using deep learning for plant diseases detection. However, their implementations may be difficult in developing countries due to several reasons. First, existing deep learning models are usually trained with images with adequate resolutions. In developing countries however, with limited internet connection, models that would perform well even when data with low resolution are used are needed. Secondly, the generated models are large. Hence, most deep learning based applications are available on-line. Unfortunately, the trend for new deep learning architectures are either have larger models or require a heavy memory usage. So, models with smaller size would be preferred. In this paper, we evaluate various existing deep learning models for plant diseases detection when low resolution data are used. They are: VGGNet, AlexNet, Resnet, Xception, and MobileNet. Our focus is deep convolutional neural network (DCNN) which is commonly applied for image data. We also propose a new DCNN architecture with two branches of concatenated residual networks. It is well known that the deeper the networks the better performance of DCNN. However, DCNN with very deep networks and large number of training parameters is prone to vanishing gradient problems. One solutions for that is to apply residual networks as branches to DCNN. While it is found that increasing the branch of the networks benefit the performance, larger memory are required to train the networks. So, we apply two concatenated residual networks only. We called it Compact Networks (ComNet). We compare our method other with six popular CNN architectures. We evaluate the performance on the PlantVillage dataset and our own dataset. We collected images of tea leaves which consist of 6 classes: 5 classes of diseases that are commonly found in Indonesia and a healthy class. Our experiments show that our method is generally better than referenced DCNN networks.