Dean Diepeveen

Plant disease adversely impacts food production and quality. Alongside detecting the disease, estimating its severity is important in managing the disease. Artificial intelligence deep learning-based techniques for plant disease detection are emerging. Unlike most of these techniques, which focus on disease recognition, this study addresses various plant disease-related tasks, including annotation, severity classification, lesion detection, and leaf segmentation. We propose a novel approach that learns the disease symptoms, which are then used to segment disease lesions for severity estimation. To demonstrate the work, a dataset of barley images was used. We captured the images of barley plants inoculated with diseases on test-bed paddocks at various growth stages. The dataset was automatically annotated at a pixel level using a trained vision transformer to obtain the ground truth labels. The annotated dataset was applied to train salient object detection (SOD) methods. Two top-performing lightweight SOD models were used to segment the disease lesion areas. To evaluate the performance of the SODs, we have tested them on our dataset and several other datasets, including the Coffee dataset, which has expert pixel-level labels that were unseen during the training step. Several morphological and spectral disease symptoms, including those akin to the widely used ABCD rule for human skin-cancer detection, i.e., asymmetry (A), border irregularity (B), colour variance (C), and diameter (D), are learned. To the best of our knowledge, this is the first study to incorporate these ABCD features in plant disease detection. We further extract visual and texture features using the grey level co-occurrence matrix (GLCM) and fuse them with the ABCD features. For the coffee dataset, our method achieved 82+% detection accuracy on the severity classification task. The results demonstrate the performance of the proposed method in detecting plant diseases and estimating their severity.

LoRa is a wireless technology suited for long-range IoT applications. Leveraging LoRa technology for image transmission could revolutionize many applications, such as surveillance and monitoring, at low costs. However, transmitting images, through LoRa is challenging due to LoRa’s limited data rate and bandwidth. To address this, we propose a pipeline to prepare a reduced image payload for transmission captured by a camera in a reasonably static background, which is common in surveillance settings. The main goal is to minimize the uplink payload while maintaining image quality. We use a selective transmission approach where dissimilar images are divided into patches, and a deep learning Siamese network determines if an image or patch has new content compared to previously transmitted ones. The data is then compressed and sent in constant packets via HARQ to reduce downlink requirements. Enhanced super-resolution generative adversarial networks and principal component analysis are used to reconstruct the images/patches. We tested our approach with two surveillance videos at two sites using LoRaWAN gateways, end devices, and a ChirpStack server. Assuming no duty cycle restrictions, our pipeline can transmit videos—converted to 1616 and 584 frames—in 7 and 26 min, respectively. Increased duty cycle restrictions and significant image changes extend the transmission time. At Murdoch Oval, we achieved 100% throughput with no retransmissions required for both sets. At Whitby Falls Farm, throughput was 98.3%, with approximately 71 and 266 packets needing retransmission for Sets 1 and 2, respectively.

Artificial intelligence (AI) and deep learning (DL) for plant disease detection are emerging research areas. DL methods generally require a large amount of annotated data for training, which is often costly, time-consuming, and infeasible. This article addresses the data scarcity problem and proposes a few-shot learning (FSL) method for barley plant disease detection. To prepare a dataset, we captured images from outdoor test-bed trials (at two different growth stages of plants across multiple paddocks) under various weather conditions, such as sunny and cloudy. The images are divided into patches and manually labelled into five classes: no-disease, net form net blotch (NFNB) (which is classified into two stages: early and severe), spot form net blotch (SFNB), and scald. We name this as the Barley dataset. We also used the publicly available cassava dataset, which has five classes. The datasets are then applied to the proposed FSL pipeline, which only uses as few as five images for each class in training. We use the Swin transformer as the network backbone. The method with the Swin-B variant as the feature extractor achieved a detection accuracy of 91.80% and 97.93% on the barley disease and cassava datasets, respectively. The result indicates that our FSL model can efficiently perform and classify barley disease with small training data.
•A few-shot learning method is developed to address data scarcity problems.•Results on the collected plant disease data warrant the model’s potential.•Cutting-edge transformers, e.g. Swin-B, perform well given only five training images•Meta-training and transfer learning significantly improve performance.•Apparent disease symptoms can be detected and used in various applications.

Existing explicit and implicit-explicit hybrid neural representations for novel view synthesis are scene-specific. In other words, they represent only a single scene and require retraining for every novel scene. Implicit scene-agnostic methods rely on large multilayer perception (MLP) networks conditioned on learned features. They are computationally expensive during training and rendering times. In contrast, we propose a novel plane-based representation that learns to represent multiple static and dynamic scenes during training and renders per-scene novel views during inference. The method consists of a deformation network, explicit feature planes, and a conditional decoder. Explicit feature planes are used to represent a time-stamped view space volume and a shared canonical volume across multiple scenes. The deformation network learns the deformations across shared canonical object space and time-stamped view space. The conditional decoder estimates the color and density of each scene constrained by a scene-specific latent code. We evaluated and compared the performance of the proposed representation on static (NeRF) and dynamic (Plenoptic videos) datasets. The results show that explicit planes combined with tiny MLPs can efficiently train multiple scenes simultaneously. The project page: https://anonpubcv.github.io/cplanes/.

Automatic weed detection and classification can significantly reduce weed management costs and improve crop yields and quality. Weed detection in crops from imagery is inherently a challenging problem. Because both weeds and crops are of similar colour (green on green), their growth and texture are somewhat similar; weeds also vary based on crops, geographical locations, seasons and even weather patterns. This study proposes a novel approach utilising object detection and meta-learning techniques for generalised weed detection, transcending the limitations of varying field contexts. Instead of classifying weeds by species, this study classified them based on their morphological families aligned with farming practices. An object detector, e.g., a YOLO (You Only Look Once) model is employed for plant detection, while a Siamese network, leveraging state-of-the-art deep learning models as its backbone, is used for weed classification. This study repurposed and used three publicly available datasets, namely, Weed25, Cotton weed and Corn weed data. Each dataset contained multiple species of weeds, whereas this study grouped those into three classes based on the weed morphology. YOLOv7 achieved the best result as a plant detector, and the VGG16 model as the feature extractor for the Siamese network. Moreover, the models were trained on one dataset (Weed25) and applied to other datasets (Cotton weed and Corn weed) without further training. The study also observed that the classification accuracy of the Siamese network was improved using the cosine similarity function for calculating contrastive loss. The YOLOv7 models obtained the mAP of 91.03 % on the Weed25 dataset, which was used for training the model. The mAPs for the unseen datasets were 84.65 % and 81.16 %. As mentioned earlier, the classification accuracies with the best combination were 97.59 %, 93.67 % and 93.35 % for the Weed25, Cotton weed and Corn weed datasets, respectively. This study also compared the classification performance of our proposed technique with the state-of-the-art Convolutional Neural Network models. The proposed approach advances weed classification accuracy and presents a viable solution for dataset independent, i.e., site-independent weed detection, fostering sustainable agricultural practices.

While LoRa overcomes the high-power consumption and deployment costs of GPS and mobile networks, it faces challenges in accuracy. This paper presents a method for LoRa-based localization and tracking. It uses unsupervised symbolization to analyze received signal features. We use partitioning, D-Markov machines for symbolization and the Chinese restaurant process to achieve unsupervised symbolization. In particular, a novel adaptive feature extraction technique is proposed in partitioning to overcome the problems of over-tracking and under-tracking. Mean spectral kurtosis analysis is performed across several partitioning techniques to assess their symbolization effectiveness. This enables the selection of the most appropriate partitioning technique. This enhances the localization and tracking accuracy of target objects by focusing on robustness to noise and multipath effects. The proposed method learns and estimates the distance range simultaneously, thereby eliminating the need for a separate offline training phase and the storage of reference coordinates. Experimental results using LoRa highlight the proposed method's efficacy in real-time localization, tracking, and superiority over the state-of-the-art method.

This paper proposes a long-range (LoRa)-based outdoor localization and tracking method. Our method presents an unsupervised localization approach that utilizes symbolized LoRa received signal features, such as RSSI, SNR, and path loss, where each symbol represents a system state. To identify the partitioning boundaries between the symbols in time series, we employ maximum entropy partitioning. The D-Markov machine is used to construct nondeterministic finite-state automata for extracting temporal patterns. We incorporate the Chinese restaurant process for online estimation, especially in scenarios with an unbounded number of probable areas around each LoRa gateway. An adaptive trilateration approach is then used to localize the target node from the estimated ranged radii of areas. The point-wise localization data was used for time-series continuous tracking. We collected a dataset using three LoRaWAN gateways, sensor nodes powered by single-use batteries, and a Chirpstack server on a sports oval. We thoroughly evaluated the proposed method from the perspectives of localization accuracy and tracking capability. Our method outperformed state-of-the-art machine learning-driven range-based and fingerprint-based localization techniques.

Frost damage significantly reduces global wheat production. Temperature development in wheat crops is a complex and dynamic process. During frost events, a vertical temperature gradient develops from soil to canopy due to the heat loss from the soil and canopy boundary. Understanding these temperature gradients is essential for improving frost management strategies in wheat crops. We hypothesise that the relationship between the temperatures of the canopy, plant and ground can be an early indicator of frost. We collected infrared thermal (IRT) images from field-grown wheat crops and extracted the temperatures from the canopy, plant and ground layers. We analysed these temperatures and applied four machine learning (ML) models to detect coldness scales leading to frost nights with different degrees of severity. We implemented a gated recurrent unit, convolutional neural network, random forest and support vector machines to evaluate the classification. Our study shows that in these three layers, temperatures have a relationship that can be used to determine frost early. The patterns of these three temperatures on a frost night differ from a cold no-frost winter night. On a no-frost night we observed that the canopy is the coldest, plant is warm, and the soil is warmest, and these three temperatures did not converge. On the other hand, on a frost night, before the frost event, the canopy and plant temperatures converged as the cold air penetrated through the canopy. These patterns in temperature distribution were translated into an ML problem to detect frost early. We classified coldness scales based on the temperatures conducive to frost formation of a certain severity degree. Our results show that the ML models can determine the coldness scales automatically with 93%–98% accuracy across the four models. The study presents a strong foundation for the development of early frost detection systems.

Aphids are persistent insect pests that severely impact agricultural productivity. The detection of aphid infestations is critical for mitigating their effects. This paper presents an artificial intelligence approach to detect aphids in crop images captured by consumer-grade RGB imaging cameras. In addition to detecting the presence of aphids, the size of the aphid is an important indicator of infestation severity. To address these, we present a Bayesian multi-task learning model to detect the presence of aphids and estimate their size simultaneously.

Our model employs a joint loss function, combining a classification loss and a customised size loss. The classification component aims to identify images containing aphids, whilst the customised size loss function estimates the size of the aphids. The latter is specifically designed to account for discrepancies between the estimated and actual ground truth sizes, enhancing the accuracy of the size estimation. The model utilizes a ResNet18 backbone, ensuring robustness and adaptability across various conditions.

The proposed model was evaluated using an agricultural pest dataset consisting of images of corn, rape, rice, and wheat crops. It achieved aphid presence detection accuracies of 75.77%, 66.39%, 70.01%, and 59% for corn, rape, rice, and wheat images, respectively. An in-depth evaluation of predictive uncertainties revealed areas of high confidence and potential inaccuracies for both size and presence of aphids in images, offering insight for future model refinement. We also conducted an ablation study to thoroughly analyse the contributions of each component in proposed model.

Our model offers a valuable tool that can be used in pest management strategies for facilitating more sustainable and efficient agricultural practices.

Summary Two‐way layouts are common in grain industry research where it is often the case that there are one or more covariates. It is widely recognised that when estimating fixed effect parameters, one should also examine for possible extra error variance structure. An exact test for heteroscedasticity, when there is a covariate, is illustrated for a data set from frost trials in Western Australia. While the general algebra for the test is known, albeit in past literature, there are computational aspects of implementing the test for the two way when there are covariates. In this scenario the test is shown to have greater power than the industry standard, and because of its exact size, is preferable to use of the restricted maximum likelihood ratio test (REMLRT) based on the approximate asymptotic distribution in this instance. Formulation of the exact test considered here involves creation of appropriate contrasts in the experimental design. This is illustrated using specific choices of observations corresponding to an index set in the linear model for the two‐way layout. Also an algorithm supplied complements the test. Comparisons of size and power then ensue. The test has natural extensions when there are unbalanced data, and more than one covariate may be present. Results can be extended to Balanced Incomplete Block Designs.