Rice is an essential food source in the Philippines, yet ensuring its quality remains challenging due to the visual similarities among rice varieties. These similarities often lead to mislabeling and adulteration, undermining consumer confidence and affecting market integrity. This study presents a deep-learning approach to automated rice variety classification and adulteration detection. A custom image dataset containing 2,400 samples was acquired, comprising four classes: Dinorado, Malagkit, Sinandomeng, and adulterated rice. Three models were evaluated - DenseNet121, EfficientNetV2S, and Mobile Vision Transformer (MobileViT). EfficientNetV2S achieved the highest performance with a test accuracy of 100%, demonstrating a superior balance of accuracy, training time, and inference speed. It was found to be 45.69% as fast in inference as DenseNet121, which achieved a 99.5% test accuracy. The lightweight MobileViT model, while indicating underfitting, provided a test accuracy of 92% and the fastest inference time, proving to be at least 31% faster than EfficientNetV2S. The results highlight EfficientNetV2S as the most effective and reliable solution for accurate rice variety identification, while also indicating the significant potential of lightweight models like MobileViT for future real-time applications in resource-constrained environments with additional optimization.

Recent advancements in computer vision, notably in 6D pose estimation with methods like DeepLabv3 and Dense Fusion, show promising results. However, a substantial research gap exists in addressing the limitations of these methods in occluded scenarios, which in need further research to enhance their effectiveness and adaptability in situations involving hidden or partially obscured objects. This study introduces a novel approach employing 6D object pose estimation and pose correction by integrating DeepLabV3 and modified dense fusion. The study exhibits enhanced accuracy in pose estimation and correction for occluded 3D object model x-ray source. Evaluation metrics, including intersection-over-union and mean average, demonstrate high accuracy percentages for detecting the body (98.91%), handle (96.57%), and aperture (89.54%). The mean IoU for each part of the 3D model portable X-ray source ranges from 65.69% to 76.42%. Pose estimation accuracy, assessed through the Average Distance Difference (ADD) metric, indicates superior performance for static pose estimation closer to the camera. Dynamic pose estimation exhibits higher average ADD metrics in scenes with total occlusion. The robustness metric reveals lower lost tracking counts in scenes without occlusion, emphasizing the algorithm's challenges in fully occluded scenarios.

Polarimetric decomposition parameters are crucial for characterizing ground objects by their unique scattering mechanisms, particularly in crop monitoring and specifically for mapping paddy rice fields. This study evaluates their impact on mapping accuracy by deriving entropy (H), anisotropy (A), and mean alpha angle (a) from a multi-temporal Sentinel-1 dataset. These parameters served as input features for training a random forest classifier to delineate rice-growing areas. Additionally, another random forest classifier was trained using backscatter intensity from the VH (Vertical Transmit, Horizontal Receive) polarization channel to provide a comparative map, enhancing the evaluation of the study outcomes. Adjustments to the temporal resolution of the input dataset were necessary due to the computational constraints of Google Earth Engine (GEE), the cloud-based platform used in this study. Despite the reduced temporal resolution, the classifier trained with polarimetric decomposition parameters performed comparably to the classifier using the full temporal resolution of backscatter intensity. The classifier utilizing backscatter intensity reported slightly higher overall accuracy and kappa coefficient, with values of 0.8460 and 0.6920, respectively. The classifier using polarimetric decomposition parameters achieved an overall accuracy of 0.8250 and a kappa coefficient of 0.6499. The performance of the two classifiers varied with respect to the calculated producer's and user's accuracies, though the differences were not significant. These results suggest that entropy (H), anisotropy (A), and mean alpha angle (a) effectively capture the distinct scattering properties of rice throughout its growth stages, as well as those of other crops and land cover types, enabling reliable rice classification. Future research could explore additional parameters and their combinations to further enhance mapping accuracy.

Mapping paddy rice fields is crucial for monitoring rice production, which is vital for food security and economic stability. Prior studies have utilized remote sensing data along with various machine learning algorithms to accomplish this task. However, the performance of these classifiers under different conditions remains an area worth investigating. In this study, rice-growing areas are delineated from other crops and land cover types using multi-temporal synthetic aperture radar (SAR) imageries recorded by Sentinel-1 and classification algorithms—Naïve Bayes, CART, and Random Forest—on the Google Earth Engine (GEE) platform. Results show that the random forest classifier outperforms the others with an overall accuracy of 85.90% and a kappa coefficient of 0.7180. This superior performance can be attributed to its capability to handle complex, high-dimensional data, which aligns well with the properties of the multi-temporal SAR data used to observe the growth cycle of paddy rice. The differing performances of the three classifiers underscored the importance of selecting an appropriate algorithm, especially when leveraging the distinctive radar signatures of paddy rice over time, to enhance the accuracy of mapping paddy rice fields.

Edible wild plants are an important source of food in many regions of the world, including in the Philippines, and their recognition is a key skill for survival in the wild. In this study, we propose a mobile platform for image recognition of edible wild plants using deep convolutional neural networks (CNNs). The proposed system is designed to be lightweight and easily deployable on mobile devices, allowing for real-time recognition of edible plants in the field. To develop the system, we first collected a dataset of images of various edible wild plants. The dataset was preprocessed and augmented for better generalization. We then trained a CNN model using transfer learning techniques on a custom specific dataset of edible wild plant images endemic to the Philippines to recognize the different species of edible plants. The trained model was then optimized for deployment on mobile devices, and the resulting mobile application was tested on a variety of wild plants. The results showed that the proposed system achieved high accuracy in identifying edible wild plants, with an average accuracy of 96.98%. The proposed system has many potential applications, including in the field of outdoor education, potential solutions to address food scarcity, and survival training. It can also be used by foragers and hikers to identify edible plants in the wild, helping to prevent the consumption of toxic plants. Additionally, it can be used by researchers to gather data on the distribution and abundance of edible plants in different regions. The proposed mobile platform for image recognition of edible wild plants using CNNs is a promising tool for enhancing the safety and sustainability of foraging and outdoor activities.

The limited physical constraints of nanosatellites due to their size, hinders their ability to transmit large amounts of image data. Because of this, the use of machine learning methods to filter data onboard has become more prominent to increase the bandwidth efficiency of these devices. By having an AI-based classification system for the images, the bandwidth necessary to transmit all these images and the tradeoff when it comes to storage, can potentially be offloaded through having a system which generates metadata that can indicate the data samples which offer the most usability, thus freeing up more space and bandwidth for these more important samples. This study explores the task of land cover classification, by utilizing one of the more prominent image segmentation models, U-Net. The model is implemented and evaluated using Pytorch using the DeepGlobe 2018 land cover classification dataset, achieving an average class IoU score of 0.68. This study seeks to support the viability of such a solution and is intended to support any future work which seeks to implement a fully automated data prefiltering system for satellite imagery.

The development of machine learning methods for onboard satellite processing is important in order to facilitate the filtering of collected data samples to maximize the use of the device's limited resources. Land cover classification can be used to focus the collected data on certain terrain types by utilizing classification methods to determine the class probabilities of individual pixels in a collected satellite image. The importance of the accuracy of the segmentation model used for such a task is important in order to avoid the trashing of data samples that offer significant information and the prioritization of data samples which offer less in terms of usable information, which in the case of land cover classification is determined by which terrain features may be prioritized over others. This study focuses on the U-Net segmentation architecture and performs an experimental study on the effects on two aspects on the training of a segmentation model for increased performance. This includes the division of the images in the dataset into smaller patches and the replacement of the CNN encoder of the segmentation architecture. The changes made to the baseline model introduced an increase in the IoU score from 0.68 to 0.7273.

Public transportation is the key economic driver of a country. The true measure of a country's progress level is scaled on the number of people using the public transportation rather than of people riding private cars. In the Philippines, Western Visayas region (Region VI) is one of the regions which needs extensive support in public transport data organization. Due to the complexity of a public transport network, handling of big data becomes a bottleneck for transport planners. Addressing this problem will help them move forward to more important tasks such as improving transport service for passengers. In this study, a framework was designed in modeling public transportation data. TigerGraph database was utilized to preconnect data and to allow acquisition of geospatial intelligence on route while Django-python was used as the web framework for the geographic insight web application. With the framework and software solution developed, the study intended to make data organization scalable, visualize data relationships, and preconnect data. Preconnecting data in public transport such as terminals, PUV stops, and facilities in conjunction with massive parallel processing (MPP) function, speeds up data analysis. This also enables expanded capability of a system to return answers to queries which need deeper analysis.

Often camera calibration in terms of lighting has become a challenge in machine learning. Large training datasets are usually required due to various light conditions that affect the colors on the images, making objects difficult to recognize. This paper proposes the utilization of K-means clustering to extract colors on the images to be used in combination of color Homography to correct colors in low light images automatically. This method aims to solve tedious camera calibration in terms of color and reduce the number of datasets.

Data mining today is much slower than before because of the advancement of computing and information systems. Relevant recommendation based on customers' preferences and needs in e-commerce gets more complicated. In the recent pandemic, people are reluctant to go out and has engaged more on internet to get their daily food and services. This phenomenon exacerbated the existing recommendation system, as the data has grown up drastically. In this study, the author recommends a relevant image quality based on the quality queries of the clothes and footwear dataset by observing their highest similarity score. Fashion MNIST images used were existing dataset for clothes and footwear. The testing on image reconstruction using training and validation approaches has shown an accurate result by showing only 0.01 loss in the dataset. Using 11 classes of the image queries, the system image has been identically reconstructed according to the queries supplied. With this result, businesses will have an implementation alternative to a faster and more efficient data mining method. Hence, this alternative will boost the speed of many recommendation systems in the e-commerce platforms and will create a better customer experience.

Fish feeding management is one of the most crucial considerations in aquaculture production. The traditional feeding method such as table-based and scheduled automated feeding schemes are inaccurate. In contrast, the automated on-demand feeding system has reduced the inaccuracies of the older feeding schemes. However, existing on-demand systems have limited accessibility because their monitoring systems are only stored by their local devices. This paper proposes an on-demand fish feeding system with online and real-time monitoring using the Internet of Things (IoT) and an accelerometer to sense the fish' demand by hitting it. An overhead surveillance camera was installed on the fish tank to automatically record and monitor the fish feeding activity on the first day of the implementation. Two groups of fish were used for the observation—the adults and pre-growth Nile tilapia (Oreochromis niloticus). Results have shown that the on-demand feeding system is highly effective on 21 pre-growth fish with an average weight of 88 grams and a standard deviation (SD) of ± 39 grams. Additionally, the feed intake ratio (FIR) of the pre-growth fish was$1.35\pm 0.69$grams, i.e., 73% to 86% lower than the recommended table-based feeding scheme. Thus, more efficient.

Visual inspection of plant health status and disease severity may yield subjective assessments due to error-prone sphere of colors and textures as affected by angular photosynthetic light source and the complexity of chlorosis. Quantification of damages on leaves due to destructive diseases is paramount for plant and pathogen interactions. To address this challenge, the proposed solution is the integration of computer vision and computational intelligence for tomato Septoria leaf spot necrotic and chlorotic region computational assessment. Dataset contains healthy and diseased tomato leaves that were captured individually. Non-vegetation pixels removal was done using CIELab color space. RGB color components and five Haralick texture features were extracted from the segmented leaf. Hybrid neighborhood component analysis and ReliefF algorithm were employed to select the important predictors resulting to RGB-entropy vector. A new tomato leaf disease index (tomLDI) optimized using artificial bee colony (ABC) was developed by normalizing visible red reflectance, and introducing red-green and red-blue reflectance ratios to enhance Septoria leaf spots pixels and reducing sensitivity to healthy green pixels. KNN bested classification tree, linear discriminant analysis and Naïve Bayes in detecting Septoria leaf disease with accuracy of 97.46%. Deep transfer image regression was tested using raw infected leaf images and the tomLDI transformed colored channels through MobileNetV2, ResNet101 and InceptionV3. Using tomLDI channel, MobileNetV2 and ResNet101 bested other networks in estimating leaf diseased region percentage and number of Septoria spots with R2values of 0.9930 and 0.9484 respectively. tomLDI channel proved to be more accurate than using raw images for regression.

Lightness signal and color reflectance constitute the reflected luminance spectra from camera captured image to camera lenses. The intensity of lightness and visible RGB signals deviates as the camera distance to object varies. The presence of uneven distribution of photosynthetic light causes angular light effect of shadowing on the focal object and light emitting objects placed on the visually noisy background added a challenge in materializing an efficient greenness index for crop phenotyping. The proposed method in this study compensates excessive relative brightness on the image by introducing lightness rectification coefficient and employing genetic algorithm to derive a novel visible tetrahedron greenness index (gvTeGI) based on normalized green waveband. Hybrid neighborhood component analysis and Pearson's correlation coefficient approach for feature selection resulted to retaining photosynthetic canopy area, and correlation and homogeneity texture features as highly important descriptors for biophysical signatures considered in this study which are lettuce fresh weight, height and number of spanning leaves. The selection, crossover and mutation rates used to optimize the genetic algorithm model are 0.2, 0.8 and 0.01 respectively. Indoor and outdoor aquaponic system was deployed for 6-week full crop life cycle cultivation. Regression machine learning models were used to estimate biophysical signatures from extracted gvTeGI channels. Optimized Gaussian processing regression model bested regression support vector machine and regression tree in estimating fresh weight, height and number of spanning leaves with R2values of 0.7939, 0.7662 and 0.7446. The proposed gvTeGI proved to be more accurate than previously published greenness index for the estimation of biophysical signatures of lettuce using consumer-grade RGB camera.

Energy production of photovoltaic (PV) system is heavily influenced by solar irradiance. Accurate prediction of solar irradiance leads to optimal dispatching of available energy resources and anticipating end-user demand. However, it is difficult to do due to fluctuating nature of weather patterns. In the study, neural network models were defined to predict solar irradiance values based on weather patterns. Models included in the study are artificial neural network, convolutional neural network, bidirectional long-short term memory (LSTM) and stacked LSTM. Preprocessing methods such as data normalization and principal component analysis were applied before model training. Regression metrics such as mean squared error (MSE), maximum residual error (max error), mean absolute error (MAE), explained variance score (EVS), and regression score function (R2score), were used to evaluate the performance of model prediction. Plots such as prediction curves, learning curves, and histogram of error distribution were also considered as well for further analysis of model performance. All models showed that it is capable of learning unforeseen values, however, stacked LSTM has the best results with the max error, R2, MAE, MSE, and EVS values of 651.536, 0.953, 41.738, 5124.686, and 0.946, respectively.

With machine learning-based innovations becoming a trend, practical resolutions of its implementation to large-scale data and computing problems must be able to cope up as well. Currently, Graphic Processing Units (GPUs) are being chosen over other available physical devices due to its powerful computing capability and easier handling. Several cloud service providers also made it possible for these to be accessible online allowing higher serviceability and lower cost upfront for businesses. With this said, the proponent would implement a common machine learning-based application, automated annotation through Mask RCNN Object Detection Model in CVAT, using AWS instance. The key purpose is to showcase the viability of deploying data and computing intensive system on the cloud.

Soil nutrients and season have direct impact on the growth and yield of a crop. Deficiency on the nutrient level of the soil may result to plant disease while applying excessive amount of soil fertilizer on the other hand, may also cause negative results to the development of the crop. Nutrients on the soil also changes as the season changes from wet season to dry season. This study aims to develop a fuzzy logic-based program that will provide an appropriate amount of fertilizer to soil. The parameters such as season, nitrogen, phosphorus and potassium level are the inputs used on the fuzzy logic system. The researchers proposed four kinds of fertilizer to use in this paper such as Complete, Urea, Solophos and Muriate of Potash. Combination and amount of these fertilizers will be based on the input parameters and fuzzy rules. These soil fertilizer recommendations can be used for rice in an inbred light soil.

This study centers on the design and development of a microcontroller based hardware interface that connects the serial camera, the processor, the WiFi module, and the LCD screen and identification software for corn plant diseases through web-server using image processing and artificial neural network. This is done by capturing and displaying the image of the leaf inside the box and transmits it to the web server as an input image; process, analyze and interpret the data through image processing. The result of the processed image will be sent to the displaying microcontroller based hardware interface through the web-server and display the Pest Management Recommendations.

Monitoring the dynamics of fish behavior is impactful both in the research for fisheries and aquaculture production. One of the most explored approaches to monitor the fish is tracking-by-detection along with computer vision. Presently, there are several challenges in this field, including underwater environment conditions and fish movement complexity. This study presents an initial investigation towards tracking the fish by predicting the trajectory 2D coordinates of fish from the sequential sampled frames in underwater videos. Here, the authors explored the Genetic Algorithm based on natural evolution selection and the Long Short-Term Memory (LSTM) algorithm. Results have shown tolerable trajectory prediction inaccuracies using the GA and LSTM. Specifically, it obtained the Mean Absolute Percentage Error at 2.8% to 30.5% and 3.33% to 17.74% for GA and LSTM, respectively. These results have allowed the authors and researchers to extend its study towards tracking the fish using these approaches.

Lettuce is one of the most popular crops for urban farming because it is easy to grow and it has high nutritional value. Moreover, it is adaptable and can be combined with other food options, or it can be eaten alone without too much preparation. Predicting lettuce growth can be crucial to find the optimum maturity and harvest time. This paper proposed to use a model of a hybrid tree-fuzzy logic approach, the classification tree was used to select the most significant features from the texture features then the fuzzy inference system was utilized in predicting the lettuce growth stage classification. The hybrid system produced accurate results with low percentage error and correct classifications. Based on these results, the most accurate prediction can be observed in the head development growth stage; the harvest growth stage has a slight variance, while the vegetative stage has the most variance. Overall, the trained hybrid system is reliable in predicting and identifying lettuce growth stage classification.

This research proposes a method in order to classify vehicles in a highly congested roads , a robust technique for vehicle classification with low computational power must be used. So, a proposed solution is to embed an AKAZE feature matching extraction which is ran in an artificial neural network will be used. AKAZE was used because it is faster than SIFT. The features extracted from the AKAZE algorithm will be grouped according to the type of vehicle where it was used and be placed to an Artificial Neural Network (ANN) for the training of the network itself. The results yielded good for real-time Vehicle Classification.

Several methods exists to identify the nutrient content of the soil. The most popular method is by using Soil Test Kit (STK). STK gives soil qualitative level of macronutrients and pH. Chemicals that change color upon reaction with soil samples can determine macronutrients such as nitrogen, phosphorus, and potassium. These chemicals are going to be processed based on the method given by the kit. With the use of different algorithms that is commonly used for classification, mostly, a vision system is required. In this study, the development of the vision system that will capture the image of the soil sample after conducting soil testing will be tackled together with the image processing and feature extraction. Using the extracted features as the input of the fuzzy logic gives accurate result in determining the nutrient level of the soil.

The anaglyph is a widely overlooked method of viewing three-dimensional images on any colored display. This is done by selectively filtering the image through colored lenses. Despite the simplicity of this system, the approach to designing anaglyph images remained largely empirical until a recent mathematical analysis by Eric Dubois. While the methods shown in the said work create good anaglyphs, they still exhibit a large amount of retinal rivalry which makes anaglyphs uncomfortable to view. This paper tackles modifications to the said approach to tackle several anaglyph issues, namely ghosting, retinal rivalry, and color reproduction, simultaneously. Subjective testing showed an improvement in viewer acceptance of images designed using the proposed method.

Stock market analysis has traditionally been proven to be difficult due to the large amount of noise present in the data. Different approaches have been proposed to predict stock prices including the use of computational intelligence and data mining techniques. Many of these methods operate on closing stock prices or on known technical indicators. Limited studies have shown that Japanese candlestick analysis serve as rich information sources for the market. In this paper decision trees based on the ID3 algorithm are used to derive short-term trading decisions from candlesticks. To handle the large amount of uncertainty in the data, both inputs and output classifications are fuzzified using well-defined membership functions. Testing results of the derived decision trees show significant gains compared to ideal mid and long-term trading simulations both in frictionless and realistic markets.