Optimising post-harvest farm yield through utilization of a machine learning based-fruit disease classification model

Abstract

This research study proposed 4 key improvements to the classical fruit disease detection models which have been proven to increase their classification and accuracy levels. The globe and, more particularly, the economically developed regions of the world are currently in the era of the Fourth Industrial Revolution (4IR). Conversely, the economically developing regions in the world (and more particularly, the African continent) have not yet even fully passed through the Third Industrial Revolution (3IR) wave. Moreover, Africa’s economy is still heavily dependent on the agricultural field. On the other hand, the state of global food insecurity is worsening on an annual basis due to exponential growth in the global human population, which continuously heightens food demand in both quantity and quality. This justifies the significance of the focus on digitizing agricultural practices to improve farm yield to meet the global steep food demand and stabilize the economies of the African continent and countries such as India that are largely dependent on the agricultural sector for revenues. Technological advances in precision agriculture are already improving farm yields, especially in the more economically developed regions of the globe, although several opportunities for further improvement still exist. Hence, this study evaluated a particular area of precision agriculture, the plant disease detection models which fall under decision support systems. The aim was to gauge the status of the research in this field, identify opportunities for further research, and propose technical amendments to the traditional plant disease detection models to improve their functional efficiency and accuracy. Hence, through reviewing the available literature, this study has realized the dearth of literature focused on the real-time monitoring of the onset signs of diseases before they spread throughout the whole plant. There is also substantially less focus on real-time mitigation measures such as actuation operations, spraying pesticides, spraying fertilizers, etc., once a disease is identified. Very little research has focused on the combination of monitoring and phenotyping functions into one model capable of multiple tasks. Most of the proposed plant disease classification models are based on a 2-Dimensional ‘view’ of the sample, which might pose challenges in the case of spherical or cylindrical plant samples such as fruits. Therefore, four key proposals were made in this research study. Proposal 1 was an improved image pre-processing technique for Machine Learning-based plant disease classification models. This technique dissolves a Red-Green-Blue (RGB) image into individual red, green and blue planes and performs the thresholding process on one or more planes and superimposes the resulting binary images. This has proven to yield better feature segmentation depending on the application. Proposals 2 and 3 aimed to grant a classification model a ‘3-Dimensional view’ of the sample to eliminate any ‘blind spot’ which might be hiding important features that would directly impact the classification decision had they not been hidden. Proposal 2 achieved this by using multiple input image cameras, while Proposal 3 employed a revolving sample stand that allows a single input camera to take multiple input images (at different angles) from the sample. Proposals 2 and 3 were tested in classifying healthy from black rot-affected oranges, and they outperformed the traditional plant disease detection model by classifying correctly even the oranges with small and uneven distribution of black rot. Proposal 4 combined crucial processes that are traditionally stand-alone in farming operation into a single hybrid model, hence the Hybrid Fruit Disease-Quality Monitoring and Sorting Model. This model offers post-harvest benefits and has also been conceptualized based on fruit plant samples. This model could detect diseases; perform quality checks (punchers, skin pills, etc.); perform grading based on these quality checks; and sort the fruits into designated bins according to their assigned class. It aims to lower the price of digital farming technology and avail it to low-budget farms. This model was tested on oranges (healthy, black rot-affected, and generally damaged) and apples (healthy, botch-affected, and generally damaged). The model managed to classify each of these diseases; perform the quality check based on the general fruit damages; and sort/grade (conceptually) these fruits according to these different classes. Its classification accuracy was 100% since all the test samples were classified correctly. Although proposals 3 and 4 of this study were electromechanical systems, their modelling and testing have been limited to the electrical aspect. A full electromechanical design still needs to be implemented for a full capability study to be done in practical settings. Another limitation of this study was collecting the fruit samples with the desired disease symptoms distribution. The test samples were collected from the local vendors and the variety was greatly limited.