For the vast majority of countries, agriculture and allied industries are their principal source of revenue. A nation's basic and vital needs include the preservation and safety of agricultural supplies. Malnutrition is a constant problem in developing countries like 'India,' and it is closely linked to nutrition security. Agriculture in any country is reliant on the number and quality of its goods, particularly plants. Diseases on plant leaves play a key role in agriculture production, having a considerable impact on crop growth output, resulting in economic losses, environmental losses, and social losses [1]. With output quantities of 21.2 million metric tonnes in 2019–2020 and 19 million metric tonnes in 2018-19, tomatoes, India's second most widely cultivated produce after potatoes, provide a reasonable view on the country's agricultural post-harvest challenges. Tomato diseases are frequent, and they are caused by a range of factors like as weather, temperature, biotic stress, and nutrition. There are up to 15 different types of tomato leaf diseases, according to current statistics, all of which have had a substantial influence on tomato output and quality, as well as inflicting significant economic losses [2]. As a result, tomato leaf disease prevention and treatment are vitally crucial in tomato production. Manually identifying sick symptoms on leaves appears to be a difficult but manageable task. Even the most experienced plant pathologists and agronomists have trouble detecting specific single or multiple diseases. As a result, erroneous assumptions and conclusions are drawn. In plant and agricultural development, the use of plant features to identify physiological changes in response to biotic and abiotic difficulties such as a lack of water, nutrients, poor illumination, bacteria, fungi, and other variables is critical.Plants that have been starved of nutrients will become sick. The nutritional requirements of plants are divided into two categories: macronutrients and micronutrients. Macronutrients include calcium, nitrogen, sulphur, potassium, phosphorus, and magnesium, among other nutrients. Plants only need trace levels of micronutrients including copper, zinc, chlorine, and iron. The potassium, nitrogen, calcium, and phosphorus deficiencies were determined by counting the leaves, measuring the plant's height, and measuring the thickness and broadness of the leaves. As a result, it will have an effect on the production and quality of the fruit.
A plant may contract a different type of infection from a variety of lesions depending on the source. Some of the most important difficulties that tomato growers face include pests and illnesses that destroy tomato trees, causing losses and lower yield. Experts are usually in charge of determining whether or not chemicals are required for production. Crop protection technologies including artificial intelligence and smartphone apps for disease categorization have become more automated in recent years [3]. As a result, a number of researchers have turned to image processing techniques to assist farmers in completing this difficult task.Tomato leaf disease detection processes now involve automated image processing, as well as expert querying of tomato disease maps. The manual identification of lesions is a time-honored procedure that is costly, time-consuming, and unreliable. Although, thanks to mapping technology, it is now possible to identify related illnesses that have a significant impact on tomato leaf growth [4].The most modern computer-aided recognition method has yet to be extensively implemented due to its extremely low accuracy, negative environmental impact, and slow recognition speed. As a result, it's vital to develop a tomato disease detection technology that can make quick and accurate decisions about tomato leaf diseases. Because tomato diseases are confined to the leaves, leaf diagnostics can assist growers in determining whether or not to spray their crops. It's worth noting that, thus far, significant progress in diagnosing diseases using leaf features has been made [5].
Among the many ways, Deep Learning (DL) techniques based on Artificial Intelligence (AI) and utilising “Convolutional Neural Network (CNN)” fared well in image identification tests. CNN is a good tool for automatically detecting and classifying diseases/lesions and pests, according to [6]. However, there are challenges with automatic detection and image classification in the field. The intricacy of the background, which matches the subject of the lesion itself, settings such as brightness changes, the angle at which the shots were taken, and the equipment used to capture the photos are all concerns [7] related to the sort of diseased images [8]. [9] has stressed the necessity of finding lesions rather than just classifying them. If the categorization is used alone, the results will be unsatisfactory until it is paired with the location of the lesion. The accuracy in tomato crop leaves was examined using non-segmented and segmented images [10, 11], and it was discovered that the network achieved an accuracy of 98.6% using segmented photos, but only 42.3% using non-segmented images. Photographs from the internet were used to create the test dataset.The segmentation of disease lesions in leaf images has a direct impact on the recognition effect of crop illnesses and the accuracy of a quantitative assessment of the severity of crop diseases. How to section damaged leaves of crops with great efficiency and good quality is a research focus. In the last two decades, traditional image processing techniques such as edge detection, colour space transformation, feature space transformation, and others have been used to extract and recognise lesions.
The creation of a method for identifying tomato leaf disease automatically has been tough. It's worth noting that both diagnosis and detection require processes that could make pinpointing the exact area within the tomato plant where symptoms appear extremely difficult. The capturing parameters are difficult to control, which makes image prediction and disease detection more difficult [12]. Furthermore, the symptoms caused by several diseases may appear to be physically identical, and discrimination methods may be based on little distinctions.In recent years, advances in computer vision have resulted in a new way for accurately diagnosing tomato blemishes [13]. This research examines the various types of biotic and abiotic plant leaf diseases that have been discovered and classified using deep learning technologies at an early stage. Many writers have studied the classification of single biotic and multi-biotic leaf diseases using pre-trained DL models. The authors employed classification models to detect the disease's spots after the initial phase of image processing, which included pre-processing the dataset.The lesions boundaries were not identified accurately with the classification model till now. Thus, this work has proposed the Hybrid-DSCNN model which segments the lesion objects after pre-processing of unstructured data for better accuracy. The proposed approach i.e., Hybrid DSCNN has used the instance and semantic segmentation to detect the foreground object from the background as shown in Fig. 1.
As a result, this work attempts to increase segmentation accuracy with a contribution of:
1. The proposed Hybrid-DSCNN model is used for the segmentation of single biotic and multi biotic leaf disease in tomato plants with the Single and Multi-Biotic Leaf Diseases (SMBLD) algorithm.
2. Hybrid-DSCNN has used a mixed approach to the U-Net and SegNet CNN models by replacing 7 Conv layers of U-Net and 13 Conv layers of SegNet CNN models in downsampling. This work has used the pooling indices (features' vector) from the encoder feature and transfers them for mapping to the corresponding up-sampling layer.
3. The proposed Hybrid-DSCNN model has converted the 5 Conv layers to 2 Conv layers with \(2 \times 2\) kernel size using a skip connection in each layer. Such skip connection method is applied to extract features’ information from lower layers.
4. This work has improved expandability as well asgeneral applicability as it is based on pixel-wiseremarks using complete deep learning segmentation networks.
Content described in the remaining paper is: Section2 discusses the literature study of segmentation; Followed section 3 is vitalbecause itdefines the method and material of the segmentation used in this work; Section 4 discusses the different evaluation metrics; resulting in part of segmented lesions and comparison of hybrid model with other models is discussed in Section 5; Section 6 discuss the conclusion part.