AN IMPROVED SWIN TRANSFORMER BASED STACKED CAPSULE AUTOENCODER MODEL FOR WEED DETECTION IN RICE FIELDS

Abstract

Weeds that grow on cultivated land compete with crop plants for essential resources such as nutrients, water, space, and sunlight. Consequently, controlling weeds is crucial for achieving successful agricultural production. One primary cause of yield decline is weed competition, which hinders the growth of paddy by competing with cultivated plants for vital resources like moisture and solar energy. Conventional methods for detecting weeds, such as UAVs and satellite imagery, are expensive to implement, while manual detection is often inefficient. To address these limitations, an enhanced Swin Transformer model for weed detection in rice fields has been proposed. Initially, the inputs are sourced from the Mendeley repository, and preprocessing is performed using image resizing and Modified Gaussian Wiener filtering (MGWF) techniques to ensure uniformity and reduce unwanted noise. The processed images are then fed into the proposed Improved Sequential Head Swin Transformer with Stacked Capsule Autoencoder (ISH-ST-SCAE) model, which detects weeds by extracting significant features. Furthermore, the model's parameters are refined using the opposition-based Orangutan Optimization Approach (O3A), enhancing its training performance. As a result, this approach efficiently detects weeds in rice fields, contributing to increased crop yields. Various evaluation metrics, including correctness, specificity, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Receiver Operating Characteristic (ROC) curve, are employed to assess the effectiveness of this framework. The proposed system achieved an accuracy of 99.45%, precision of 99.50%, recall of 99.69%, F1-score of 99.59%, specificity of 99.94%, MAE of 0.0138, and RMSE of 0.1964.