Waste in rivers is a significant problem that can impact human and aquatic life systems. This paper looks at the implementation and steps of classifying waste in rivers using Computer Vision and Deep Learning. Due to technological improvements in computer processing systems Deep Learning has improved in terms of accuracy and efficiency and hence Computer Vision has become more viable in real-world applications. This research investigated the use of the Convolutional Neural Network DenseNet-201 for the classification of real-world waste data in rivers. In order to develop the most accurate model, DenseNet-201 was trained and tested under different conditions. A custom waste dataset was created for the purposes of the study and different pre-processing pipelines were explored for the purpose of investigating its effects on classification accuracy. The first dataset was unmodified with 1,600 images of 100×100 size and 6 waste classes, the second dataset had aspect ratio adjustment, the third dataset was augmented using ImageDataGenerator, and the fourth was augmented using Albumentations augmentation pipeline and resulted in 24,000 images in the dataset. Two variants of the DenseNet-201 model were trained. The first variant was fully trained on each dataset to investigate the accuracy of the DenseNet-201 neural network fully trained on the four datasets without Transfer Learning. The second variant of DenseNet-201 was Fine-Tuned and trained on the four datasets three times each to investigate three levels of compute resources and their effects on classification accuracy and model performance. The resulting model was successfully used to classify the real-world data collected from a local stream.