Satellite Image Segmentation for Agriculture Mapping using Mask R-CNN based Auto-encoder

Abstract

This research study focuses on agricultural mapping, which involves the use of satellite techniques to map and monitor agricultural lands, crops, and vegetation without physical contact with the Earth's surface. Satellite imaging, aerial photography, and other satellite methods are employed to gather information about crop growth patterns, soil moisture, and other environmental factors influencing crop health and yield. However, obtaining high-quality satellite images for agricultural mapping is challenging due to the introduction of noise during image acquisition. In this thesis, we propose a novel approach called Satellite Image Segmentation for Agriculture Mapping using Mask R-CNN based Auto-encoder. The proposed method consists of two components: an auto-encoder for image denoising and a Mask R-CNN for region-based segmentation. The auto-encoder comprises an encoder that encodes the image into a latent space representation and a decoder that reconstructs the image from the latent space. RGB images are utilized in the model, and the Mask R-CNN is integrated with the decoder of the auto-encoder. Furthermore, the "blind-spot" training technique known as Noise2Void is adopted for data augmentation in image segmentation. This technique encourages the neural network to learn how to denoise the image based on the available information instead of memorizing the training data. The simulation results demonstrate that the proposed Mask R-CNN approach achieves higher accuracy and prediction in agricultural mapping.