Self-supervised representation learning framework for remote crop monitoring using sparse autoencoder

Abstract

Remote crop monitoring is one of the emerging technologies required for precision agriculture. The advent of remote monitoring techniques accumulates huge amount of real-time image data in cloud storage. Preserving this big data without suppressing its significant details is important for the further crop investigation process. Feature learning with autoencoder helps in learning significant data without compromising the original dataset variance. This work aims to develop a sparse autoencoder model to obtain high abstract level features from the image such that the original image could be reconstructed from the reduced feature with minimum error. Sparse autoencoder is an unsupervised back propagation neural network with sparse network connections that converts high-dimensional image into low-dimensional features. The performance of autoencoder with respect to image reconstruction is evaluated in terms of MSE and PSNR. Also, the effect of dimensionality reduction is qualitatively analyzed with PCA plots that confirms the reduced dataset maintain the expanded variance required for further crop investigations.