Visual harvest: self-supervised learning for lettuce growth analysis

Accurately estimating lettuce growth parameters is crucial for optimizing vertical farming systems, yet existing methods often rely on labour-intensive manual measurements or labelled data sets which can be scarce and costly to acquire. In this work, we propose a novel approach that leverages self-supervised learning techniques to estimate lettuce growth parameters (dry weight, fresh weight, height, diameter, and leaf area) using image data collected throughout the plant’s growth cycle. Our methodology consists of a two-part pipeline. First, we implement a self-supervised pre-training step using unlabelled lettuce images obtained at different weeks since seeding. The second part involves fine-tuning the learned weights of a ResNet18 architecture (from the self-supervised pre-training step) on a smaller labelled lettuce data set for the regression of the 5 growth parameters. We adapt and extend two popular self-supervised learning algorithms, plantSimCLR and plantBT, tailored specifically for plant imagery. Firstly, we propose SimCLR for plants (plantSimCLR) by creating positive and negative pairs based on the time elapsed since seeding. Secondly, we introduce BarlowTwins for plants (plantBT) by applying the redundancy reduction principle to self-supervision. We apply random spatial transformations to the lettuce images to obtain two distorted versions of the original image. The self-supervised pre-training task promotes the representations of the distorted lettuce versions to be close to each other using either contrastive learning (plantSimCLR) or cross-correlation (plantBT). We evaluate the quality of the learned representation against ImageNet pre-trained weights. Our evaluation demonstrates that both plantSimCLR and plantBT provide a more effective starting point for estimating the lettuce growth parameter. Additionally, our method significantly decreases the time and effort needed for manually measuring the lettuce growing characteristics. By leveraging self-supervised learning techniques tailored to plant imagery, our work offers a promising avenue for advancing automated monitoring and optimization of vertical farming systems, ultimately contributing to sustainable and efficient agricultural practices.