Using AI to Estimate Height of Plants through Surveillance Cameras at an Industrial Scale : CNNs on Basil Plants with Robel Poles

University essay from KTH/Skolan för elektroteknik och datavetenskap (EECS)

Abstract: This report presents the results of investigations into whether, and how well, Artificially Intelligent (AI) algorithms can be used to estimate the height of plants by using images from regular surveillance cameras, setup over one of Svegros basil farms. The project is of great economical importance as too tall basil plants will not fit the shelves at stores and too small plants will disappoint customers. This is a part of a bigger movement at Svegro to automate the monitoring and caring for the growing plants, aiming at lowering energy consumption and minimizing waste. To measure the heights, rulers (Robel poles) were placed behind the plants that moved on conveyor belts under cameras so the plants’ heights could manually be established from the number of visible lines on the Robel pole, not covered by the plant. The research problem was to engineer an AI based solution to predict how many lines were visible above the plant. After two months of gathering images and manually annotating them, three Convolutional Neural Network (CNN) models of varying complexity were trained on the images of individual Robel poles from the basil field. Results obtained with Grad-CAM showed that the networks do not learn to count the lines but to correlate the leafs size and shape to the height. The best score was a Mean Absolute Error of 0.74 and a Mean Squared Error of 0.83, where a MAE of 2.53 and MSE of 11.11 corresponded to just predicting the data sets median. This was achieved with EfficientNet0B. The results were compared with a human being’s performance which showed that the human still performed better but due to the noisy data, the results are considered impressive and the score exceeded the expectations of the team at Svegro so the final model is now used there today. It was also shown that reasonably good results could be obtained even without the Robel pole in the training images, meaning the Svegro team could stop setting out the Robel poles but with a slight loss in precision. Suggestions for improvements, like changing the design of the Robel poles, are presented to aid future research to fully automate the process with higher accuracy. 

  AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)