Combining Machine Learning with Computer Vision for Precision Agriculture Applications

Abstract

Financial and social elements of modern societies are closely connected to the cultivation of corn. Due to its massive production, deficiencies during the cultivation process directly translate to major financial losses. Existing field monitoring solutions utilize aerial and ground means towards identifying sectors of the farmland presenting under-performing crops. Nevertheless, an inference element is still absent; that is the automated diagnose of the cause and severity of the deficiency. The early detection and treatment of crops deficiencies and the frequent evaluation of their growth status are thus tasks of great significance. Towards an automated health condition assessment, this thesis introduces schemes for the computation of plant health indices. First, we propose a methodology to detect nitrogen (N) deficiencies in corn fields and assess their severity at an early stage using low-cost RGB sensors. The introduced methodology is twofold. First, a low complexity recommendation scheme identifies candidate plants exhibiting nitrogen deficiency and second, a detection elimination step completes the inference loop by deciding which of the candidate plants are actually exhibiting that condition. Experimental results on a diverse real-world dataset achieve a 90.6% accuracy for the detection of N-deficient regions and support the extension of this methodology to other crops and deficiencies that show similar visual characteristics. Second, based on the 3D reconstruction of small batches of corn plants at growth stages between ''V3'' and ''V6'', an automated alternative to existing manual and cumbersome phenotype estimation methodologies is presented. The use of 3D models provides an elevated information content, when compared to planar methods, mainly due to the alleviation of leaf occlusions. High-resolution images of corn stalks are collected and used to obtain 3D models of plants of interest. Based on the extracted 3D point clouds, the calculation of a plethora of phenotypic characteristics for each 3D reconstruction are obtained such as the number of plants depicted with 88.1% accuracy, Leaf Area Index (LAI) with 92.48% accuracy, the height with 89.2% accuracy, the leaf length with 74.8% accuracy, and the location and the angles of leaves with respect to the stem. The last two variables are connected by showing the trend of the angles to change with respect to the leaf position on the stem as the crops grow. An experimental validation using both artificially made corn plants emulating real-world scenarios and real corn plants in different growth stages supports the efficacy of the proposed methodology. Although the proposed methodologies are agnostic to the platform that performs the data collection, for the presented experiments a MikroKopter Okto XL equipped with a Nikon D7200 RGB sensor and a DJI Matrice 100 with a Zenmuse X3 and a Zenmuze Z3 RGB high-resolution cameras were used. The flight altitude ranged between 6 and 15 m and the resolution of the images varies within a range of 0.2 to 0.47 cm/pixel. Thorough data collection and interpretation leads to a better understanding of the needs not only of the farm as a whole but to each individual plant providing a much higher granularity to potential treatment strategies. Through the thoughtful utilization of modern computer vision techniques, it is possible to achieve positive financial and environmental results for these tasks. The conclusions of this work, suggest a fully automated scheme for information gathering in modern farms capable of replacing current labor-intensive procedures, thus greatly impacting the timely detection of crop deficiencies.