Convolutional Neural Network diseases detection of grapevines and UAV autonomous precision spray control

Nowadays, Precision Agriculture (PA) and Digital Agriculture (DA) are becoming fundamental instruments to oppose and prevent the upcoming agricultural sector crisis due to fertile soils scarcity, climate change, famine, lack of water and demographic expansion. Conventional individuation of crops pandemic clusters and pathological status rely on manual inspection, affected by high subjectivity as well as being costly and time wasting. Furthermore, intensive spraying of Plant Protection Products (PPP) has been for decades the unique method to ensure large-scale productions, with dramatic consequences in terms of eutrophication, soil toxicity and resources wasting. The combination of automated health status detection and automated precision spray allows to increase the soil productivity, to use fertilizer and pesticides only where is needed and drastically cut down the costs. This thesis presents an Unmanned Aerial Vehicles (UAV) implementation in disease recognition and precision aerial spraying of grapevines. A diseases recognition algorithm has been proposed, based on leaves images in the visible spectrum. Moreover, this algorithm is based on the transfer learning of MobileNet2, a Convolutional Neural Network (CNN), that has been pre-trained on the big image database ImageNet to classify one-thousand objects. The CNN is proposed as features extractor, then Linear Support Vector Machine (LSVM) and Logistic Regression (LR) are used for the final classification of grapevine leaves presenting black rot, esca, and leaf blight (Isariopsis leaf spot) symptoms. The outcome of the algorithm is a cartesian map providing information about the individual request of PPP and the relative plant position, called waypoints. A path planning routine, aimed at maximizing the autonomy of the quadrotor, was designed to solve a modified Travelling Salesman Problem finding the optimal sequence of waypoints, using Artificial Intelligence (AI). The solver minimizes both the travel distance and the time-averaged carried payload, by means of a Genetic Algorithm (GA). A Linear Quadratic Regulator (LQR) has been proposed, with yaw-gain scheduling to control a 25 kg quadrotor with a tank of 10 L, and an Extended Kalman Filter (EKF) as state estimator. A sensor-free wind estimation algorithm was developed, allowing a local estimate of wind magnitude and direction. Furthermore, experimental relations between nozzle pressure, rotors downwash, air-UAV relative velocity and the drift entity of a hollow-cone nozzle were retrieved by image analysis. Intersecting wind data with drift experimental curves, was then developed an algorithm to correct the UAV path in presence of wind to maximize some irroration quality factors.