Precision agriculture is a new concept that has been introduced worldwide to increase production, reduce labor and ensure efficient management of fertilizers and irrigation processes. Computer vision is an essential component of precision agriculture and plays an important role in many agricultural tasks. It serves as a perceptual tool for the mechanical interface between robots and environments or sensed objects, as well as for many other tasks such as crop yield prediction. Another important consideration is that some vision applications must run on edge devices, which typically have very limited processing power and memory. Therefore, the computer vision models that are to run on edge devices must be optimized to achieve good performance. Due to the significant impact of Deep Learning and the advent of mobile devices with accelerators, there has been increased research in recent years on computer vision for general purpose applications that have the potential to increase the efficiency of precision agriculture tasks. This thesis explore how deep learning models running on edge devices are affected by optimizations, i.e., inference accuracy and inference time. Lightweight models for weed segmentation, peach fruit detection, and fruit disease classification are cases of studies. First, a case study of peach fruit detection with the well-known Faster R-CNN object detector using the breakthrough AlexNet Convolutional Neural Network (CNN) as the image feature extractor is performed. A detection accuracy of 0.90 was achieved using metric Average Precision (AP). The breakthrough AlexNet CNN is not an optimized model for use in mobile devices. To explore a lightweight model, a case study of peach fruit disease classification is next conducted using the MobineNet CNN. The MobileNet was trained on a small dataset of images of healthy, rotten, mouldy, and scabby peach fruit and achieved a performance of 0.96 F1. Lessons learned from this work led to using this model as a baseline CNN for other computer vision applications (e.g., fruit detection and weed segmentation). Next, a study was conducted on robotic weed control using an automated herbicide spot sprayer. The DeepLab semantic segmentation model with the MobileNet backbone was used to segment weeds and determine spatial coordinates for the mechanism. The model was optimized and deployed on the Jetson Nano device and integrated with the robotic vehicle to evaluate real-time performance. An inference time of 0.04 s was achieved, and the results obtained in this work provide insight into how the performance of the semantic segmentation model of plants and weeds degrades when the model is adapted through optimization for operation on edge devices. Finally, to extend the application of lightweight deep learning models and the use of edge devices and accelerators, the Single Shot Detector (SSD) was trained to detect peach fruit in three different varieties and was deployed in a Raspberry Pi device with an integrated Tensor Unity Processor (TPU) accelerator. Some variations of MobileNet as a backbone were explored to investigate the tradeoff between accuracy and inference time. MobileNetV1 yielded the best inference time with 21.01 Frame Per Second (FPS), while MobileDet achieved the best detection accuracy (88.2% AP). In addition, an image dataset of three peach cultivars from Portugal was developed and published. This thesis aims to contribute to future steps in the development of precision agriculture and agricultural robotics, especially when computer vision needs to be processed on small devices.

Este trabalho apresenta aspetos de Redes de Sensores sem Fios aplicadas à promoção da agicultura de precisão, inteligente e autónoma. O sistema de irrigação aqui desenvolvido com base em conceitos da Internet das Coisas (IoT) destina-se a combater os desperdícios, que ocorrem num modelo de rega tradicional e aprimorar a eficiência energética nas comunicações. Para desenvolver este sistema de irrigação foi utilizado o Raspberry Pi 4, juntamente com os módulos XBEE – PRO S1 da norma do Institute of Electrical and Electronic Engineering (IEEE) 802.15.4. Testaram-se os alcances de comunicação pretendidos. As medições da humidade do terreno, condições atmosféricas e horários apropriados de rega fornecem dados que são processados pelo sistema de monitorização, com o objetivo de realizar uma rega inteligente, autónoma, eficaz e sem desperdicios. Considerou-se a plataforma de interface gráfica Blynk para criar um painel que permite a visualização dos níveis de humidade e temperatura. A configuração do sistema de rega envolve a utilização de uma mini bomba submersível em água, de um sensor de temperatura e de um sensor de humidade do solo ligados a um Arduíno Uno. Este sensor comunica os valores da humidade para o Raspberry Pi. Os resultados subjacentes comprovam as vantagens da solução e mostram que a eficiência da comunicação entre nós XBEE é diferenciada dependendo da distância. Através da compação de dois modos de operação (Application Programming Interface, API, e modo transparente, AT) conclui-se que o modo AT é mais fácil de utilizar e económico, pois realiza configurações diretas e comunicações básicas.

A deteção de frutos é de fundamental importância em sistemas de estimação de produção. Neste trabalho, são apresentados os resultados preliminares da utilização do método de deteção de objetos Faster R-CNN na deteção de imagens de pêssegos. O estudo consiste na avaliação do desempenho do método em imagens RGB obtidas em ambiente real num pomar. Embora este método de deteção tenha sido aplicado noutros trabalhos com o objetivo de detetar frutos, ainda não foi utilizado na deteção de pêssegos. A cor, a sua distribuição na árvore e a clusterização são características intrínsecas aos pêssegos. Os resultados obtidos, ainda que preliminares, mostram um elevado potencial da utilização do método na deteção destes frutos. Todavia, os resultados também mostram a necessidade de melhoria no desempenho. Isso pode ser alcançado com o aumento na quantidade de imagens de treino e também por definir um melhor critério de anotação dos frutos oclusos.

Os mapas de produtividade refletem na maior parte das vezes a variação espacial e temporal dos diferentes fatores condicionadores do processo produtivo, sendo que o conhecimento de tais variações é importante na tomada de decisão desse mesmo processo produtivo. Neste estudo pretendemos identificar e caraterizar a variabilidade espacial e temporal (4 anos) da produtividade de milho, considerando para tal uma parcela regada por um Center Pivot e utilizando a técnica multivariada de análise de componentes principais. A aplicação da análise de componentes principais aos dados multitemporais da produtividade de milho permitiu perceber que: i) o 1º componente está relacionado com a produtividade média temporal, podendo a partir desta e dos respetivos loadings, perceber se a variabilidade temporal da produtividade média é baixa ou alta; ii) o 2º componente identifica as manchas da parcela com maior ou menor resiliência à variação temporal da produtividade; e iii) o 3º componente, fornece-nos pois a carta da estabilidade temporal da produtividade.The yield maps reflect in most cases the spatial and temporal variation of the different conditioning factors of the production process, so the knowledge of such variations is important in decision making of that same production process. In this study we intend to identify and characterize the spatial and temporal variability (over 4 years) of maize yield, considering for such a parcel irrigated by a center pivot and the multivariate technique of principal components analysis. The application of principal components analysis to a multitemporal data of maize productivity allowed to realize that: i) the 1st component is related to the average productivity of time and from this and the respective loadings we can see if the temporal variability of the average productivity is low or high; ii) the 2nd component identifies the portion of parcel more or less resilient to temporal variation, iii) the 3rd component provides us the maps of the temporal stability of productivity.

A wireless sensor network was used to monitor the light interception in pastures. The focusoftheworkwastheconstructionofalightmeterlowcostasaalternativetotheLICORLI-1776,high-costequipmentcurrentlyusedbyEmbrapa,andalsotheconstruction ofasensornodethatfunctionedproperlyinpastures,withouttheneedforskilledlaborto operateit. It have been analysed prototypes of luminosity meters of existing low cost and proposed ameter,withitscorrespondingsensor node , suitable low cost. This work presents theoretical and practical studies, implementing two prototypes of sensor nodesusingthephotoelectriccomponentphotodiodeS7686. Tests were performed and evaluated experimentally the performance of a wireless sensor network with star topology. In these tests, the proposed network showed the expected resultsandtheprototypedevelopedshowedsatisfactoryresultswithrespecttothecostof its production.

The next decades will be marked by the increase of the demand of foods and consumption of chemical fertilizers, mainly in the soils of low fertility of the Brazilian savannah. The precision agriculture (PA) involves soil sample tests and decisions in input application rates from computerized and automated machines. The soil fertility is fundamental for recommendation, and the soil sample tests are used as basic information. The study evaluated the variation sources in the process of soil samplimg, soil sample tests and soil fertility maps using kriging. The study considered the difference of two sampling methods in the vertical distribution of the nutrients at from 0 cm to 10 cm and 10 cm to 20 cm, it tested the number of soil samples necessary to make composed sample, it evaluated the effect of different instruments to collect soil sample tests and showed the variation found when sending the same sample for six differents laboratories. The field experiment was installed at 17o 46'54”S, 49o 28 ' 20”W in two different soils, a oxissol very loamy and a oxissol loamy. The field has 494,97 meters in length and 777,81 m of width and useful area 38,5 hectares, subdivided in 77 square cells with 70,71 m of a squared grid. In the center of each grid cell trenches of 50 cm of depth were open with aid of a mattock, in which the samples were collected in two depths, 00 to 20 cm and 20 to 40 cm, in all the 77 points of the grid. The standard method was the standard trench (STT). In eight points raffled in the grid, two samples were collected in the layers from 0 cm to 10 cm and 10 cm to 20 cm, to compare the detection capacity of recognize soil fetility surface concentration. In other eight raffled points were open more eight trenches, in a radius of ten meters from STT, in which samples were collected and mixed in a composed sample that was compared with the simple sample from STT. In more ten other points raffled in the grid the instruments used to collect samples were tested. The following instruments were tested: Big Dutch Auger (BDA); Small Dutch Auger (SDA); Thread Auger (THA); Probe Auger (PRA); Standard Trench (STT). In more three raffled points from the grid bigger samples were collected for sending the same sample for five different laboratories. Besides that the temporary variation of two laboratories was compared from Feb/2005 to Jun/2008. The results of all STT were analyzed by geoestatiscs programs and fertility maps were made from the calculation of the values of the nugget effect, sill and range of the samples. The area was was planted with soy, Glyvine max (L.), on 24/11/2006 and the crop was accomplished in 21/03/2007 in the same points of STT, where samples of the soy plants were taken for determination of the productivity of the area. The productivity presented pronounced regional effect, accompanying indexes integrators of the variations as the organic matter, the clay and pedologic components, and those components seem to be important in the definition of the income of the cultures and be probably the main factors for spatialy related to the productivity. The elements in an isolated way didn't show similar structure variation of the productivity, with variation in shorter distances revealing the need of an individualized site treatment of the elements, combinin sampling strategies with interpolation according to the behavior of the element of interest. Using PA concepts to collect soil samples to generate fertilty maps and applications maps with variable rates, a grid in the dimensions of the rehearsal should be made being respected the global tendency observed linked to the soil type, organic matter and clay, instead of that the average of all results it would avoid the high expenses with analyses. The concentration of nutritious available in the layer 0 cm to 10 cm it was not detected by the sampling in the layer 0 cm to 20 cm. The collection of simple samples for composition a composed sample is unnecessary procedure and showed no increase of precision in collection method. The instruments used in soil sample collections can cause variations significant and significant in cost of inputs, being factor important in variation process of diagnosis of soil fertility. The choice of the laboratory is what more contributed to the total variations found in the study of all other sources studied variations.

The deficiency in the management of agricultural areas causes problems of economic and environmental order. Many times the terraces dimensioning methods have results with reduced horizontal spacing, discouraged their adoption and causing erosion problems, and consequently loss of soil, water and nutrients. The soil loss calculation models can be used for the design of these structures, considering the displacement of soil lower than its natural formation rate, controlling the erosive process. However, the use of these models is hampered by the lack of data bases and tools that simplify their use. Concomitantly conservation planning can be integrated into agronomic management, enabling the optimization of the use of agricultural land and of the equipments, conciliating the economic and environmental aspects. In this sense the present study aimed to evaluate the cartographic quality of  altimetric surveys carried with GNSS RTK receiver onboard in quadricycle; develop a terrace in level spacing calculation method, using the soil loss calculated by USLE; a model of dimensioning agricultural systems, involving the use of terraces as rural roads, plots division, and generation and planting lines; analyze the accumulation of runoff on sugar cane lines for different furrow depths and the spatial variability of relief; and finally, develop a computational tool using the method and model. The altimetric data collected with quadricycle met the Class A pattern of cartographic accuracy, for the preparation of cartographic products with vertical equidistance of 0.30m. The terrace spacing calculation method reduced by up to 83.0% the total length of terraces; the agricultural systems dimensioning model enabled a reduction of up to 51.6% in the number of operations as well as increasing the average length of the lines and reducing the area for the use of rural roads, which increases the efficiency in the machinery use and in the use of agricultural area. The evaluation of the runoff volume accumulation in the planting lines due to the furrow depth found large storage capacity, reaching values of up to 158.1 m³ ha -1 , averaging 95.8 m³ ha -1 to furrows with 0.3m depth, resulting in a reduction of up to 10.1centimeters at the height of the terraces. The developed computational tool proved to be easy to use, and successfully completed the design of the plots, planting rows and terraces.

The evolution of precision agriculture increased agricultural production and provided major challenges in terms of efficiency, food safety, sustainability and environmental impact. Specifically, the pest management is extremely important in healthy crop. This work proposes a neural network architecture for detecting and counting the whitefly pest at different stages of nymphs and its adult life. Data are being considered in controlled environments through microscopic images and open environments in the field with sticky traps. Systematic review, data acquisition and analysis steps are completed, and results obtained so far point to a viability of a solution delivered in a software application and an intelligent trap to be used in the field.