Research and implementation of robotic, digital and A.I. solutions for the management of agricultural and urban green spaces

In the context of the ongoing digital and ecological transition promoted by European strategies, agriculture and green area management are increasingly required to adopt sustainable and technologically advanced approaches. The landscape results from the interaction of natural, cultural, anthropogenic, and productive factors, making it a complex system to manage and preserve to enhance the resilience of agroecosystems. Modern agroecosystems management face the dual challenge of reducing chemical inputs, as ones usually employed for weed control, and resource consumption while maintaining productivity and enhancing biodiversity. For all these reasons it is necessary to identify and integrate different supporting strategies and technologies to achieve a real digital transition in agriculture and green area management. In this framework, technologies such as robotics, digital imaging, and artificial intelligence (AI) can play a crucial role in optimizing maintenance operations, improving resource efficiency, and promoting biodiversity conservation. 9 trials were conducted addressing different aspects of digital and robotic innovation applied to agroecosystem management, structured into three main thematic areas: benchmarking and validation of agricultural robots, management and operational performance of robotic systems, ecological and biodiversity-related impacts of robotic management. 1. The design of the field campaign of ACRE competition 2023 was first conceived and then realized, with a special focus on the timeline and on all the necessary operations to set up a fitting experimental design for autonomous weeding robots performance evaluation based on objectives benchmarks: functionality BenchMark (FBM) of plant discrimination and field navigation. The test field designated for the campaign covered an area of 2.8 hectares and it was divided into 35 distinct plots, each serving its specific purpose. These plots were sown with three different crops:18 rows of corn (Zea Mays L.), 16 rows of bean (Phaseolus vulgaris L.), and a larger plot of beet (Beta vulgaris L.). The two main crops were sown along rows with different distances between them, 75 cm for the corn and 37.5 the bean. Specifically, three shifting rows were in the bean plots, while two were observed in the corn plots. The same number of crop plants in each row, with the same space between each other, was chosen to ensure the same conditions for all participants. To facilitate manoeuvres and ensure clear demarcations between plots, designated corridors were intentionally left between them. Manual weeding efforts were carried out in each plot, except for the B. Vulgaris L. plots. It is worth noting that existing Lolium perenne L. plants were intentionally left untouched within the P. vulgaris L. plots. Furthermore, as part of the experimental design, 15,000 weed plants encompassing three different species (5,000 for each species) were manually transplanted. The selected species were wild mustard (Sinapis arvensis L.), ryegrass (Lolium perenne L.), and chamomile (Matricaria chamomilla L.). The weeds were selected to provide different patterns and shapes. During the competition a mobile rover developed by Agilex was employed to perform the functionality BenchMarks and a large amount of RGBD images were collected. 2. In light of the anticipated surge in the global population to 9-10 billion by 2050, there is a pressing need for agricultural production methods that are both efficient and environmentally sustainable. To help meet these challenging demands, Artificial Intelligence and Robotics technologies are being increasingly adopted in the agricultural sector to support precision agriculture tasks. This work focuses on the specific task of autonomous weeding, which requires expertise at the intersection of Computer Vision and Agronomy. One key prerequisite to autonomous weeding is the capability to segment weeds and crops from robot-collected images. However, robust segmentation performance is typically achieved, in Computer Vision, by relying on extensive and densely labelled datasets whose creation can be both expensive and time-intensive. Common detectors like YOLO perfectly exemplify this paradigm. By contrast, Few-Shot Learning (FSL) methods can learn from minimal annotated examples and drastically mitigate the costs and challenges associated with acquiring large datasets. In this paper, we show that HDMNet, a cost-effective FSL architecture, can already ensure 73 to 80% of the upper-bound performance of data-hungry YOLOv5 and YOLOv8 models for the detection of Bean and Corn crop regions. Crucially, these results can be achieved with as few as one support example, effectively zeroing out the data annotation cost. In the lack of reference data on the human labour required to collect expensive annotations that require the knowledge of experts in agronomy, this paper provides a novel analysis of human labour costs associated with agricultural dataset curation that is grounded in the real use-case of labelling robot-collected images with crop and weed regions. As such, the paper offers a layered analysis of the trade-offs between human effort, dataset size, label quality, and prediction accuracy when designing and building datasets for agricultural image analysis. Quantitative performance and cost results are complemented by a qualitative analysis of common error causes in the model predictions, leading to partial crop and weed region coverage mainly due to the incidence of small and sparse weed regions, crops overshadowing small-sized weeds, as well as insufficiently granular ground truth annotations generated via semi-autonomous pipelines (Segment Anything). Concurrently, we curate and contribute a comprehensive dataset for crop and weed segmentation, the 2023 ACRE Competition dataset, which includes an 'Early Dataset' and a densely annotated 'Refined Dataset', acquired under different geographical, soil, robot configuration and lighting conditions that previously released datasets in the literature. These novel contributions guide the development of weed-crop segmentation solutions that are both high-performance and cost-effective. 3. Sustainable turfgrass management is essential for maintaining healthy and visually appealing green spaces. Autonomous mowers have emerged as an innovative solution, but the efficiency and quality of mowing operations depend on several factors. This study investigates the impact of mowing patterns and cutting heights on the performance of an autonomous mower through updated custom-built software. Three different mowing patterns (vertical, diagonal, and horizontal) and two cutting heights (3 cm and 6 cm) were analysed to analyse mowing efficiency, coverage, and cutting uniformity. The vertical pattern emerged as the most effective, maximizing speed (0.52 m s-1) and efficiency (0.77), while minimizing overlap (4.27 cm) and uncut areas (0.014 m2). In contrast, the horizontal and diagonal patterns showed lower efficiency (0.71 and 0.76) and less coverage percentage (97.05% and 96.71%) compared to the vertical pattern (98.57%). Cutting height influenced performance, with higher heights sometimes requiring adjustments to prevent inefficiencies. The interaction between the mowing pattern and cutting height was critical for optimizing both operational efficiency and cutting quality. These findings highlight the importance of selecting an appropriate mowing pattern and cutting height tailored to the specific operational goals. 4. Turfgrasses deliver essential ecosystem services, including soil protection, temperature mitigation, and aesthetic enhancement of green areas. Mowing is a key practice to ensure these benefits, particularly when performed in a proper way, as with autonomous mowers. However, mowing patterns and the resulting trampling can influence the visual quality of turfgrass, especially through overlapping and uncut areas caused by turn trajectories. This study aimed to validate the measurement accuracy of a new function in a custom-built software (v2.5.0.0) designed to analyse mower trajectories via RTK-GPS data, and to assess the influence of three mowing patterns (vertical, horizontal, diagonal) on overlapping and uncut areas. The software outputs were validated against two manual tracking methods: chalk powder and wire. The experiment was conducted on a mature Bermudagrass stand, and trajectory data were analysed using ANOVA. Results showed that the measurement method significantly affected straight trajectory overlap, with the wire method underestimating overlap (2.36 cm) due to field-related operational limits. In contrast, chalk powder (3.72 cm) and software (4.71 cm) produced comparable results. Significant differences were also observed in turn trajectory no-cut areas, especially with the diagonal mowing pattern, which generated the largest uncut zones (0.059 m2). 5. In the framework of the ongoing digital and ecological transition in agriculture, this trial was designed to assess the real applicability of autonomous robotic systems for vineyard management. By comparing a traditional tractor-based approach with an autonomous robot for under row weed control, the study aims to provide objective data on performance, environmental sustainability, operational efficiency and evaluate the feasibility of the autonomous approach through Business Process Model and Notation BPMN-CBA approach. The evaluation of both systems focused on several key performance indicators. The quality and efficacy of the intervention were assessed through pre- and post-operation measurements of weed biomass (g d.m. m−2), coverage (%), height (cm), and floristic composition (species richness, Shannon diversity index and evenness), considering both the inter-vine spaces and the areas immediately surrounding the trunks. Soil compaction was analysed. The working width of the under-row portion managed by the robot and the crop damage rate, expressed as the percentage of injured plants per row, were also recorded. Operational efficiency was analysed by measuring the total working time, the energy consumption, and the weeding possibility zone. Environmental performance was quantified through CO₂ emissions (kg ha⁻¹) and comparative energy-use analysis between the robotic and tractor-based systems. Finally, working capacity (ha h⁻¹) and the degree of automation were evaluated, considering the time spent by operators for planning, supervision, and manual control, providing an integrated understanding of the technological and operational sustainability of each management approach. All the data collected were used to perform the BPMN-CBA analysis. Preliminary results show no difference in terms of weed control efficacy, compaction and in terms of weeding possibility zone. The BPMN-CBA highlight a positive NPV for the adoption of autonomous solution on an area of 50 ha. 6. Robotic solutions and technological advances for turf management demonstrated excellent results in terms of quality, energy, and time consumption. Two battery-powered autonomous mowers (2 WD and 4 WD) with random patterns were evaluated according to different trampling levels (control, low, medium, high) on a typical warm season turfgrass at University of Pisa, Italy. Data on the percentage of area mowed, the distance travelled, the number of passages, and the number of intersections were collected through RTK devices and processed by a custom-built software (1.8.0.0). The main quality parameters of the turfgrass were also analysed by visual and instrumental assessments. Soil penetration resistance was measured through a digital penetrometer. The efficiency significantly decreased as the trampling level increased (from 0.29 to 0.11). The over-trampled areas were mainly detected by the edges (on average for the medium level: 18 passages for the edges vs. 14 in the central area). The trampling activity caused a reduction in turf height (from about 2.2 cm to about 1.5 cm). The energy consumption was low and varied from 0.0047 to 0.048 kWh per cutting session. Results from this trial demonstrated suitable quality for a residential turf of the Mediterranean area (NDVI values from 0.5 to 0.6), despite the over-trampling activity. Soil penetration data were low due to the reduced weight of the machines, but slightly higher for the 4 WD model (at 5 cm of depth, about 802 kPa vs. 670 kPa). 7. Autonomous mowers navigation pattern plays a crucial role in turfgrass quality, influencing both aesthetic and functional performance. This study aimed to evaluate the impact of three different autonomous mower navigation patterns (random, vertical, and chessboard) on operational performance and the effect of trampling activity on turfgrass. Each pattern was tested in terms of data on number of passages, distance travelled (m), number of intersections and percentage of area mowed using a remote sensing system and an updated custom-built software. Green coverage percentage was assessed weekly using image analysis (Canopeo app) to evaluate the turfgrass quality. The green coverage percentage, together with the number of passages, are analysed and correlated. The random pattern generated the highest number of passages and intersections, leading to lower average green coverage (64%) compared with the chessboard (80%) and vertical (81%) patterns. The effective number of passages to reach 60% green cover (EP60) was 56, 87, and 155 for random, vertical, and chessboard, respectively. Future studies should extend this approach to other species and environmental conditions, integrating the effective dose (in terms of passages) method into decision-support systems for smart mowing management. 8. The quality of sports facilities, especially football pitches, has gained significant attention due to the growing importance of sports globally. This study examines the effect of two different cutting systems, a traditional ride-on mower and an autonomous mower, on the quality and functional parameters of a municipal football field. The analysis includes visual assessments, measurements of grass height, and evaluations of surface hardness, comparing the performance of the two cutting systems. Additionally, studies of turfgrass composition and machine learning techniques, particularly with YOLOv8s and YOLOv8n, are conducted to test the capability of assessing weed and turfgrass species distribution. The results indicate significant differences in grass colour based on the position (5.36 in the corners and 3.69 in the central area) and surface hardness between areas managed with a traditional ride-on mower (15.25 Gmax) and an autonomous mower (10.15 Gmax) in the central region. Higher height values are recorded in the area managed with the ride-on mower (2.94 cm) than with the autonomous mower (2.61 cm). Weed presence varies significantly between the two cutting systems, with the autonomous mower demonstrating higher weed coverage in the corners (17.5%). Higher overall performance metrics were obtained through YOLOv8s. This study underscores the importance of innovative management practices and monitoring techniques in optimizing the quality and playability of a football field while minimizing environmental impact and management efforts. 9. Planning and management of green spaces play a key role for the biodiversity conservation and enhancement in urban context. Among these green areas, lawns can provide different ecosystem services including the biodiversity promotion. Furthermore, residential, and public lawns are crucial for bees and other pollinator populations. However, lawn management practices, especially intensive mowing, can negatively affect plants diversity and abundance disrupting pollinators activity. Since the many limits connected to the use of traditional mowing machines autonomous mowers are becoming very popular in both public and private green areas. Nevertheless, very little is known about their influence on biodiversity, and particularly on pollinators. In this perspective the use of Phyla nodiflora, a creeping plant species frequently visited by pollinators, obtained positive results when managed with an autonomous mower. In this study, P. nodiflora plants are transplanted in spots (8 spots for each plot) within a turf mainly composed of bermudagrass (Cynodon dactylon x transvaalensis “Patriot”) with the aim of promoting greater floral diversity and the conservation of pollinators. The research aim is to evaluate two different autonomous mowers navigation systems (random vs systematic) and two different cutting height (2 cm vs 6 cm) in terms of the effect on P. nodiflora. Phyla nodiflora flowers height, number of Phyla nodiflora flowers, Phyla nodiflora coverage per plot and turfgrass height were assessed. Furthermore, the development of a script enabling automatic flower counting was completed; the reliability and consistency of automatic counts was assessed. Data on monitoring of pollinators were used to calculate the family richness, the Shannon diversity index and the Evenness. The treatment, the survey data and their interaction had a significant effect (p<0.001) on flower height, number of flowers and coverage percentage. Highest values of flower height were recorded in treatment with systematic pattern, the highest one in treatment 2 (6.08). The same trend was recorded for others parameters with highest value in treatment with systematic pattern (i.e. treatment 1 with 29.46 number of flowers and 6.23 % of coverge). Regarding the automatic flower counting YOLOv8s and YOLOv11s exhibited better performance respect to the nano models, but, overall YOLOv8s represent a preferable choice. The family richness exhibited the highest value in treatment 2, the evenness in the treatment 4.