Advancing Sustainability: Research Initiatives at the Signals and Images Lab

Abstract

partners. In this paper, we aim to briefly survey the relations of the work conducted at the Signals and Images Lab of CNR-ISTI, Pisa, with the themes of sustainability. We explore both the broader implications and the application-specific aspects of our work, highlighting references to published research and collaborative projects undertaken with key stakeholders and industrial ∗Corresponding author.

Keywords

• ment goals outlined by the United Nations
• surpassing

1. Introduction

The relationship between Sustainability and Artificial In

telligence (AI) is multifaceted. On the one hand, Artificial Intelligence can aid in addressing the challenges of modernization that may potentially conflict with sustainability goals. However, on the other hand, energy-intensive methods employed in artificial intelligence applications many paradigms and models within artificial intelligence are dificult to manage and maintain in a sustainable stationary nature of societal and technological changes.

In this context, the association with the principles of nal of diferent methods was devised, tested and validated; they can be tracked back mainly to general machine learning theory, computer vision and pervasive computing, topological data analysis and reliability analysis of observed data.

CEUR

ceur-ws.org b0 weak learners trained to overfit on their respective subsets. In the study, we explored the limits of ensemble size by utilizing only two weak learners. We found that this adaptive ensemble strategy remains eficient, even when extended to include up to five weak learners. Additionally, we identified potential avenues for further improvements, such as implementing various bagging strategies (e.g., training weak learners on subsets categorized by class dimensionality, clustering, or diferent colour space mappings of inputs). The basic idea of the method is depicted in Figure 1; in particular, ensembling is accomplished by an innovative strategy of performing bagging at the deep feature level. Namely, only the convolutional layers of each trained weak model are kept, while the final decisional layers are neglected; in this way, each weak model is turned into an extractor of deep features. The deep features of each weak model are then concatenated and fed to a trainable final decision layer.

These findings lay the groundwork for exploring similar strategies in various domains, such as Object Detection (by performing the ensemble at the feature extraction backbone level) and Segmentation (by conducting the ensemble on the encoding within typical encoderdecoder architectures).

Combination Layer Features1

Features2 Input

Input

Input

Output Module Features1

Feature Extractor

Output Module

Features2 Feature Extractor

2. Research topics

2.1. Adapting ensembling for green AI

In the realm of greenAI, computer vision stands out as a ifeld greatly benefiting from deep learning, continually advancing the state-of-the-art through the utilization of convolutional neural networks (CNNs) and visual transformers. Across various computer vision scenarios, in the last years, complexity has appeared to escalate exponentially, even for marginal enhancements, afecting 2.2. AI and sustainable agriculture both the number of parameters and of Floating Point The early detection of plant stress stands as a pivotal pracOperations (FLOPs). Among various machine learning tice in agriculture. Plant stress can be categorized into methodologies, ensembling emerges as a technique that biotic stress, caused by living organisms such as viruses, fuses multiple models, termed weak learners, to produce bacteria, fungi, nematodes, insects, arachnids, and weeds, a model with superior performance than any individ- and abiotic stress, resulting from environmental factors ual weak learner. Typically, this amalgamation involves like drought, heat, cold, strong winds, flooding, and nutriaggregating the weak learners’ outputs, often through ent deficiencies. Both forms of stress significantly impact voting or averaging for classification or regression tasks, crop yield and quality, leading to substantial economic respectively. Factors such as ensemble size (i.e., the num- losses when stress thresholds are surpassed. Despite adber of weak learners) and ensemble techniques (e.g., bag- vancements in genetics providing cultivars increasingly ging, boosting, stacking) play crucial roles in achieving resistant to various stresses, yield and quality losses resatisfactory results. However, ensembling necessitates main critical globally, particularly with the concurrent training several models, rendering the overall validation occurrence of abiotic and biotic stresses due to climate process more resource-intensive, with model complexity change. Presently, most plant inspections rely on mangrowing at least linearly relative to the ensemble size. ual methods involving direct visual analysis, which may Additionally, ensembling is time-consuming, thus posing not always facilitate accurately identifying diseases and a significant barrier to its widespread adoption, particu- stress types, especially in underdeveloped areas of the larly in computer vision applications. world. Farmers typically rely on naked-eye inspections,

Contrary to these challenges, in [ 1 ] we introduced a necessitating constant observation, high levels of skill, technique aimed at reversing the prevailing trend in im- and experience. While some farmers may utilize guideage classification, characterized by marginal performance lines containing basic concepts and visual aids (such as gains accompanied by substantial increases in complexity. pictures or notes) to distinguish between biotic and abiSpecifically, our approach involves a reimagined form otic injuries and determine appropriate solutions, others of ensembling designed to surpass the state-of-the-art may require technical assistance for formal and comprewhile maintaining constrained complexity in terms of hensive diagnoses. However, these methodologies are both parameter count and FLOPs. We demonstrated the often time-consuming and expensive, posing challenges feasibility of this approach by implementing bagging on for large farms or those with limited resources. Furtwo distinct subsets of data, employing two EficientNet- thermore, identifying weed types, whether broadleaf or grassy, during their early stages (from germination to the High Atlas watershed [ 7 ]. Furthermore, investigations development of the first four/six leaves) presents signifi- have been conducted to infer how excess irrigation water cant dificulties, precisely when intervention would be from rice cultivation can mitigate saltwater intrusion and most efective. Consequently, there is a growing need for potentially contribute to recharging the north of the Nile automated infection recognition methods to enhance ac- Delta Aquifer [ 8 ]. The findings indicate that strategically curacy and eficiency. The increasing public concern for placing rice cultivation and utilizing natural recharge environmental conservation and the imperative for more methods can significantly reduce aquifer vulnerability. eficient agriculture to accommodate population growth amidst shrinking arable land underscores the importance 2.4. Computer vision for medical waste of developing cost-efective and sustainable solutions to support farmers. In this context, machine learning tech- Medical waste (MW) poses a significant challenge as niques can potentially revolutionize the timely suppres- healthcare facilities generate considerable quantities of sion of harmful plant organisms, thereby maintaining hazardous waste daily. Proper handling and treatment chemical treatment and other interventions at econom- are essential, necessitating specialized management pracically and ecologically justified levels. In the research tices. Typically, medical waste is deposited into desigactivities at the lab, the methods detailed in Section 2.1 nated containers at the point of origin, which are subsehave been applied to the agricultural domain. Initially, quently sealed and require prompt disposal. Disposal state-of-the-art results were attained on the PlantVillage methods, primarily through incineration or waste-todataset [ 2 ]. Subsequently, these methods were leveraged energy processes, take place in specialized treatment fato form the intelligent core of the mobile application Gra- cilities often situated at a considerable distance from the noscan, developed within the Agrosat project and made point of waste generation. Moreover, due to the potential available by Barilla on popular iOS and Android plat- biological hazards, manual handling of the waste postforms [ 3 ]. Additionally, outstanding outcomes, coupled disposal is prohibited, with human intervention limited with a new dataset for weed classification, were docu- to critical situations and requiring specialized personal mented in [ 4 ]. Furthermore, an application targeting protective equipment. Recent years have seen a rise in Olive tree diseases, a vital sector in Italy’s agricultural medical waste production attributed to factors such as production, was presented in [ 5 ]. the expansion of the healthcare sector and increased utilization of disposable medical equipment. The emergence 2.3. Hydrological modelling for resilience of COVID-19 further exacerbated this trend, resulting in surges in the production of potentially hazardous and and equitable access to water infectious waste. Consequently, there is a pressing need Many basins in the Mediterranean and the MENA region for scalable and sustainable methods to manage medical (Middle-East North Africa) are currently facing recurring waste. Notably, medical devices are often composed of droughts and periods of water scarcity, often exacerbated high-quality virgin materials, predominantly polymers, by intermittent extreme events. Consequently, urgent alongside significant proportions of glass, textiles, and action is required to implement efective measures for metals. Therefore, sorting medical waste for recycling the sustainable management of water resources, with presents an opportunity for more sustainable waste mana particular emphasis on groundwater. In this context, agement practices. However, primary sorting is essential Artificial Groundwater Recharge (AGR) has emerged as a as manual intervention at later stages is impractical due viable solution for addressing water scarcity challenges. to the associated hazards. Secondary sorting may only By capturing and storing surplus water during periods occur after the infectious and hazardous nature of the of intense precipitation or increased surface water avail- medical waste has been neutralized, typically through ability, artificial recharge can replenish depleted aquifers sterilization. Manual primary sorting of medical waste and serve as a dependable water source during drought is laborious and time-intensive, potentially overburdenperiods. In collaboration with Ibn Tofail University (Ken- ing healthcare facility operators and increasing the risk itra, Morocco), considerable eforts have been directed of human error. Thus, there is a need for systems to towards enhancing the accuracy of hydrological models assist specialized staf during sorting, streamlining the and reducing uncertainty. This has involved incorporat- process and minimizing errors. To address this need, in ing various remote sensing-based Actual Evapotranspi- the framework of the project Medical Waste Treating ration (AET) products and assessing their efectiveness 4.0, in cooperation with industries and recycling experts, in improving the simulation of hydrological responses, we proposed research in the development of an artifiboth in single and multi-variable scenarios [ 6 ]. Simi- cial intelligence (AI) model leveraging deep learning and larly, the Soil and Water Assessment Tool (SWAT) model computer vision techniques to classify various types of and remotely sensed products have been incorporated to items comprising standard medical waste. Building upon characterize the dynamics of streamflow and snow in the previous work as reported in Section 2.1, we proposed a convolutional neural network based on the EficientNet family. To train the network, a specialized dataset mimicking a waste collection table equipped with a stereo camera has been curated (see Figure 2). While the dataset is still expanding, initial results indicate promising performance [ 9 ]. Unlike existing studies on medical waste, the dataset is made publicly available to foster further research [ 10 ].

2.5. Smart cameras for urban mobility

Recent advancements in remote sensing have proven to

In contemporary cities and urban environments, the sus- be highly efective in the realm of marine observation. tainability of mobility is posing growing concerns due This technology furnishes experts with an extensive array to several factors. These include the rise in the number of data gleaned from satellite sensors, necessitating the of cars on the roads, budget constraints leading to a re- development of automatic or semi-automatic methods for duction in free parking spaces, and economic challenges their analysis. In this context, one of the focuses of the SI during times of crisis, which diminish the eficiency of Lab lies in the categorization of upwelling regimes within public transportation systems. However, the increased the Iberia/Canary Current System (ICCS), an area that availability and power of computing facilities and de- remains relatively understudied within the domain of vices ofers an opportunity to address two of the most upwelling ecosystems. Of particular interest among the pressing issues: congestion and trafic resulting from various underlying processes are mesoscale events such vehicles searching for parking spaces. In our research, as upwellings, countercurrents, and filaments. These we have explored approaches centred around wireless events play a crucial role in transporting nutrient-rich smart cameras for monitoring open-air public parking waters from deeper regions to the surface, thereby sigareas and urban roads. More specifically, a smart camera nificantly influencing the biological parameters of the is a vision sensor equipped with artificial vision-logic for habitat and augmenting local biodiversity. Essentially, the on-board interpretation of acquired images and video there exists a correlation between the biogeochemical streams, along with a network interface to communicate and physical processes within a marine biological system. the processing outcome. A network of smart cameras is Sea Surface Temperature (SST), which measures the a sensor network whose nodes consist of a set of smart temperature of the water at the ocean’s surface, is a vicameras that cooperate for scene analysis, pervasively tal metric in this context. This temperature is assessed covering an area of interest. The primary strengths of the using satellite instruments that capture the energy emitproposed solution lie in the autonomy and scalability of ted from the ocean surface across diferent wavelengths the proposed architecture. Indeed, autonomy is achieved on a global scale. Subsequently, these data are corrobthrough the intelligent capabilities of smart cameras that orated with temperature readings obtained from ships can deploy powerful AI methods on board, ensuring inde- and buoys. The acquisition, processing, and comprehenpendence from external systems. Scalability is facilitated sion of sea surface data (see Figure 3) hold significant by low-cost single nodes that can be integrated into a importance as they provide insights into the underlying broader sensor network without requiring expensive in- reasons behind regional temperature variations and their stallation requirements. Starting with very low resource repercussions on the marine ecosystem. Environmental cameras leveraging classical image processing methods changes can have profound efects on the lives of species [ 11 ], approaches have moved to large-scale Field Opera- inhabiting these regions, influencing factors such as food tional Tests (FOT) [ 12 ], also integrating and comparing accessibility, migration patterns, and mating opportunimultimodal technologies [ 13 ] for parking lot monitoring; ties. To tackle the challenging task of categorizing marine a recent survey and way forwards are discussed in [ 14 ]. mesoscale patterns using remote sensing data, a pipeline of topological machine learning (TML) techniques has we addressed this gap by introducing an app founded on been used [ 15 ]. Such methods can be traced back to Al- a crowdsensing approach to enhance safety and awaregebraic Topology (AT), a branch of mathematics dealing ness at sea. The app seamlessly integrates into broader with shapes. In a nutshell, AT assigns an algebraic ob- systems and frameworks for environmental monitoring, ject to a topological space. Then, it can compute several aligning with our envisioned future endeavors that are invariants and descriptors (the Euler characteristic, the currently explored in the EU Funded Nautilos project. Betti numbers, and the homology groups). Such invari- Additional research eforts are being spent in identifying ants and descriptors are used in computational topology AI methods to derive information from satellite images to to compare shapes and to define distances between them. observe and categorize marine trafic, possibly leading to One of the most used tools from computational topology understanding patterns as well as in identifying malicious is persistent homology, which is a method for comput- behaviors, such as Illegal, Unreported, and Unregulated ing topological features of a space across diferent scales. (IUU) fishing or prohibited discharges [ 20 ]. In this conMore persistent features are detected over a wide range text, Synthetic Aperture Radar (SAR) images allow for of spatial resolutions (have a long “lifespan”) and are vessel detection under most weather conditions. Image deemed more likely to represent important, or true fea- processing and computer vision methods have been used tures of the underlying space, rather than artefacts of in the framework of the ESA-funded OSIRIS and OSIRISsampling, or noise. Applications of persistent homol- FO projects to (i) detect the presence of vessel targets in ogy span from computer vision and shape analysis to the input imagery, (ii) estimate the vessel types based on biomedical imaging and complex network analysis. In their geometric and scatterometric features, (iii) estimate this perspective, one of the most promising trends is the the vessel kinematics, and (iv) classify the navigation merging of persistent homology with machine and deep behavior of the vessel and predict its route [ 21 ]. Optical learning. In our work, TML is used to tackle the task images, although they can’t be acquired on cloudy days of categorizing four distinct mesoscale patterns using and at night, may convey fine-grained information suitremote sensing data—specifically patterns that can be able for the analysis and classification of small vessels, found in SST maps from the southwestern region of the potentially characterizing both industrial and small-scale Iberian Peninsula within the ICCS. Our initial investiga- fisheries, which is currently under consideration within tion attains a classification accuracy of 56% across the the PNRR NBFC Project. four labels [ 16 ]. These findings are promising, particularly given the presence of noise in the data and instances of low-quality or strong missing data (also due to weather 3. Conclusions conditions and cloudy sky).

2.7. AI for the marine environment: from citizen science to maritime trafic monitoring

The Sustainable Development Goals (SDGs) outlined by the United Nations encompass the marine and maritime domain, emphasizing the conservation and sustainable utilization of oceans, seas, and their resources. Concurrently, there is a pressing need for enhanced navigation safety, particularly in coastal regions. Presently, operational services leverage advanced technologies, including remote sensing and in situ monitoring networks, to aid navigation and environmental preservation eforts. However, the potential benefits of crowdsensing remain largely untapped. A citizen science approach to monitoring oil spills was proposed in [ 17 ] within the framework of a more comprehensive Maritime Information System (MIS), featuring proactive decision support services [ 18 ]. However, the potential of using Volunteered Geographic Information (VGI) and, more generally, crowdsensing information has not been fully exploited to address broader safety and security-related events. With this aim, in [ 19 ],

In this short paper, we provide an overview of the re

search conducted at SI Lab, focusing on themes related to sustainability. The breadth of applications showcased highlights the considerable value of AI as a tool to sensibly address the challenges of the future and facilitate ecological transition.

Acknowledgments

This work has been partially supported by the CNR FOE

Project DIT.AD022.207 “La Scienza per le TRansizioni Industriale, Verde, Energetica - STRIVE” and by the project SAC.AD002.014 “Accordo bilaterale Marocco-CNRST”. This work is part of a project that has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 101000825 – NAUTILOS Project.

Funder: ProFunder: Project funded under the National Recovery and Resilience Plan (NRRP), Mission 4 Component 2 Investment 1.4 - Call for tender No. 3138 of 16 December 2021, rectified by Decree n.3175 of 18 December 2021 of Italian Ministry of University and Research funded by the European Union – NextGenerationEU. Award Number: Project code CN_00000033, Con

References

[1] A. Bruno , D. Moroni , M. Martinelli , Eficient adaptive ensembling for image classification , Expert Systems ( 2023 ).

[2] A. Bruno , D. Moroni , R. Dainelli , L. Rocchi , S. Morelli , E. Ferrari, P. Toscano , M. Martinelli , Improving plant disease classification by adaptive minimal ensembling , Frontiers in Artificial Intelligence 5 ( 2022 ) 868926 .

[3] R. Dainelli , A. Bruno , M. Martinelli , D. Moroni , L. Rocchi , S. Morelli , E. Ferrari, M. Silvestri , S. Agostinelli , P. La Cava , et al., Granoscan: an ai-powered mobile app for in-field identification of biotic threats of wheat , Frontiers in Plant Science 15 ( 2024 ) 1298791 .

[4] R. Dainelli , M. Martinelli , A. Bruno , D. Moroni , S. Morelli , M. Silvestri , E. Ferrari , L. Rocchi , P. Toscano , Recognition of weeds in cereals using ai architecture , in: Precision agriculture'23 , Wageningen Academic Publishers, 2023 , pp. 401 - 407 .

[5] A. Bruno , D. Moroni , M. Martinelli , Eficient deep learning approach for olive disease classification , in: 2023 18th Conference on Computer Science and Intelligence Systems (FedCSIS) , IEEE, 2023 , pp. 889 - 894 .

[6] S. Taia , A. Scozzari , L. Erraioui , M. Kili , A. Mridekh , S. Haida , J. Chao , B. El Mansouri , Comparing the ability of diferent remotely sensed evapotranspiration products in enhancing hydrological model performance and reducing prediction uncertainty , Ecological Informatics 78 ( 2023 ) 102352 .

[7] S. Taia , L. Erraioui , Y. Arjdal , J. Chao , B. El Mansouri , A. Scozzari, The application of SWAT model and remotely sensed products to characterize the dynamic of streamflow and snow in a mountainous watershed in the High Atlas , Sensors 23 ( 2023 ) 1246 .

[8] I. Abd-Elaty , G. A. Sallam , L. Pugliese , A. M. Negm , S. Straface , A. Scozzari , A. Ahmed , Managing coastal aquifer salinity under sea level rise using rice cultivation recharge for sustainable land cover , Journal of Hydrology: Regional Studies 48 ( 2023 ) 101466 .

[9] A. Bruno , C. Caudai , G. R. Leone , M. Martinelli , D. Moroni , F. Crotti , Medical waste sorting: a computer vision approach for assisted primary sorting , in: 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW) , IEEE, 2023 , pp. 1 - 5 .

[10] A. Bruno , M. Martinelli , D. Moroni , Medical waste 4.0 dataset , 2022 . doi: 10 .5281/zenodo.7643417.

[11] M. Magrini , D. Moroni , G. Palazzese, G. Pieri, G. Leone , O. Salvetti , Computer vision on embedded sensors for trafic flow monitoring , in: Intelligent Transportation Systems (ITSC) , 2015 IEEE 18th International Conference on, IEEE , 2015 , pp. 161 - 166 .

[12] G. R. Leone , D. Moroni , G. Pieri, M. Petracca , O. Salvetti , A. Azzarà , F. Marino , An intelligent cooperative visual sensor network for urban mobility , Sensors 17 ( 2017 ) 2588 .

[13] M. Alam , D. Moroni , G. Pieri, M. Tampucci , M. Gomes , J. Fonseca , J. Ferreira , Real-time smart parking systems integration in distributed its for smart cities , Journal of Advanced Transportation 2018 ( 2018 ).

[14] C. Biyik , Z. Allam , G. Pieri , D. Moroni , M. O'fraifer, E. O'Connell , S. Olariu , M. Khalid , Smart parking systems: Reviewing the literature, architecture and ways forward , Smart Cities 4 ( 2021 ) 623 - 642 .

[15] F. Conti , D. Moroni , M. A. Pascali , A topological machine learning pipeline for classification , Mathematics 10 ( 2022 ) 3086 .

[16] F. Conti , O. Papini , D. Moroni , G. Pieri, M. Reggiannini , M. A. Pascali , Analysis of sea surface temperature maps via topological machine learning , in: 2023 IX International Conference on Information Technology and Nanotechnology (ITNT) , IEEE, 2023 , pp. 1 - 4 .

[17] M. Martinelli , D. Moroni , Volunteered geographic information for enhanced marine environment monitoring , Applied Sciences 8 ( 2018 ) 1743 .

[18] D. Moroni , G. Pieri, M. Tampucci , Environmental decision support systems for monitoring small scale oil spills: Existing solutions, best practices and current challenges , Journal of Marine Science and Engineering 7 ( 2019 ) 19 .

[19] D. Moroni , G. Pieri, M. Reggiannini , M. Tampucci , A mobile crowdsensing app for improved maritime security and awareness , in: 2022 IEEE International Conference on Pervasive Computing and Communications Workshops and other Afiliated Events (PerCom Workshops) , IEEE, 2022 , pp. 103 - 105 .

[20] M. Reggiannini , M. Righi , M. Tampucci , A. Lo Duca , C. Bacciu , L. Bedini , A. D'Errico , C. Di Paola , A. Marchetti , M. Martinelli , et al., Remote sensing for maritime prompt monitoring , Journal of Marine Science and Engineering 7 ( 2019 ) 202 .

[21] M. Reggiannini , E. Salerno , C. Bacciu , A. D'Errico , A. Lo Duca , A. Marchetti , M. Martinelli , C. Mercurio , A. Mistretta , M. Righi , et al., Remote sensing for maritime trafic understanding , Remote Sensing 16 ( 2024 ) 557 .