Soft decision making for agri-food 4.0
Kuupäev
2024-09-26
Autorid
Ajakirja pealkiri
Ajakirja ISSN
Köite pealkiri
Kirjastaja
Abstrakt
Väitekiri esitab ulatusliku uurimuse uuenduslikest, andmepõhistest lähenemistest, mis käsitlevad jätkusuutliku põllumajanduse kriitilisi väljakutseid, eriti Agri-Food 4.0 kujunemisjärgus paradigmas. Uurimus keskendub masinõppe (ML) mudelite täpsuse, efektiivsuse ja tõlgendatavuse parandamisele, et ennustada olulisi põllumajanduslikke parameetreid, nagu evapotranspiratsioon, mulla orgaaniline süsinik (SOC), Ex-ante elutsükli hindamine nisu tootmisel ja põllukultuuride haiguste tuvastamine (roostetriibu nisuhaiguse tuvastamine) kasutades pilte, eesmärgiga panustada jätkusuutlike ja vastupidavate agri-toidusüsteemide arengusse.
Uurimistöö Eesmärgid
Põllumajandus seisab silmitsi kasvava survega, mis tuleneb kliimamuutustest, ressursside piirangutest ja toiduga kindlustatuse vajaduse suurenemisest. Need väljakutsed nõuavad arenenud otsustustööriistu, mis optimeerivad ressursikasutust ja parandavad saagikust, säilitades samal ajal keskkonna jätkusuutlikkuse. Agri-Food 4.0 kontseptsioon hõlmab selliste tehnoloogiate integreerimist nagu asjade internet (IoT), pilvearvutus ja tehisintellekt (AI), et revolutsiooniliselt muuta põllumajandust. See väitekiri on kooskõlas nende eesmärkidega, pakkudes välja tõlgendatavad ja jätkusuutlikud otsustamistööriistad, mis käsitlevad nii tõhusust kui ka läbipaistvust, mis on vajalikud põllumajanduse otsustustugisüsteemides (DSS).
Uurimistöö eesmärgid on:
Arendada tõlgendatavaid masinõppemudeleid, mis tasakaalustavad täpsust ja tõlgendatavust, mis on olulised põllumajanduse sidusrühmadele.
Parandada põllumajandustavade jätkusuutlikkust, integreerides masinõppemudelid valdkonnapõhiste teadmistega.
Rakendada automatiseeritud masinõppe (AutoML) tehnikaid põllukultuuride haiguste tuvastamise optimeerimiseks, keskendudes konkreetselt nisuhaiguste tuvastamisele.
Mulla Orgaanilise Süsiniku (SOC) Ennustamine
Mulla orgaaniline süsinik (SOC) on mulla tervise oluline komponent, mis mõjutab selle struktuuri, veemahutavust ja toitainete kättesaadavust, mis kõik on olulised jätkusuutliku põllukultuuride tootmise jaoks. SOC on ka oluline tegur ülemaailmses süsinikuringes, kuna mullad on suurimad maismaa süsiniku reservuaarid. Väikesed muutused SOC tasemetes võivad märkimisväärselt mõjutada atmosfääri CO2 kontsentratsioone, mõjutades seega kliimamuutusi.
Väitekiri uurib erinevaid masinõppe tehnikaid SOC tasemete täpsemaks ja tõhusamaks ennustamiseks. Traditsioonilised SOC hindamise meetodid, mis sageli hõlmavad keerulist ja töömahukat proovide võtmist ja analüüsi, on täiustatud andmepõhiste mudelitega nagu ANFIS-T (Adaptive Neuro-Fuzzy Inference System koos murdosatikhhonovi regulatsiooniga). Uurimistöö näitab, et ANFIS-T mitte ainult ei paranda ennustuste täpsust, vaid säilitab ka kõrge tõlgendatavuse taseme, mis on praktilises põllumajanduses oluline. See on eriti oluline, kuna see on kooskõlas ülemaailmsete jõupingutustega, nagu EL-i mullastrateegia aastaks 2030, mis rõhutab SOC suurendamist kasvuhoonegaaside vähendamise eesmärkide täitmiseks.
Evapotranspiratsiooni Ennustamine
Evapotranspiratsioon (ET) on vee aurustumise ja taimede transpiratsiooni summa. See on kriitiline parameeter veeressursside juhtimisel, eriti põllumajanduses, kus tõhusad niisutusstrateegiad on vajalikud vee kasutamise optimeerimiseks. Põllumajandus moodustab märkimisväärse osa ülemaailmsest veetarbimisest, kus ET esindab suurimat osa veekaost põllumajandussüsteemides.
Traditsioonilised ET mõõtmise ja hindamise meetodid, nagu lüsimeetrid ja evaporimeetrid, on sageli piiratud nende kõrge maksumuse ja raskustega nende ulatuslikul kasutuselevõtul. Väitekiri hindab masinõppemudelite, eriti ANFIS-T, kasutamist ET ennustamisel. Need mudelid pakuvad täpseid ennustusi, olles samas piisavalt tõlgendatavad, et neid saaksid kasutada põllumehed ja põllumajandusjuhid teadlike otsuste tegemiseks niisutuspraktikate kohta. Parandades ET ennustuste täpsust, aitab uurimistöö kaasa põllumajanduse jätkusuutlikumale veekasutusele, mis on kliimamuutuste ja veepuuduse kontekstis üha olulisem.
Ex-Ante Elutsükli Hindamine (LCA) Nisu Tootmisel
Elutsükli hindamine (LCA) on kõikehõlmav metoodika toodete keskkonnamõjude hindamiseks alates tooraine kaevandamisest kuni tootmise, kasutamise ja kõrvaldamiseni. Traditsioonilised LCA-d viiakse tavaliselt läbi pärast toote elutsükli lõppu (ex-post), piirates nende kasulikkust proaktiivsete ja jätkusuutlike otsuste tegemisel.
Väitekiri edendab ex-ante LCA kasutamist, mis hindab võimalikke keskkonnamõjusid teadus- ja arendustegevuse etapis, võimaldades varasemaid sekkumisi ja jätkusuutlikumaid tootedisainilahendusi. Nisu tootmise kontekstis integreerib väitekiri masinõppe LCA-ga, et ennustada ja leevendada keskkonnamõjusid, nagu kasvuhoonegaaside heitkogused ja ressursikasutus. See lähenemisviis on eriti kooskõlas Euroopa Liidu rohelise kokkuleppega, mille eesmärk on vähendada süsinikuheidet kõigis sektorites, sealhulgas põllumajanduses. ML-i integreerimine LCA-sse mitte ainult ei paranda LCA ennustamisvõimekust, vaid muudab selle ka tõlgendatavamaks ja sidusrühmadele rakendatavaks.
Põllukultuuride Haiguste Tuvastamine
Põllukultuuride haigused, eriti need, mis mõjutavad põhikultuure nagu nisu, kujutavad endast märkimisväärseid ohte ülemaailmsele toiduga kindlustatusele. Nisu roostetriibuhaigus, mille põhjustajaks on seenpatogeen Puccinia striiformis, on üks kõige laastavamaid nisukultuuri mõjutavaid haigusi. Traditsioonilised haiguste tuvastamise meetodid, mis tuginevad suures osas käsitsi inspekteerimisele, ei ole ulatuslikud ja on vigadele altid.
Väitekiri pakub välja uue AutoML-i raamistiku, mis hõlmab kontekstitundlikku tunnusjoonte inseneeriat, et parandada nisu roostetriibuhaiguse tuvastamise täpsust ja efektiivsust. Kasutades UAV-ga kogutud pilte ja arenenud pilditöötlusvõtteid, ületab raamistik täpsuse ja arvutusliku efektiivsuse osas traditsioonilised süvaõppemudelid nagu ResNet-18. See lähenemisviis mitte ainult ei lihtsusta haiguste tuvastamise protsessi, vaid suurendab ka jälgimissüsteemide mastaapsust, muutes nende rakendamise suurtele põllumajanduslikele aladele teostatavaks.
Kokkuvõte
Väitekiri järeldab, et arenenud masinõppetehnikate integreerimine valdkonnapõhiste teadmistega parandab oluliselt võimet käsitleda põllumajanduse võtmeprobleeme. Parandades nende mudelite tõlgendatavust, täpsust ja efektiivsust, panustab uurimistöö jätkusuutlikumate ja vastupidavamate agri-toidusüsteemide arengusse. Tulevased uurimused võiksid uurida nende tehnikate laiemat rakendamist erinevatel põllukultuuridel ja keskkondades, täiendavalt täiustades mudeleid ja raamistikke, et käsitleda ülemaailmse põllumajanduse arenguvajadusi. Lisaks on tehisintellekti mudelite energiatarbimise vähendamisele suunatud jätkuvad jõupingutused, nagu rõhutatakse väitekirjas, kriitilise tähtsusega, et viia põllumajanduslikud uuendused vastavusse keskkonna jätkusuutlikkuse eesmärkidega.
The thesis presents a comprehensive exploration of innovative, data-driven approaches for addressing critical challenges in sustainable agriculture, particularly within the emerging paradigm of Agri-Food 4.0. The study focuses on enhancing the accuracy, efficiency, and interpretability of machine learning (ML) models in predicting key agricultural parameters like evapotranspiration, soil organic carbon (SOC), Ex-ante Life Cycle Assessment for wheat production, and crop disease detection (Rust stripe wheat disease detection) using images, aiming to contribute to the development of sustainable and resilient agri-food systems. Research Objectives Agriculture is facing mounting pressures due to climate change, resource limitations, and the growing need for food security. These challenges necessitate advanced decision-making tools that can optimize resource use and improve crop yields while maintaining environmental sustainability. The concept of Agri-Food 4.0 involves integrating technologies such as the Internet of Things (IoT), cloud computing, and artificial intelligence (AI) to revolutionize agriculture. This thesis aligns with these objectives by proposing soft decision-making frameworks that are interpretable and sustainable, addressing both the efficiency and transparency needed in agricultural decision support systems (DSS). The research aims to: • Develop interpretable machine learning models that balance accuracy with interpretability, essential for stakeholders in agriculture. • Enhance the sustainability of agricultural practices by integrating Machine Learning models with domain-specific knowledge. • Apply automated machine learning (AutoML) techniques to optimize the detection of crop diseases, specifically focusing on wheat disease detection. Soil Organic Carbon (SOC) Prediction: Soil Organic Carbon (SOC) is a crucial component of soil health, influencing its structure, water retention capacity, and nutrient availability, all of which are vital for sustainable crop production. SOC is also a significant factor in the global carbon cycle, as soils are the largest terrestrial carbon reservoir. Small changes in SOC levels can significantly affect atmospheric CO2 concentrations, thereby impacting climate change. The thesis explores various machine learning techniques to predict SOC levels more accurately and efficiently. Traditional methods for SOC estimation, which often involve complex, labour-intensive sampling and analysis, are enhanced by data-driven models like ANFIS-T (Adaptive Neuro-Fuzzy Inference System with fractional Tikhonov regularization). The research demonstrates that ANFIS-T not only improves prediction accuracy but also maintains a high level of interpretability, which is essential for practical agricultural applications. This is particularly important as it aligns with global efforts, such as the EU's soil strategy for 2030, which emphasizes increasing SOC to meet greenhouse gas reduction targets. Evapotranspiration Prediction: Evapotranspiration (ET) is the sum of water evaporation from the soil and transpiration from plants. It is a critical parameter in water resource management, especially in agriculture where efficient irrigation strategies are needed to optimize water use. Agriculture accounts for a significant portion of global water consumption, with ET representing the largest component of water loss in agricultural systems. Traditional methods for measuring and estimating ET, such as lysimeters and evaporimeters, are often limited by their high cost and the difficulty of deploying them at scale. The thesis evaluates the use of machine learning models, particularly ANFIS-T, to predict ET. These models are shown to provide accurate predictions while being interpretable enough to be used by farmers and agricultural managers to make informed decisions about irrigation practices. By improving the accuracy of ET predictions, the research contributes to more sustainable water use in agriculture, which is increasingly important in the context of climate change and water scarcity. Ex-Ante Life Cycle Assessment (LCA) for Wheat Production: Life Cycle Assessment (LCA) is a comprehensive methodology for assessing the environmental impacts of products, from raw material extraction through production, use, and disposal. Traditional LCAs are typically conducted after a product's life cycle is complete (ex-post), limiting their utility in making proactive, sustainable decisions. The thesis advances the use of ex-ante LCA, which evaluates potential environmental impacts during the research and development phase, allowing for earlier interventions and more sustainable product designs. In the context of wheat production, the thesis integrates machine learning with LCA to predict and mitigate environmental impacts such as greenhouse gas emissions and resource use. This approach is particularly aligned with the European Union’s Green Deal, which aims to reduce carbon emissions across all sectors, including agriculture. The integration of ML into LCA not only enhances the predictive capabilities of LCA but also makes it more interpretable and actionable for stakeholders. Crop Disease Detection: Crop diseases, particularly those affecting staple crops like wheat, pose significant threats to global food security. Wheat stripe rust, caused by the fungal pathogen Puccinia striiformis, is one of the most devastating diseases affecting wheat. Traditional methods of disease detection, which rely heavily on manual inspection, are not scalable and are prone to errors. The thesis proposes a novel AutoML framework that incorporates context-aware feature engineering to improve the accuracy and efficiency of detecting wheat stripe rust. By leveraging UAV-acquired imagery and advanced image processing techniques, the framework outperforms traditional deep learning models like ResNet-18 in both accuracy and computational efficiency. This approach not only streamlines the disease detection process but also enhances the scalability of monitoring systems, making it feasible to deploy on large agricultural fields. Conclusion: The thesis concludes that the integration of advanced ML techniques with domain-specific knowledge significantly enhances the ability to address key challenges in agriculture. By improving the interpretability, accuracy, and efficiency of these models, the research contributes to the development of more sustainable and resilient agri-food systems. Future research could explore the broader application of these techniques across different crops and environments, further refining the models and frameworks to address the evolving needs of global agriculture. Additionally, continued efforts in reducing the energy consumption of AI models, as emphasized in the thesis, will be crucial for aligning agricultural innovation with environmental sustainability goals.
The thesis presents a comprehensive exploration of innovative, data-driven approaches for addressing critical challenges in sustainable agriculture, particularly within the emerging paradigm of Agri-Food 4.0. The study focuses on enhancing the accuracy, efficiency, and interpretability of machine learning (ML) models in predicting key agricultural parameters like evapotranspiration, soil organic carbon (SOC), Ex-ante Life Cycle Assessment for wheat production, and crop disease detection (Rust stripe wheat disease detection) using images, aiming to contribute to the development of sustainable and resilient agri-food systems. Research Objectives Agriculture is facing mounting pressures due to climate change, resource limitations, and the growing need for food security. These challenges necessitate advanced decision-making tools that can optimize resource use and improve crop yields while maintaining environmental sustainability. The concept of Agri-Food 4.0 involves integrating technologies such as the Internet of Things (IoT), cloud computing, and artificial intelligence (AI) to revolutionize agriculture. This thesis aligns with these objectives by proposing soft decision-making frameworks that are interpretable and sustainable, addressing both the efficiency and transparency needed in agricultural decision support systems (DSS). The research aims to: • Develop interpretable machine learning models that balance accuracy with interpretability, essential for stakeholders in agriculture. • Enhance the sustainability of agricultural practices by integrating Machine Learning models with domain-specific knowledge. • Apply automated machine learning (AutoML) techniques to optimize the detection of crop diseases, specifically focusing on wheat disease detection. Soil Organic Carbon (SOC) Prediction: Soil Organic Carbon (SOC) is a crucial component of soil health, influencing its structure, water retention capacity, and nutrient availability, all of which are vital for sustainable crop production. SOC is also a significant factor in the global carbon cycle, as soils are the largest terrestrial carbon reservoir. Small changes in SOC levels can significantly affect atmospheric CO2 concentrations, thereby impacting climate change. The thesis explores various machine learning techniques to predict SOC levels more accurately and efficiently. Traditional methods for SOC estimation, which often involve complex, labour-intensive sampling and analysis, are enhanced by data-driven models like ANFIS-T (Adaptive Neuro-Fuzzy Inference System with fractional Tikhonov regularization). The research demonstrates that ANFIS-T not only improves prediction accuracy but also maintains a high level of interpretability, which is essential for practical agricultural applications. This is particularly important as it aligns with global efforts, such as the EU's soil strategy for 2030, which emphasizes increasing SOC to meet greenhouse gas reduction targets. Evapotranspiration Prediction: Evapotranspiration (ET) is the sum of water evaporation from the soil and transpiration from plants. It is a critical parameter in water resource management, especially in agriculture where efficient irrigation strategies are needed to optimize water use. Agriculture accounts for a significant portion of global water consumption, with ET representing the largest component of water loss in agricultural systems. Traditional methods for measuring and estimating ET, such as lysimeters and evaporimeters, are often limited by their high cost and the difficulty of deploying them at scale. The thesis evaluates the use of machine learning models, particularly ANFIS-T, to predict ET. These models are shown to provide accurate predictions while being interpretable enough to be used by farmers and agricultural managers to make informed decisions about irrigation practices. By improving the accuracy of ET predictions, the research contributes to more sustainable water use in agriculture, which is increasingly important in the context of climate change and water scarcity. Ex-Ante Life Cycle Assessment (LCA) for Wheat Production: Life Cycle Assessment (LCA) is a comprehensive methodology for assessing the environmental impacts of products, from raw material extraction through production, use, and disposal. Traditional LCAs are typically conducted after a product's life cycle is complete (ex-post), limiting their utility in making proactive, sustainable decisions. The thesis advances the use of ex-ante LCA, which evaluates potential environmental impacts during the research and development phase, allowing for earlier interventions and more sustainable product designs. In the context of wheat production, the thesis integrates machine learning with LCA to predict and mitigate environmental impacts such as greenhouse gas emissions and resource use. This approach is particularly aligned with the European Union’s Green Deal, which aims to reduce carbon emissions across all sectors, including agriculture. The integration of ML into LCA not only enhances the predictive capabilities of LCA but also makes it more interpretable and actionable for stakeholders. Crop Disease Detection: Crop diseases, particularly those affecting staple crops like wheat, pose significant threats to global food security. Wheat stripe rust, caused by the fungal pathogen Puccinia striiformis, is one of the most devastating diseases affecting wheat. Traditional methods of disease detection, which rely heavily on manual inspection, are not scalable and are prone to errors. The thesis proposes a novel AutoML framework that incorporates context-aware feature engineering to improve the accuracy and efficiency of detecting wheat stripe rust. By leveraging UAV-acquired imagery and advanced image processing techniques, the framework outperforms traditional deep learning models like ResNet-18 in both accuracy and computational efficiency. This approach not only streamlines the disease detection process but also enhances the scalability of monitoring systems, making it feasible to deploy on large agricultural fields. Conclusion: The thesis concludes that the integration of advanced ML techniques with domain-specific knowledge significantly enhances the ability to address key challenges in agriculture. By improving the interpretability, accuracy, and efficiency of these models, the research contributes to the development of more sustainable and resilient agri-food systems. Future research could explore the broader application of these techniques across different crops and environments, further refining the models and frameworks to address the evolving needs of global agriculture. Additionally, continued efforts in reducing the energy consumption of AI models, as emphasized in the thesis, will be crucial for aligning agricultural innovation with environmental sustainability goals.
Kirjeldus
Väitekirja elektrooniline versioon ei sisalda publikatsioone