Weeds pose significant challenges in agricultural crop production, causing substantial short- and long-term damage. Today, weed control heavily relies on chemical and, to a lesser extent, mechanical interventions, incurring high costs for both weed control products and labour. Indeed, weed management stands out as one of the most substantial expenses in agricultural production. To enhance the sustainability of agricultural crop production by minimizing chemical usage, it becomes imperative to increase the effectiveness of treatments through precise species-level identification and accurate monitoring of spatial distribution. The utilization of drones for weed detection emerges as a promising method, offering the potential to create automated weed distribution maps using specialized software and algorithms. This thesis concentrates on optimizing the procedures for weed detection through drones, with the primary goal of demonstrating that by refining weed detection from drones, not only can the exact location be identified, but also weeds can be discriminated at the species level. The central focus of this thesis revolves around enhancing the resolution of drone-acquired images. This was achieved through dedicated flights and the acquisition of images at various altitudes over soybean and sunflower crop fields. Altitudes of 5 m, 7.5 m, 10 m, 12.5 m, 15 m, and 30 m were chosen, utilizing the DJI Mavic 3 Enterprise drone equipped with an RGB sensor. In an effort to achieve species-level classification of weeds from the acquired images, various software tools (ArcGIS and SAGA) and dedicated algorithms such as Artificial Neural Networks (ANN), Maximum Likelihood Classifier (MLC), and Deep Learning (DL) were tested. Results indicated that not all algorithms could effectively distinguish weed species from the crops, specifically ANN and MLC. The most promising results were obtained through DL, displaying high overall accuracy. For the soybean field, the accuracy at 5 m was 94%, at 7.5 m it was 79%, at 10 m it was 80.8%, at 12.5 m the accuracy was 83%, at 15 m it reached 88%, and at 30 m, it was 85.5%. Concerning the sunflower field, the overall accuracy was 98.8% at 5 m, 90.2% at 7.5 m, 86.1% at 10 m, 77.3% at 12.5 m, and 91.4% at 15 m. Recognizing the limited accuracy at 30 m, a specific classification was implemented for the sunflower field, focusing on the differentiation between broadleaf and narrowleaf species. This modification resulted in a significant improvement, achieving an overall accuracy of 96.2%. The outcomes suggest that by optimizing image resolution and employing advanced software and algorithms, species-level classification of weeds is achievable. The accuracy achieved in species-level discrimination across the extensive dataset covered in this thesis reaffirms the feasibility of using these technologies for weed surveys. Implementing this methodology for weed classification could enable the creation of high-precision prescription maps, providing information not only on weed location for site-specific distribution through machinery like tractors or robots with appropriate weed bars but also guiding the selection of herbicides and their targeted application. The proposed methodology, supported by the obtained results, holds the potential to significantly enhance the efficiency of weed control operations, reducing associated costs while preserving the environment.
Le piante infestanti rappresentano uno dei maggiori problemi delle colture agrarie, in quanto provocano danni alla produzione sia nel breve che nel lungo periodo. Attualmente il controllo delle piante infestanti fa affidamento principalmente sul diserbo chimico e in minor misura sul controllo meccanico, operazioni che comportano elevati costi sia per i prodotti impiegati sia per la manodopera. Infatti, la gestione delle infestanti rappresenta uno dei maggiori costi della produzione agricola. Al fine di ridurre l’impiego di prodotti chimici e rendere più sostenibili le coltivazioni agrarie, è importante rendere il trattamento più efficace grazie all’identificazione delle malerbe a livello di specie e al monitoraggio accurato della loro distribuzione spaziale. Per raggiungere questo obiettivo, l’utilizzo dei droni per il rilevamento delle infestanti sembra essere un metodo promettente, consentendo la creazione di mappe di distribuzione delle infestanti realizzate in maniera automatica da specifici software e algoritmi. La tesi in oggetto si concentra sull’impiego dei droni per il rilevamento delle malerbe e mira a ottimizzare le procedure di identificazione. L’obiettivo principale è dimostrare che tramite l’ottimizzazione del rilevamento da drone è possibile, oltre che identificare la posizione, anche discriminare le infestanti a livello di specie. La tesi si è focalizzata sull’ottimizzazione della risoluzione delle immagini acquisite tramite drone. Ciò è stato realizzato attraverso voli dedicati e l'acquisizione di fotografie a diverse altitudini, sorvolando campi coltivati di soia e girasole. Le altitudini selezionate sono state 5 m, 7.5 m, 10 m, 12.5 m, 15 m e 30 m, utilizzando il drone DJI Mavic 3 Enterprise equipaggiato con un sensore RGB. Nel tentativo di ottenere la classificazione a livello della specie delle infestanti delle immagini acquisite sono stati testati software (ArcGIS e SAGA) ed algoritmi dedicati quali Reti Neurali Artificiali (ANN), Classificatore delle Massime Verosimiglianze (MLC) e Deep Learning (DL). I risultati hanno dimostrato come non tutti gli algoritmi sono stati in grado di distinguere le specie dalla coltura, nello specifico ANN e MLC. I risultati migliori sono stati ottenuti tramite DL. I risultati prodotti usando questo algoritmo avevano l’accuratezza complessiva (overall accuracy) abbastanza elevata. Per il campo di soia i risultati ad altezza di 5 m erano pari a 94%, a 7,5 m pari a 79%, a 10 m pari a 80,8%, a 12,5 l’accuratezza era di 83%, a 15 m 88% e a 30 m 85,5%. Per quanto riguarda il campo di girasole l’accuratezza complessiva era di 98,8% a 5 m, 90,2% a 7,5 m, 86,1% a 10 m, 77,3% a 12,5 m, 91,4% a 15 m. Data la limitata accuratezza nella classificazione a 30 metri, è stata implementata una classificazione specifica per il campo di girasole, basata sulla distinzione tra specie a foglia larga e specie a foglia stretta. Questa modifica ha portato a un miglioramento significativo, con un'accuratezza complessiva del 96,2%. I risultati indicano che ottimizzando la risoluzione delle immagini ed applicando i software e gli algoritmi all’avanguardia, è possibile ottenere la classificazione a livello di specie delle malerbe. L’accuratezza ottenuta relativa alla discriminazione a livello di specie nella stragrande maggioranza delle foto oggetto di tesi conferma la possibilità di utilizzo di tali tecnologie per i rilievi malerbologici e per la creazione di mappe di prescrizione, che offrirebbero informazioni relative sia alla posizione delle malerbe nel campo per la distribuzione sito-specifica, operata da macchinari, ma anche quelle relative all’approccio più adeguato come ad esempio la scelta degli erbicidi da applicare. La metodologia proposta in questa tesi, confermata dai risultati ottenuti, potrebbe portare ad un notevole aumento dell’efficacia delle operazioni di diserbo, riducendo, i costi relativi al controllo, salvaguardando l’ambiente.
Ottimizzazione delle tecniche di rilevamento delle piante infestanti da drone
COSCIA, DOMENICO GIUSEPPE
2022/2023
Abstract
Weeds pose significant challenges in agricultural crop production, causing substantial short- and long-term damage. Today, weed control heavily relies on chemical and, to a lesser extent, mechanical interventions, incurring high costs for both weed control products and labour. Indeed, weed management stands out as one of the most substantial expenses in agricultural production. To enhance the sustainability of agricultural crop production by minimizing chemical usage, it becomes imperative to increase the effectiveness of treatments through precise species-level identification and accurate monitoring of spatial distribution. The utilization of drones for weed detection emerges as a promising method, offering the potential to create automated weed distribution maps using specialized software and algorithms. This thesis concentrates on optimizing the procedures for weed detection through drones, with the primary goal of demonstrating that by refining weed detection from drones, not only can the exact location be identified, but also weeds can be discriminated at the species level. The central focus of this thesis revolves around enhancing the resolution of drone-acquired images. This was achieved through dedicated flights and the acquisition of images at various altitudes over soybean and sunflower crop fields. Altitudes of 5 m, 7.5 m, 10 m, 12.5 m, 15 m, and 30 m were chosen, utilizing the DJI Mavic 3 Enterprise drone equipped with an RGB sensor. In an effort to achieve species-level classification of weeds from the acquired images, various software tools (ArcGIS and SAGA) and dedicated algorithms such as Artificial Neural Networks (ANN), Maximum Likelihood Classifier (MLC), and Deep Learning (DL) were tested. Results indicated that not all algorithms could effectively distinguish weed species from the crops, specifically ANN and MLC. The most promising results were obtained through DL, displaying high overall accuracy. For the soybean field, the accuracy at 5 m was 94%, at 7.5 m it was 79%, at 10 m it was 80.8%, at 12.5 m the accuracy was 83%, at 15 m it reached 88%, and at 30 m, it was 85.5%. Concerning the sunflower field, the overall accuracy was 98.8% at 5 m, 90.2% at 7.5 m, 86.1% at 10 m, 77.3% at 12.5 m, and 91.4% at 15 m. Recognizing the limited accuracy at 30 m, a specific classification was implemented for the sunflower field, focusing on the differentiation between broadleaf and narrowleaf species. This modification resulted in a significant improvement, achieving an overall accuracy of 96.2%. The outcomes suggest that by optimizing image resolution and employing advanced software and algorithms, species-level classification of weeds is achievable. The accuracy achieved in species-level discrimination across the extensive dataset covered in this thesis reaffirms the feasibility of using these technologies for weed surveys. Implementing this methodology for weed classification could enable the creation of high-precision prescription maps, providing information not only on weed location for site-specific distribution through machinery like tractors or robots with appropriate weed bars but also guiding the selection of herbicides and their targeted application. The proposed methodology, supported by the obtained results, holds the potential to significantly enhance the efficiency of weed control operations, reducing associated costs while preserving the environment.File | Dimensione | Formato | |
---|---|---|---|
COSCIA_DOMENICO.pdf
accesso aperto
Dimensione
12.14 MB
Formato
Adobe PDF
|
12.14 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/59118