Crop traits retrieval and disease detection in potato: UAV optical sensing for plant monitoring

Marston Héracles Domingues Franceschini

Research output: Thesisinternal PhD, WU


The increase in food demand associated to the projected population growth in the next decades comes with different challenges and one of them is how to increase productivity while reducing or limiting environmental impacts of agricultural activities. For that, two crucial aspects for more efficient production systems comprise crop monitoring, targeting stress factors that can undermine productivity, and the selection of new cultivars under field conditions, which depends on phenotyping protocols. In both cases, remote or proximal sensing can potentially provide an alternative to conventional methods of crop traits evaluation. In particular, optical imaging systems on-board of Unmanned Aerial Vehicles (UAVs) have proved to be one of the most accessible and flexible sensing options for crop characterization. A range of UAV-compatible cameras are currently available for data acquisition in the optical domain, from high-resolution RGB to multispectral and hyperspectral systems, with images in tens to hundreds of spectral bands. Despite the successful use of such imaging systems in different cases, their inclusion in specific applications targeting the assessment of disease incidence and severity is still lacking in multiple scenarios. In this context, an attempt is made in this thesis to tackle key components of crop characterization and disease detection based on UAV optical imaging systems, with focus on hyperspectral sub-decimeter imagery, which still have unexplored potential for application in these topics.

In Chapter 2, the use of narrow-band hyperspectral imagery coupled with straightforward modelling methods (Vegetation Indices – VIs and univariate linear regression) is evaluated in the context of crop traits retrieval in organic potato production systems. In addition, results obtained with the hyperspectral camera are compared with those of ground-based spectral measurements (hand-held sensor) and the possible interchangeable use of both approaches is assessed. It is demonstrated that relatively accurate retrieval is possible based on narrow-band VIs even if simple models are used, with RMSE of 6.07 µg·cm−2, 0.67 m2·m−2, 0.24 g·m−2 and 5.5% obtained for leaf chlorophyll, leaf area index, canopy chlorophyll and canopy cover, for five UAV-based data acquisitions, from 43 to 99 Days After Planting (DAP). Furthermore, comparing VIs derived from UAV and hand-held narrow-band sensor indicated that interchangeable use would be possible for later season, when canopy is almost fully closed, and if changes to be monitored are expected to be uniform in the given experimental plot or field patch of interest. However, better results obtained with the hyperspectral images indicate that increased spatial sampling also generally increases prediction performance, even when changes in crop traits are relatively uniform, at the experimental plot scale.

Chapter 3 comprises the evaluation of sub-decimeter UAV hyperspectral imagery for severity assessment of late blight in potato. Focus was given to low disease severity levels (up to 15.0% of leaf area affected), which characterize stages allowing control practices by farmers. The analysis used consisted of Simplex Volume Maximization (SiVM), to summarize the spectral variability of healthy and diseased plants, coupled with log-likelihood ratio (LLR), to stablish the association of individual spectral signatures (pixel-wise) to characteristic spectral behaviour of different cultivation systems and disease severity classes. Spectral changes associated to disease severity levels as low as between 2.5 and 5.0% of affected leaf area could be detected using the UAV images. Comparison with ground based imagery (acquired with the same camera in handheld mode, resulting in higher spatial resolution) indicate that similar patterns were described in both cases, despite differences concerning the distribution of affected areas detected within the sampling units and an attenuation in the signal measured using the UAV platform. In addition, while aggregated information at sampling unit level provided discriminative potential for relatively high levels of disease incidence, relating the spectral information at the pixel level with late blight occurrence increased the discriminative potential of the data acquired.

Next, in Chapter 4 the classification of potato plants with respect to late blight severity using sub-decimeter UAV hyperspectral imagery was investigated. Again, special focus was given to low diseased severity levels. A specific modelling framework was applied to include the image spatial component in the analysis. This framework was composed of Dictionary Learning and Sparse Representation applied to overlapping image patches of 9 by 9 pixels, which were in turn used to represent individual sampling units (corresponding to areas of 0.75 by 1.00 m). This method was selected considering its simplicity and the possibly to be applied when the number of observations is limited. Prediction accuracy varied for assessments made 64 and 78 DAP, with balanced accuracy ranging between 0.61 and 0.63. Spatial patterns of the pathogen spread in the experimental field could be described, to a certain extent, by point-wise predictions derived using the best classifiers for each date. In addition, characteristic elements (spectral-spatial features) corresponding to the different disease severity classes could be derived. Particular spectral-spatial patterns associated to each class were identified for observations made 78 DAP. Finally, class-specific ‘attention maps’ allowed to identify areas related to the different severity intensities within each sampling unit.

Finally in Chapter 5, the classification of blackleg in seed potato using sub-decimeter UAV hyperspectral imagery was evaluated. In this case, spectral features (i.e., VIs) are compared to structural features, which were derived from point clouds obtained using high-resolution RGB images (i.e., using Structure from Motion – SfM) and LiDAR. The integration of spectral and structural features was also tested. Since blackleg in potato has symptoms that comprise changes in canopy structure and chemical composition it was expected that integrating optical and structural features would provide better classification results. This was confirmed with balanced accuracy of 0.782, 0.761, 0.708, 0.859 and 0.789 for datasets including VIs, LiDAR, SfM, VIs+LiDAR and VIs+SfM features, respectively. Therefore, despite the fact that structural features, especially those derived from LiDAR point cloud, provided discriminative information regarding healthy and diseased plants, using spectral information alone (with 2.0 cm spatial resolution) was sufficient to achieve relatively accurate predictions in this context.

Therefore, the content of this thesis provided relevant contribution to the field of crop monitoring, in particular regarding the assessments of disease severity in potato with focus on late blight and blackleg. Different sensing approaches (i.e., hyperspectral and high-resolution RGB imagery as well as LiDAR point cloud) and distinct classification methods were evaluated in order to better understand potential and limitations of UAV-based data acquisition in these cases. The results presented can assist future developments in terms of sensor systems and modelling approaches to be adopted, however the main contribution is perhaps the verification of the potential of centimeter-resolution optical datasets for the assessment of disease incidence in potato under field conditions.

Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • Wageningen University
  • Kooistra, Lammert, Promotor
  • Bartholomeus, Harm, Co-promotor
Award date12 Apr 2021
Place of PublicationWageningen
Print ISBNs9789463957410
Publication statusPublished - 12 Apr 2021


Dive into the research topics of 'Crop traits retrieval and disease detection in potato: UAV optical sensing for plant monitoring'. Together they form a unique fingerprint.

Cite this