Mário Cunha

Abstract

Abstract
Roads and roadsides provide dispersal channels for non-native invasive alien plants (IAP), many of which hold devastating impacts in the economy, human health, biodiversity and ecosystem functionality. Remote sensing is an essential tool for efficiently assessing and monitoring the dynamics of IAP along roads. In this study, we explore the potentialities of object based image analysis (OBIA) approach to map several invasive plant species along roads using very high spatial resolution imagery. We compared the performance of OBIA approaches implemented in one open source software (OTB/Monteverdi) against those available in two proprietary pro-grams (eCognition and ArcGIS). We analysed the images by two sequential processes. First, we obtained a land-cover map for 15 study sites by segmenting the images with the algorithms Mean Shift Segmentation (MSS) and Multiresolution Segmentation (MRS), and by classifying the segmented images with the algorithms Support Vector Machine (SVM), Nearest Neighbour Classifier (NNC) and Maximum Likelihood Classifier (MLC). We created a mask using the polygons classified as non-vegetation to crop the images of the 15 study sites. Second, we repeated the previous segmentation and classification steps over the 15 masked images of vegetated areas using the same algorithms. OTB/Monteverdi, with MSS and SVM algorithms, showed to be a good software for land-cover mapping (OA = 87.0%), as well as ArcGIS, with MSS and MLC algorithms (OA = 84.3%). However, these two programs, using the same segmentation algorithms, did not achieve good accuracy results when mapping IAP species (OA(OTB/Monteverdi) = 63.3%; OAA(cos = 45.7%). eCognition, with MRS and NNC algorithms, reached better classification results in both land-cover and IAP maps (OA(Land-cover )= 95.7%; OA(Invasive-plant )= 92.8%). 'Bare soil' and 'Road', and 'A. donax' were the classes with best and worst overall accuracy, respectively, when mapping land-cover classes in the three programs. 'Other trees' was the class with the most accurate and significant differences in the three programs when mapping IAP species. The separation of each invasive species should be improved with a phenology-based design of field surveys. This study demonstrates the effectiveness of sequential segmentation and classification of RS data for mapping and monitoring plant invasions along linear infrastructures, which allows to reduce the time, cost and hazard of extensive field campaigns along roadsides.

Abstract
Wildfire disturbances can cause modifications in different dimensions of ecosystem functioning, i.e., the flows of matter and energy. There is an increasing need for methods to assess such changes, as functional approaches offer advantages over those focused solely on structural or compositional attributes. In this regard, remote sensing can support indicators for estimating a wide variety of effects of fire on ecosystem functioning, beyond burn severity assessment. These indicators can be described using intra-annual metrics of quantity, seasonality, and timing, called Ecosystem Functioning Attributes (EFAs). Here, we propose a satellite-based framework to evaluate the impacts, at short to medium term (i.e., from the year of fire to the second year after), of wildfires on four dimensions of ecosystem functioning: (i) primary productivity, (ii) vegetation water content, (iii) albedo, and (iv) sensible heat. We illustrated our approach by comparing inter-annual anomalies in satellite-based EFAs in the northwest of the Iberian Peninsula, from 2000 to 2018. Random Forest models were used to assess the ability of EFAs to discriminate burned vs. unburned areas and to rank the predictive importance of EFAs. Together with effect sizes, this ranking was used to select a parsimonious set of indicators for analyzing the main effects of wildfire disturbances on ecosystem functioning, for both the whole study area (i.e., regional scale), as well as for four selected burned patches with different environmental conditions (i.e., local scale). With both high accuracies (area under the receiver operating characteristic curve (AUC) > 0.98) and effect sizes (Cohen's |d| > 0.8), we found important effects on all four dimensions, especially on primary productivity and sensible heat, with the best performance for quantity metrics. Different spatiotemporal patterns of wildfire severity across the selected burned patches for different dimensions further highlighted the importance of considering the multi-dimensional effects of wildfire disturbances on key aspects of ecosystem functioning at different timeframes, which allowed us to diagnose both abrupt and lagged effects. Finally, we discuss the applicability as well as the potential advantages of the proposed approach for more comprehensive assessments of fire severity.

Abstract
Soil erosion is a severe and complex issue in the agriculture area. The main objective of this study was to assess the soil loss in two regions, testing different methodologies and combining different factors of the Revised Universal Soil Loss Equation (RUSLE) based on Geographical Information Systems (GIS). To provide the methodologies to other users, a GIS open-source application was developed. The RUSLE equation was applied with the variation of some factors that compose it, namely the slope length and slope steepness (LS) factor and practices factor (P), but also with the use of different sources of information. Eight different erosion models (M1 to M8) were applied to the two regions with different ecological conditions: Montalegre (rainy-mountainous) and Alentejo (dry-flat), both in Portugal, to compare them and to evaluate the soil loss for 3 potential erosion levels: 0–25, 25–50 and >50 ton/ha·year. Regarding the methodologies, in both regions the behavior is similar, indicating that the M5 and M6 methodologies can be more conservative than the others (M1, M2, M3, M4 and M8), which present very consistent values in all classes of soil loss and for both regions. All methodologies were implemented in a GIS application, which is free and available under QGIS software. © 2021 by the authors. Licensee MDPI, Basel, Switzerland.

Abstract
The development of robotic solutions for agriculture requires advanced perception capabilities that can work reliably in any crop stage. For example, to automatise the tomato harvesting process in greenhouses, the visual perception system needs to detect the tomato in any life cycle stage (flower to the ripe tomato). The state-of-the-art for visual tomato detection focuses mainly on ripe tomato, which has a distinctive colour from the background. This paper contributes with an annotated visual dataset of green and reddish tomatoes. This kind of dataset is uncommon and not available for research purposes. This will enable further developments in edge artificial intelligence for in situ and in real-time visual tomato detection required for the development of harvesting robots. Considering this dataset, five deep learning models were selected, trained and benchmarked to detect green and reddish tomatoes grown in greenhouses. Considering our robotic platform specifications, only the Single-Shot MultiBox Detector (SSD) and YOLO architectures were considered. The results proved that the system can detect green and reddish tomatoes, even those occluded by leaves. SSD MobileNet v2 had the best performance when compared against SSD Inception v2, SSD ResNet 50, SSD ResNet 101 and YOLOv4 Tiny, reaching an F1-score of 66.15, an mAP of 51.46 and an inference time of 16.44ms with the NVIDIA Turing Architecture platform, an NVIDIA Tesla T4, with 12 GB. YOLOv4 Tiny also had impressive results, mainly concerning inferring times of about 5ms.