Confusion matrix for images classified for scenario (ii).

A confusion matrix for the comparison of controls with responses from the crowd.

Land Cover Map 2021 (LCM2021) is a suite of geospatial land cover datasets (raster and polygon) describing the UK land surface in 2021. These were produced at the UK Centre for Ecology & Hydrology by classifying satellite images from 2021. Land cover maps describe the physical material on the surface of the country. For example grassland, woodland, rivers & lakes or man-made structures such as roads and buildingsThis is a 10 m Classified Pixel dataset, classified to create a single mosaic of national cover. Provenance and quality:UKCEH’s automated land cover classification algorithms generated the 10m classified pixels. Training data were automatically selected from stable land covers over the interval of 2017 to 2019. A Random Forest classifier used these to classify four composite images representing per season median surface reflectance. Seasonal images were integrated with context layers (e.g., height, aspect, slope, coastal proximity, urban proximity and so forth) to reduce confusion among classes with similar spectra.Land cover was validated by organising the pixel classification into a land parcel framework (the LCM2021 Classified Land Parcels product). The classified land parcels were compared to known land cover producing confusion matrix to determine overall and per class accuracy.View full metadata information and download the data at catalogue.ceh.ac.uk

These products were developed to provide scientific and correspondingly spatially explicit information regarding the distribution and abundance of conifers (namely, singleleaf pinyon (Pinus monophylla), Utah juniper (Juniperus osteosperma), and western juniper (Juniperus occidentalis)) in Nevada and portions of northeastern California. Encroachment of these trees into sagebrush ecosystems of the Great Basin can present a threat to populations of greater sage-grouse (Centrocercus urophasianus). These data provide land managers and other interested parties with a high-resolution representation of conifers across the range of sage-grouse habitat in Nevada and northeastern California that can be used for a variety of management and research applications. We mapped conifer trees at 1 x 1 meter resolution across the extent of all Nevada Department of Wildlife Sage-grouse Population Management Units plus a 10 km buffer. Using 2010 and 2013 National Agriculture Imagery Program digital orthophoto quads (DOQQs) as our reference imagery, we applied object-based image analysis with Feature Analyst software (Overwatch, 2013) to classify conifer features across our study extent. This method relies on machine learning algorithms that extract features from imagery based on their spectral and spatial signatures. Conifers in 6230 DOQQs were classified and outputs were then tested for errors of omission and commission using stratified random sampling. Results of the random sampling were used to populate a confusion matrix and calculate the overall map accuracy of 84.3 percent. We provide 5 sets of products for this mapping process across the entire mapping extent: (1) a shapefile representing accuracy results linked to our mapping subunits; (2) binary rasters representing conifer presence or absence at a 1 x 1 meter resolution; (3) a 30 x 30 meter resolution raster representing percentage of conifer canopy cover within each cell from 0 to 100; (4) 1 x 1 meter resolution canopy cover classification rasters derived from a 50 meter radius moving window analysis; and (5) a raster prioritizing pinyon-juniper management for sage-grouse habitat restoration efforts. The latter three products can be reclassified into user-specified bins to meet different management or study objectives, which include approximations for phases of encroachment. These products complement, and in some cases improve upon, existing conifer maps in the western United States, and will help facilitate sage-grouse habitat management and sagebrush ecosystem restoration. These data support the following publication: Coates, P.S., Gustafson, K.B., Roth, C.L., Chenaille, M.P., Ricca, M.A., Mauch, Kimberly, Sanchez-Chopitea, Erika, Kroger, T.J., Perry, W.M., and Casazza, M.L., 2017, Using object-based image analysis to conduct high-resolution conifer extraction at regional spatial scales: U.S. Geological Survey Open-File Report 2017-1093, 40 p., https://doi.org/10.3133/ofr20171093. References: ESRI, 2013, ArcGIS Desktop: Release 10.2: Environmental Systems Research Institute. Overwatch, 2013, Feature Analyst Version 5.1.2.0 for ArcGIS: Overwatch Systems Ltd.

This is a web map service (WMS) for the 10-metre Land Cover Map 2023. The map presents the and surface classified into 21 UKCEH land cover classes, based upon Biodiversity Action Plan broad habitats.UKCEH’s automated land cover algorithms classify 10 m pixels across the whole of UK. Training data were automatically selected from stable land covers over the interval of 2020 to 2022. A Random Forest classifier used these to classify four composite images representing per season median surface reflectance. Seasonal images were integrated with context layers (e.g., height, aspect, slope, coastal proximity, urban proximity and so forth) to reduce confusion among classes with similar spectra.Land cover was validated by organising the 10 m pixel classification into a land parcel framework (the LCM2023 classified land parcels product). The classified land parcels were compared to known land cover producing a confusion matrix to determine overall and per class accuracy.

Robust risk assessment requires accurate flood intensity area mapping to allow for the identification of populations and elements at risk. However, available flood maps in West Africa lack spatial variability while global datasets have resolutions too coarse to be relevant for local scale risk assessment. Consequently, local disaster managers are forced to use traditional methods such as watermarks on buildings and media reports to identify flood hazard areas. In this study, remote sensing and Geographic Information System (GIS) techniques were combined with hydrological and statistical models to delineate the spatial limits of flood hazard zones in selected communities in Ghana, Burkina Faso and Benin. The approach involves estimating peak runoff concentrations at different elevations and then applying statistical methods to develop a Flood Hazard Index (FHI). Results show that about half of the study areas fall into high intensity flood zones. Empirical validation using statistical confusion matrix and the principles of Participatory GIS show that flood hazard areas could be mapped at an accuracy ranging from 77% to 81%. This was supported with local expert knowledge which accurately classified 79% of communities deemed to be highly susceptible to flood hazard. The results will assist disaster managers to reduce the risk to flood disasters at the community level where risk outcomes are first materialized.

Fruit and vegetable plants are vulnerable to diseases that can negatively affect crop yield, causing planters to incur significant losses. These diseases can affect the plants at various stages of growth. Planters must be on constant watch to prevent them early, or infestation can spread and become severe and irrecoverable. There are many types of pest infestations of fruits and vegetables, and identifying them manually for appropriate preventive measures is difficult and time-consuming.This pretrained model can be deployed to identify plant diseases efficiently for carrying out suitable pest control. The training data for the model primarily includes images of leaves of diseased and healthy fruit and vegetable plants. It can classify the multiple categories of plant infestation or healthy plants from the images of the leaves.Using the modelFollow the guide to use the model. Before using this model, ensure that the supported deep learning libraries are installed. For more details, check Deep Learning Libraries Installer for ArcGIS. Fine-tuning the modelThis model can be fine-tuned using the Train Deep Learning Model tool. Follow the guide to fine-tune this model.Input8 bit, 3-band (RGB) image. Recommended image size is 224 x 224 pixels. Note: Input images should have grey or solid color background with one full leaf per image. OutputClassified image of the leaf with any of the plant disease, healthy leaf, or background classes as in the Plant Leaf Diseases dataset.Applicable geographiesThis model is expected to work well in all regions globally. However, results can vary for images that are statistically dissimilar to training data.Model architectureThis model uses the ResNet50 model architecture implemented in ArcGIS API for Python.Accuracy metricsThis model has an overall accuracy of 97.88 percent. The confusion matrix below summarizes the performance of the model on the validation dataset. Sample resultsHere are a few results from the model:Ground truth: Apple_black_rot / Prediction: Apple_black_rotGround truth: Potato_early_blight / Prediction: Potato_early_bightGround truth: Raspberry_healthy / Prediction: Raspberry_healthyGround truth: Strawberry_leaf_scorch / Prediction: Strawberry_leaf_scorch

A robust early warning system can alert to the presence of food crises and related drivers, informing decision makers on food security. To date, decision-makers in Zimbabwe still rely on agriculture extension personnel to generate information on wheat production and monitor the crop. Such traditional methods are subjective, costly and their accuracy depends on the experience of the assessor. This study investigates Sentinel-2 NDVI and time series utility as a wheat-monitoring tool over the wheat-growing areas of Zimbabwe's Bindura, Shamva, and Guruve districts. NDVI was used to classify and map the wheat fields. The classification model's evaluation was done by creating 100 reference pixels across the classified map and constructing a confusion matrix with a resultant kappa coefficient of 0.89. A sensitivity test, receiver operating characteristic (ROC) and area under the curve (AUC) were used to measure the model's efficiency. Fifty GPS points randomly collected from wheat fields in the selected districts were used to identify and compute the area of the fields. The correlation between the area declared by farmers and the calculated area was positive, with an R2 value of 0.98 and a Root Mean Square Error (RMSE) of 2.23 hectares. The study concluded that NDVI is a good index for estimating the area under wheat. In this regard, NDVI can be used for early warning and early action, especially in monitoring programs like ‘Command Agriculture’ in Zimbabwe. In current and future studies, the use of high-resolution images from remote sensing is essential. Furthermore, ground truthing is always important to validate results from remote sensing at any spatial scale.

Spatial predictions of the fractions of mud, sand and gravel as continuous response variables for the north-west European continental shelf. Mud, sand and gravel fractions range from 0-1 (i.e. 0-100%). These fractions were generated from two additive log-ratios (ALR), ALRs and ALRm which are independent, unconstrained response variables. These raw predictions as rasters are also included presented in the attached dataset. Predicted fractions have been combined to predict the likely sediment classification based on the EUNIS level 3 sediment classification for broadscale habitats, Folk 5, Folk 7, Folk 11 and Folk 15 classification schemes. These are available as raster tif files with an ArcGIS layer file indicating the appropriate class for each raster value. For all predictions an accompanying map of the spatial distribution of error/accuracy is also included as a separate raster. For the three components of the sediment fraction a smoothed Root-Mean-Squared-Error layer is available. For the classification maps a smoothed local accuracy map is available. Spatial predictions of mud, sand and gravel were generated for the north-west European continental shelf. Based on these fractions sediment classification maps were also generated for the study site. To support the interpretation of these layers maps of the spatial distribution of error/accuracy were also generated. In short, analysis combined the eight continuous predictive layers (Bathymetry, Bathymetric position index at a 50-pixel radii, Bathymetric position index at a 434-pixel radii, Distance from coast, Current speed at the seabed, Wave peak orbital velocity at the seabed, and suspended inorganic particulate matter for summer and winter as two separate variables) with sediment observation data in a statistical regression model to make spatial predictions of the fractions of mud, sand and gravel. Spatial predictions were generated based on two additive log-ratios that could then be back transformed to produce spatial predictions for each fraction. From these spatial predictions any classification scheme based on the percentages of mud, sand and gravel can be applied. Included here are the five classification shemes generated from these maps. The maps of accuracy were also generated to support interpretation. For the maps of the fractions of mud, sand and gravel map error was calculated based on the Root-Mean-Squared-Error of the observed vs predicted fractions from the test samples. A smoothed surface of local RMSE was then generated using the Inverse Distance Weighted (IDW) technique in ArcGIS. Each pixels’ RMSE was determined based on the closest 50 points (up to a maximum distance of 200 km). A weighting power function was applied in the IDW tool (set at 0.3) so nearer points contributed more to the pixel than distant points. For the classified maps spatial accuracy was calculated using a locally constrained confusion matrix. The IDW technique was applied to calculate a local thematic accuracy value. As above, this was applied based on the closest 50 points (maximum distance of 200 km) with a weighting power function of 0.3.

This is a 50 x 50 meter cell sized (unit of analysis is 50x50 meters, the imagery used to develop this was 0.6 m NAIP) depiction of woody vegetation canopy cover throughout the AVCA Area of Interest. This dataset was developed from a detailed binary classification of the 2017 National Agricultural Imagery Program (NAIP) 60 centimeter imagery throughout the AVCA AOI. The base classification has an confusion matrix overall accuracy of 95%, a Kappa Statistic of 0.84, and no user nor producer's accuracy falling below 80%. From this classification dataset, we were able to determine that the overall woody vegetation canopy coverage throughout the AVCA AOI was 20.3% As a check, a point analysis (500 points) was conducted. Each point was evaluated to determine if it fell atop woody vegetation. Of the 500 points, 108 or 21.6% fell atop woody vegetation. Since this is a sample, we calculated the margin of error and can state that we are 95% certain that the proportion of woody vegetation canopy is between 18.0% and 25.2% [1]. Which supports our derived classification percentage of 20.3%. Of note, since this dataset was produced, the USGS released their series of brush analysis datasets (based on 2016 data). Based on this 2016 USGS brush percentage dataset, the percent of brush throughout the AVCA AOI was 19.9%.

Red indicates WV >40%

Orange 30-40%

Yellow: 20-30%

Light Green: 10-20%

Dark Green: 0-10%

Project funded by Western SARE.

[1] Caryl L. Elzinga, Daniel W. Salzer, and John W. Willoughby, Measuring & Monitoring Plant Populations (Denver, CO: Bureau of Land Management, 1998), 368