Is deeper always better? Evaluating deep learning models for yield forecasting with small data.

Predicting crop yields, and especially anomalously low yields, is of special importance for food insecure countries. In this study, we investigate a flexible deep learning approach to forecast crop yield at the provincial administrative level based on deep 1D and 2D convolutional neural networks using limited data. This approach meets the operational requirements-public and global records of satellite data in an application ready format with near real time updates-and can be transferred to any country with reliable yield statistics. Three-dimensional histograms of normalized difference vegetation index (NDVI) and climate data are used as input to the 2D model, while simple administrative-level time series averages of NDVI and climate data to the 1D model. The best model architecture is automatically identified during efficient and extensive hyperparameter optimization. To demonstrate the relevance of this approach, we hindcast (2002-2018) the yields of Algeria's three main crops (barley, durum and soft wheat) and contrast the model's performance with machine learning algorithms and conventional benchmark models used in a previous study. Simple benchmarks such as peak NDVI remained challenging to outperform while machine learning models were superior to deep learning models for all forecasting months and all tested crops. We attribute the poor performance of deep learning to the small size of the dataset available.

Abstracts of Presentations to the Working Session on Improving Predictive Modeling of Mycotoxin Risk for Africa Held at the 3rd ASM2022 on 7 September 2022, in Stellenbosch, South Africa.

In 2008, the African Postharvest Losses Information Systems project (APHLIS, accessed on 6 September 2022) developed an algorithm for estimating the scale of cereal postharvest losses (PHLs). The relevant scientific literature and contextual information was used to build profiles of the PHLs occurring along the value chains of nine cereal crops by country and province for 37 sub-Saharan African countries. The APHLIS provides estimates of PHL figures where direct measurements are not available. A pilot project was subsequently initiated to explore the possibility of supplementing these loss estimates with information on the aflatoxin risk. Using satellite data on drought and rainfall, a time series of agro-climatic aflatoxin risk warning maps for maize was developed covering the countries and provinces of sub-Saharan Africa. The agro-climatic risk warning maps for specific countries were shared with mycotoxin experts from those countries for review and comparison with their aflatoxin incidence datasets. The present Work Session was a unique opportunity for African food safety mycotoxins experts, as well as other international experts, to meet and deepen the discussion about prospects for using their experience and their data to validate and improve agro-climatic risk modeling approaches.

Comparing land surface phenology of major European crops as derived from SAR and multispectral data of Sentinel-1 and -2.

The frequent acquisitions of fine spatial resolution imagery (10 m) offered by recent multispectral satellite missions, including Sentinel-2, can resolve single agricultural fields and thus provide crop-specific phenology metrics, a crucial information for crop monitoring. However, effective phenology retrieval may still be hampered by significant cloud cover. Synthetic aperture radar (SAR) observations are not restricted by weather conditions, and Sentinel-1 thus ensures more frequent observations of the land surface. However, these data have not been systematically exploited for phenology retrieval so far. In this study, we extracted crop-specific land surface phenology (LSP) from Sentinel-1 and Sentinel-2 of major European crops (common and durum wheat, barley, maize, oats, rape and turnip rape, sugar beet, sunflower, and dry pulses) using ground-truth information from the "Copernicus module" of the Land Use/Cover Area frame statistical Survey (LUCAS) of 2018. We consistently used a single model-fit approach to retrieve LSP metrics on temporal profiles of CR (Cross Ratio, the ratio of the backscattering coefficient VH/VV from Sentinel-1) and NDVI (Normalized Difference Vegetation Index from Sentinel-2). Our analysis revealed that LSP retrievals from Sentinel-1 are comparable to those of Sentinel-2, particularly for winter crops. The start of season (SOS) timings, as derived from Sentinel-1 and -2, are significantly correlated (average r of 0.78 for winter and 0.46 for summer crops). The correlation is lower for end of season retrievals (EOS, r of 0.62 and 0.34). Agreement between LSP derived from Sentinel-1 and -2 varies among crop types, ranging from r = 0.89 and mean absolute error MAE = 10 days (SOS of dry pulses) to r = 0.15 and MAE = 53 days (EOS of sugar beet). Observed deviations revealed that Sentinel-1 and -2 LSP retrievals can be complementary; for example for winter crops we found that SAR detected the start of the spring growth while multispectral data is sensitive to the vegetative growth before and during winter. To test if our results correspond reasonably to in-situ data, we compared average crop-specific LSP for Germany to average phenology from ground phenological observations of 2018 gathered from the German Meteorological Service (DWD). Our study demonstrated that both Sentinel-1 and -2 can provide relevant and at times complementary LSP information at field- and crop-level.

Near real-time vegetation anomaly detection with MODIS NDVI: Timeliness vs. accuracy and effect of anomaly computation options.

For food crises early warning purposes, coarse spatial resolution NDVI data are widely used to monitor vegetation conditions in near real-time (NRT). Different types of NDVI anomalies are typically employed to assess the current state of crops and rangelands as compared to previous years. Timeliness and accuracy of such anomalies are critical factors to an effective monitoring. Temporal smoothing can efficiently reduce noise and cloud contamination in the time series of historical observations, where data points are available before and after each observation to be smoothed. With NRT data, smoothing methods are adapted to cope with the unbalanced availability of data before and after the most recent data points. These NRT approaches provide successive updates of the estimation of the same data point as more observations become available. Anomalies compare the current NDVI value with some statistics (e.g. indicators of central tendency and dispersion) extracted from the historical archive of observations. With multiple updates of the same datasets being available, two options can be selected to compute anomalies, i.e. using the same update level for the NRT data and the statistics or using the most reliable update for the latter. In this study we assess the accuracy of three commonly employed 1 km MODIS NDVI anomalies (standard scores, non-exceedance probability and vegetation condition index) with respect to (1) delay with which they become available and (2) option selected for their computation. We show that a large estimation error affects the earlier estimates and that this error is efficiently reduced in subsequent updates. In addition, with regards to the preferable option to compute anomalies, we empirically observe that it depends on the type of application (e.g. averaging anomalies value over an area of interest vs. detecting "drought" conditions by setting a threshold on the anomaly value) and the employed anomaly type. Finally, we map the spatial pattern in the magnitude of NRT anomaly estimation errors over the globe and relate it to average cloudiness.

ASAP: A new global early warning system to detect anomaly hot spots of agricultural production for food security analysis.

Monitoring crop and rangeland conditions is highly relevant for early warning and response planning in food insecure areas of the world. Satellite remote sensing can obtain relevant and timely information in such areas where ground data are scattered, non-homogenous, or frequently unavailable. Rainfall estimates provide an outlook of the drivers of vegetation growth, whereas time series of satellite-based biophysical indicators at high temporal resolution provide key information about vegetation status in near real-time and over large areas. The new early warning decision support system ASAP (Anomaly hot Spots of Agricultural Production) builds on the experience of the MARS crop monitoring activities for food insecure areas, that have started in the early 2000's and aims at providing timely information about possible crop production anomalies. The information made available on the website (https://mars.jrc.ec.europa.eu/asap/) directly supports multi-agency early warning initiatives such as for example the GEOGLAM Crop Monitor for Early Warning and provides inputs to more detailed food security assessments that are the basis for the annual Global Report on Food Crises. ASAP is a two-step analysis framework, with a first fully automated step classifying the first sub-national level administrative units into four agricultural production deficit warning categories. Warnings are based on rainfall and vegetation index anomalies computed over crop and rangeland areas and are updated every 10 days. They take into account the timing during the crop season at which they occur, using remote sensing derived phenology per-pixel. The second step involves the monthly analysis at country level by JRC crop monitoring experts of all the information available, including the automatic warnings, crop production and food security-tailored media analysis, high-resolution imagery (e.g. Landsat 8, Sentinel 1 and 2) processed in Google Earth Engine and ancillary maps, graphs and statistics derived from a set of indicators. Countries with potentially critical conditions are marked as minor or major hotspots and a global overview is provided together with short national level narratives.

Remote sensing monitoring of land restoration interventions in semi-arid environments with a before-after control-impact statistical design.

Restoration interventions to combat land degradation are carried out in arid and semi-arid areas to improve vegetation cover and land productivity. Evaluating the success of an intervention over time is challenging due to various constraints (e.g. difficult-to-access areas, lack of long-term records) and the lack of standardised and affordable methodologies. We propose a semi-automatic methodology that uses remote sensing data to provide a rapid, standardised and objective assessment of the biophysical impact, in terms of vegetation cover, of restoration interventions. The Normalised Difference Vegetation Index (NDVI) is used as a proxy for vegetation cover. Recognising that changes in vegetation cover are naturally due to environmental factors such as seasonality and inter-annual climate variability, conclusions about the success of the intervention cannot be drawn by focussing on the intervention area only. We therefore use a comparative method that analyses the temporal variations (before and after the intervention) of the NDVI of the intervention area with respect to multiple control sites that are automatically and randomly selected from a set of candidates that are similar to the intervention area. Similarity is defined in terms of class composition as derived from an ISODATA classification of the imagery before the intervention. The method provides an estimate of the magnitude and significance of the difference in greenness change between the intervention area and control areas. As a case study, the methodology is applied to 15 restoration interventions carried out in Senegal. The impact of the interventions is analysed using 250-m MODIS and 30-m Landsat data. Results show that a significant improvement in vegetation cover was detectable only in one third of the analysed interventions, which is consistent with independent qualitative assessments based on field observations and visual analysis of high resolution imagery. Rural development agencies may potentially use the proposed method for a first screening of restoration interventions.