A user friendly web-based solution for crop mapping for different contexts of in-situ data availability

The unprecedented availability of free and open Earth Observations (EO) data and accessibility to free and low-cost cloud computing provide the ideal conditions for implementing scalable solutions for improved agriculture monitoring and agricultural water management that can be used in operational contexts. However, the uptake of EO data in National Statistics Offices is still limited, especially in developing countries. The main reason for this is the common lack in countries of sufficient and high quality of in-situ data which is required to provide ground truth information for the training of the classification algorithms and for validation of crop maps.

In this context, FAO in partnership with hydrosolutions ltd have developed a user-friendly platform (named EOSTAT CropMapper) for high resolution mapping of crop types at country-scale using earth observations. All processing steps are implemented in Google Earth Engine. The system provides smooth access to crop maps, crop statistics and irrigation water requirements and works in three different contexts of in-situ data availability:

• Scenario 1: a large and accurate in-situ data is available. The system relies on a traditional Random Forest (RF) classifier.
• Scenario 2: a limited amount of in-situ data is available. The system relies on the use of a Dynamic Time Warping (DTW) algorithm to classify pixels into crop types based on only a few reference samples per crop type that represent the characteristic phenologies.
• Scenario 3: no in-situ data is available. The system relies on K-means clustering to map clusters of crop pixels. Subsequently the user is requested to associate each cluster to a crop label based on his expert knowledge.

In this contribution we present an overview of the methodology, of the functionalities of the tool and the architecture, and we provide results of the mapping workflow and the accuracy measures. The system has been first deployed in Afghanistan, but can be easily transferred to any place where samples of geotagged crop type information are available. Here we present an implementation of the EOSTAT CropMapper for Kashkadarya Region in Uzbekistan and an accuracy assessment of the crop type classification based on a dataset of ground-truth data (Remelgado et al., 2020). The reference ground-truth dataset consists of 2’172 crop type samples collected in the year 2018.

We demonstrate that the crop classification with DTW based on few carefully checked training samples can outperform conventional RF classification with at least two times more samples. With five times more training samples, RF outperforms DTW in terms of overall accuracy. The main condition for obtaining good results with DTW is a comprehensive quality assurance and quality control of the training data points. While the full ground-truth dataset consists of 2’172 samples, we used only 40-80 samples to train the DTW algorithm. It is understood that the quality assurance and control of such small samples sizes requires less time and is a more cost effective solution. RF is less sensitive to noise in the training data, and a large training data set can compensate the mistakes in the labeling of the ground-truth data.