Using mixed labels and a multi-stage approach to map crop types over smallholder-dominated agricultural systems

Abstract:

Accurate and timely crop type maps at field scales are critical inputs for estimating crop yields, and in turn provide key information for understanding agricultural practices and food security. However, developing such maps are particularly difficult for smallholder-dominated agricultural systems, where fields are small and irregular and necessary ground-truth data are sparse and hard to obtain. Newly released open training datasets such as Radiant MLHub’s crop type labels provide new opportunities to help overcome this information deficit. However, the structure and collection method of these datasets are not consistent. Uneven format and unbalanced distribution limits the application of advanced algorithms of Artificial Intelligence/Machine Learning. To overcome this problem and to make optimal use of these data, we used a three-stage modeling procedure. In the first stage, we used PlanetScope imagery and a new land cover mapping platform based on active learning to map individual field boundaries. In the second stage, we used the point-based labels to train a second RF model to predict crop types using temporal features extracted from Sentinel-1 time series, raw bands of PlanetScope seasonal composites, vegetation indices derived from Sentinel-2 seasonal composites, and other ancillary features such as altitude. We used the field boundaries from Stage 1 as masks for predicting crop types using the Stage 2 model. From these results, we select the predictions with the most regular geometries and highest classification certainties as new training labels. For Stage 3, we combine these model-generated labels with the polygonised field labels to train a DeepLabv3+ model, using the same image features in Stage 2. DeepLabv3+ is an extension of DeepLabv3 that adds a decoder module to refine the segments especially along their boundaries. However, low-quality labels limit the power of Deep Learning algorithms, prolong the training process, or increase the computational requirements, thus the first two stages were required to generate a training dataset of suitable size and quality. The results demonstrate the usefulness of an approach to ensemble inconsistent ground-truth labels and do crop type classification over smallholder-dominated agricultural systems.