SAR and Optical Data Applied to Early-Season Mapping of Integrated Crop–Livestock Systems Using Deep and Machine Learning Algorithms

Abstract

Regenerative agricultural practices are a suitable path to feed the global population. Integrated Crop–livestock systems (ICLSs) are key approaches once the area provides animal and crop production resources. In Brazil, the expectation is to increase the area of ICLS fields by 5 million hectares in the next five years. However, few methods have been tested regarding spatial and temporal scales to map and monitor ICLS fields, and none of these methods use SAR data. Therefore, in this work, we explored the potential of three machine and deep learning algorithms (random forest, long short-term memory, and transformer) to perform early-season (with three-time windows) mapping of ICLS fields. To explore the scalability of the proposed methods, we tested them in two regions with different latitudes, cloud cover rates, field sizes, landscapes, and crop types. Finally, the potential of SAR (Sentinel-1) and optical (Sentinel-2) data was tested. As a result, we found that all proposed algorithms and sensors could correctly map both study sites. For Study Site 1(SS1), we obtained an overall accuracy of 98% using the random forest classifier. For Study Site 2, we obtained an overall accuracy of 99% using the long short-term memory net and the random forest. Further, the early-season experiments were successful for both study sites (with an accuracy higher than 90% for all time windows), and no significant difference in accuracy was found among them. Thus, this study found that it is possible to map ICLSs in the early-season and in different latitudes by using diverse algorithms and sensors.

Mapping Integrated Crop–Livestock Systems Using Fused Sentinel-2 and PlanetScope Time Series and Deep Learning

“…Double crop and ICLS encompass soybean that is primarily cultivated as a summer crop subsequently followed by either another crop (e.g., corn, millet, or crotalaria) in Double crop or pasture in ICLS. Previous studies have also emphasized the challenge of discriminating Double crop and ICLS [13][14][15]. Our results showed an accuracy improvement compared to previous works focused on mapping ICLS that used data from MODIS sensor [13,14] and very close accuracy obtained in the study of Toro et al [15], which used S1 and S2 SITS.…”

“…Previous studies have also emphasized the challenge of discriminating Double crop and ICLS [13][14][15]. Our results showed an accuracy improvement compared to previous works focused on mapping ICLS that used data from MODIS sensor [13,14] and very close accuracy obtained in the study of Toro et al [15], which used S1 and S2 SITS. The quality of the prediction maps obtained from the fused time series was high for SA1, where the field size is smaller.…”

“…Kuchler et al [14] also suggested testing new machine learning algorithms that can better account for the temporal information required for detecting patterns in highly dynamic targets such as ICLS. In addition, Toro et al [15] investigated the application of deep learning algorithms on Sentinel-1 (S1) and Sentinel-2 (S2) time series to map ICLS early in the season. Despite S2 data generating the best mapping results, Toro et al [15] suggested analyses over multiple years and the use of other deep learning networks more specialized in learning spectro-temporal patterns.…”

“…In addition, Toro et al [15] investigated the application of deep learning algorithms on Sentinel-1 (S1) and Sentinel-2 (S2) time series to map ICLS early in the season. Despite S2 data generating the best mapping results, Toro et al [15] suggested analyses over multiple years and the use of other deep learning networks more specialized in learning spectro-temporal patterns. All previous works emphasized the difficulty of mapping ICLS due to their complexity and heterogeneity.…”

“…Motivated by the need to map sustainable agricultural production areas with ICLS and enhance their monitoring, this research aims to demonstrate an end-to-end workflow for mapping ICLS using existing state-of-the-art techniques for data fusion and deep learning algorithms applied to SITS. The RF classifier was used as a baseline for this research since it was tested in previous works for ICLS mapping [14,15]. In addition, we used three deep learning architectures designed to work with SITS; namely, 1D convolutional neural networks, represented in this research by Temporal Convolutional Neural Network (TempCNN) [22], Residual Network (ResNet) [23], and an attention-based model, called Lightweight Temporal Attention Encoder (L-TAE) [24].…”