Residue Density Segmentation for Monitoring and Optimizing Tillage Practices
RResidue Density Segmentation for Monitoring andOptimizing Tillage Practices
Jennifer Hobbs ∗ Intelinair, Inc.Champaign, IL 61820 [email protected]
Ivan Dozier
Intelinair, Inc.Champaign, IL 61820 [email protected]
Naira Hovakimyan
University of Illinois at Urbana-ChampaignIntelinair, Inc.Champaign, IL 61820 [email protected]
Abstract “No-till” and cover cropping are often identified as the leading simple, best man-agement practices for carbon sequestration in agriculture. However, the root of theproblem is more complex, with the potential benefits of these approaches dependingon numerous factors including a field’s soil type(s), topography, and managementhistory. Instead of using computer vision approaches to simply classify a field astill vs. no-till, we instead seek to identify the degree of residue coverage across afield through a probabilistic deep learning segmentation approach to enable moreaccurate analysis of carbon holding potential and realization. This approach willnot only provide more precise insights into currently implemented practices, butalso enable a more accurate identification process of fields with the greatest po-tential for adopting new practices to significantly impact carbon sequestration inagriculture.
Carbon sequestration is one of the primary topics raised in discussions around agriculture and climatechange. Soils have the capacity to be enormous carbon sources or sinks with farm management prac-tices significantly impacting how much carbon is held in the soil [8]. Past agricultural managementdecisions in the US have depleted global soil organic carbon (SOC) by as much as . ± . billionUS tn [12]. US cropland which covers roughly Mha is estimated to have the capacity to sequester . − . Mgha of carbon/year for a total potential of − T g carbon/year [2, 11]. Altering pastpractices to bring carbon out of the atmosphere and into the soil helps to both mitigate greenhousegasses, reduce negative agricultural contributions to the environmental, and improves soil health andwater holding capacity [15, 2, 17].Many initiatives around carbon sequestration for cropland are heavily focused around tillage prac-tices [3]. Residues consist of crop biomass such as dried leaves and stalks leftover from harvest; theseresidues contain key nutrients that the plants had absorbed during the season. By reincorporatingthese residues back into the soil, usually via tilling, farmers are able to recycle those nutrients: asresidues decompose, nutrients re-enter the soil, fueling the next year’s crops. In contrast, “no-till”and alternative tillage practices limit the amount of tillage conducted. Maintaining surface residueshas numerous benefits including increasing SOC and water capacity, increasing porosity, preventing ∗ corresponding authorTackling Climate Change with Machine Learning workshop at NeurIPS 2020. a r X i v : . [ c s . C V ] F e b igure 1: (Top Left) Aerial image of a field with highly variable residue density. Such variability isdue to a combination of management practices, soil composition, and topography (Bottom Left). Atground-level (Right), we see these same factors as well as crop-type and time-since-tillage impact thefield’s appearance.erosion, and enhancing soil stability, especially when used in combination with cover crops [17].Switching to no-till additionally requires one-less step in the farming life-cycle, saving labor time aswell as reducing fuel usage. However, residues can cause keep nutrients tied up in unusable forms,harbor pest and diseases, and inhibit emergence, leading to significant loss of yield if not adequatelymanaged [7].As a result, adoption of no-till and reduced-tillage practices vary widely across regions and crops,with only ∼ of farmland using no-till practices continuously [17, 4]. While many associateno-till and cover cropping as the key, beneficial approaches in carbon sequestration and erosionprevision, the impact of various tillage practices is far more complicated; the amount of carbon whichcan be sequestered with these practices can vary widely based on soil composition, moisture-levels,topography, and other management decisions [16, 2]. The economic benefit of these practices mustbe established in an accurate, personalized manner for each farm in order to promote widespread trustand adoption.The visual impact of these management choices manifest themselves in complex ways across thefield as seen in Figure 1. Capturing and understanding all of these contributing factors is criticalfor accurately assessing the impact of tillage practices on a particular farm as well as encouragingbroad adoption of these management practices; a simple classification of till/no-till is not enough.Therefore we segment the field into different levels of residue coverage to provide a fine-grained mapof this biomass layer. This information can be combined with soil, hydrological, and other modelsto more accurately determine the opportunity for carbon sequestration on a given farm parcel andthe impact of tillage practices to capitalize on those storage capacities. Identifying the density ofresidue further helps farmers more effectively manage residues across their field and enables novelprecision tillage practices. As opposed to treating tillage as a binary management practice, farmerscan take a targeted approach to till specific areas of their fields in the way that best addresses carbonsequestration, yield, nutrient, and erosion risks: casting these tillage practices not as a choice betweeneconomic and environmental needs, but as a strategic plan which maximizes both. The appearance of residues in a field is strongly dictated by the crop type. Identification of thecrop planted in the previous season can be accomplished from publicly available low-resolution( > may be feasible from low-resolution satellite imagery, segmenta-tion of residue levels is not. Therefore we will collect high-resolution ( < > soil visible), moderate ( − soil visible), heavy ( < soil visible), and ponding ( soilvisible with obvious multi-layer buildup) as seen in Figure 2. Since this is an inherently challengingannotation task due to the ambiguity over where specific borders end, each image will be annotatedby multiple annotators so that the distribution over annotators can be learned [9]. Fully convolutional encoder-decoder neural networks have proven highly successful in many segmen-tation tasks, including those in agriculture [5, 6, 13]. Following these and the probabilistic U-Netapproach above, we learn the distribution over the the plausible level segmentation; a five-channel(RGBN + topography) image serves as input into the model and a five-channel (one per level) imageis returned. The best approach to fuse topography and imagery will be explored [14].This mapping of residue levels alone provides value as it can be used to alert the farmer to areaswhich may experience emergence issues due to excessive residues and ponding. Furthermore, thisresidue map would be combined with other sources of information such as soil makeup, weather,topography, etc. and passed to downstream calculations and models to compute both the potential aswell as achieved carbon sequestered for the given farmland(Figure 2) [16, 1].
The sustainability and conservation goals that alternative and reduced tillage practices promise areonly attainable if widely adopted in significant areas. As with many conservation initiatives, till vs.no-till is often seen as a binary choice between maximizing economic or environmental outcomes,leading to slow or minimal adoption. This approach enables us to provide farmers with intelligenceabout their farm, enabling them to make the best decisions about where and how much to till for longterm sustainability in regards to both the environment and yield. The residue map can be furtherused to alert farmers to ponding boundaries and other areas which could be susceptible to disease,pests, and emergence suppression. Recent years have shown how precision agriculture practicesaround chemical and water applications have led to both economic and environment advances, andthis approach will enable the same for tillage practices.
Reframing the discussion around carbon sequestration for agriculture, not in the overly simplisticterms of till vs. no-till, but as precision residue management, is crucial for identifying the best tacticsfor a given farm, accurately quantifying the impact of those decisions, as well as promoting adoption.3igh-resolution aerial imagery and deep learning approaches will allow us to accurately determinelevels of residue across the field; because annotating densities is a difficult challenge, we will usea probabilistic segmentation approach to learn density levels over annotators. The final residuemap combined with other data layers such as topography and soil type will enable a more completeunderstanding of the potential as well as realized carbon sequestration opportunities for that field.
References [1] Humberto Blanco-Canqui and Rattan Lal. No-tillage and soil-profile carbon sequestration: An on-farmassessment.
Soil Science Society of America Journal , 72(3):693–701, 2008.[2] Adam Chambers, Rattan Lal, and Keith Paustian. Soil carbon sequestration potential of us croplandsand grasslands: Implementing the 4 per thousand initiative.
Journal of Soil and Water Conservation ,71(3):68A–74A, 2016.[3] CV Cole, J Duxbury, J Freney, O Heinemeyer, K Minami, A Mosier, K Paustian, N Rosenberg, N Sampson,D Sauerbeck, et al. Global estimates of potential mitigation of greenhouse gas emissions by agriculture.
Nutrient cycling in Agroecosystems , 49(1-3):221–228, 1997.[4] Elizabeth Creech. Saving money, time and soil: The economics of no-till farming.
USDA Conservationblog, November , 30, 2017.[5] Mulham Fawakherji, Ali Youssef, Domenico Bloisi, Alberto Pretto, and Daniele Nardi. Crop and weedsclassification for precision agriculture using context-independent pixel-wise segmentation. In , pages 146–152. IEEE, 2019.[6] Jorge Fuentes-Pacheco, Juan Torres-Olivares, Edgar Roman-Rangel, Salvador Cervantes, Porfirio Juarez-Lopez, Jorge Hermosillo-Valadez, and Juan Manuel Rendón-Mancha. Fig plant segmentation from aerialimages using a deep convolutional encoder-decoder network.
Remote Sensing , 11(10):1157, 2019.[7] Hero T Gollany, Jean-Alex E Molina, C Edward Clapp, Raymond R Allmaras, Milegua F Layese, John MBaker, and HH Cheng. Nitrogen leaching and denitrification in continuous corn as related to residuemanagement and nitrogen fertilization.
Environmental Management , 33(1):S289–S298, 2004.[8] Richard A Houghton. Why are estimates of the terrestrial carbon balance so different?
Global changebiology , 9(4):500–509, 2003.[9] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, KlausMaier-Hein, SM Ali Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-netfor segmentation of ambiguous images. In
Advances in Neural Information Processing Systems , pages6965–6975, 2018.[10] Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov. Deep learning classificationof land cover and crop types using remote sensing data.
IEEE Geoscience and Remote Sensing Letters ,14(5):778–782, 2017.[11] Rattan Lal. Sequestering carbon and increasing productivity by conservation agriculture.
Journal of Soiland Water Conservation , 70(3):55A–62A, 2015.[12] R Lal et al. Soil management and restoration for c sequestration to mitigate the accelerated greenhouseeffect.
Progress in Environmental Science , 1(4):307–326, 1999.[13] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedicalimage segmentation. In
International Conference on Medical image computing and computer-assistedintervention , pages 234–241. Springer, 2015.[14] Hao Sheng, Xiao Chen, Jingyi Su, Ram Rajagopal, and Andrew Ng. Effective data fusion with generalizedvegetation index: Evidence from land cover segmentation in agriculture. In
Proceedings of the IEEE/CVFConference on Computer Vision and Pattern Recognition Workshops , pages 60–61, 2020.[15] Pete Smith. Soils and climate change.
Current opinion in environmental sustainability , 4(5):539–544,2012.[16] Amy Swan, Mark Easter, Kevin Brown, Mark Layer, and Keith Paustian. Comet-planner: Carbon andgreenhouse gas evaluation for usda-nrcs conservation practice planning. 2018.[17] Tara Wade, Roger Claassen, and Steven Wallander. Conservation-practice adoption rates vary widely bycrop and region. Technical report, 2015.[18] Liheng Zhong, Lina Hu, and Hang Zhou. Deep learning based multi-temporal crop classification.
Remotesensing of environment , 221:430–443, 2019., 221:430–443, 2019.